Paper KkfN9BAN

Is there smart money?
How information in the futures market is

priced into the cross-section of stock returns with delay
Steven Wei Ho∗ Alexandre R. Lauwers§
First draft: November 17th, 2017

This draft: January 6th , 2019
Abstract
We document a new empirical phenomenon in which the positions of managed
money (MM) traders, who are sophisticated speculators in the commodity futures
market, as disclosed by the CFTC Disaggregated Commitments of Traders (DCOT)
reports, can predict the cross-section of commodity producers’ stock returns in the
subsequent week. The results are more pronounced in firms with higher information
asymmetry, proxied by analyst dispersion and historical volatility. We thus provide
more empirical evidence to the literature on investor specialization, market segmen-
tation, informational friction and gradual information diffusion. Our results are also
stronger in non-NBER recession periods. The Monday-to-Friday results are weaker
than the Wednesday-to-Tuesday results.
JEL codes: G11, G14
Keywords: Return Predictability, Limits to Arbitrage, Commodity, Market Seg-

mentation, Investor Specialization
‡ We are indebted to Brian Henderson (discussant) as well as the participants at the AFA 2019
Atlanta Annual Meeting for insights. We are also grateful for Alminas Zaldokas (discussant) and
seminar participants at the Paris Winter 2018 Conference. We thank Michael Johannes, Patrick
Bolton, Ana-Maria Fuertes, Lawrence Glosten, Matthieu Gomez, Lars Peter Hansen, Robert
Hodrick, Harrison Hong, Wei Jiang, Harry Mamaysky, Rahul Mukherjee, José Scheinkman, Cédric
Tille, Neng Wang and Laura Veldkamp for comments and suggestions. All errors are our own. We
thank Columbia University for research support and Wharton Research Data Services (WRDS)
was used in preparing firm level stock return and firm characteristics data. This subscribed
service and the data available thereon constitute valuable intellectual property of WRDS and/or
its third-party suppliers.
∗ Adjunct Assistant Professor, Department of Economics, Columbia University, 420 West
118th Street Office 1028A, Mail Code 3308, New York, NY 10027, email: [email protected].
§ PhD Student, The Graduate Institute, Geneva, Case postale 1672, 1211 Genève 1, Switzerland,
email: [email protected]
1. Introduction
Is there smart money in the commodity futures market? We answer the question by studying
the positions that Managed Money (MM) take in the weekly Disaggregated Commitments
of Traders (DCOT) published by the Commodity Futures Trading Commission (CFTC),
using a sample period from August 2006 to September 2018. The Managed Money (MM)
category of traders in the DCOT reports consists of hedge funds, mutual funds, large ETFs,
registered commodity trading advisers (CTAs), registered commodity pool advisers (CPOs),
and unregistered funds identified by the CFTC.
Hong and Stein (1999) established that the gradual diffusion of information can explain a
variety of predictability patterns of returns, as some investors can process only a subset of
publicly available information. The theoretical work by van Nieuwerburgh and Veldkamp
(2010) suggests that market participants cannot specialize in every asset as information
acquisition and processing is costly. In addition, Menzly and Ozbas (2010) have found
that due to investor specialization and market segmentation, value-relevant information
diffuses gradually in financial markets; specifically, stocks that are in economically related
supplier and customer industries cross-predict each other’s returns. Thus, we posit our main
conjecture as gradual information diffusion. Sophisticated investors (specifically, MM) who
are specialized in commodities and who trade in the commodity futures market, would react
to information updates related to commodity fundamentals faster than their unspecialized
counterparts and thus their voices would be diluted in the stock market by investors who are
not specialized in commodities. Therefore, one would expect that information extracted from
the commodity futures market can predict stock returns of commodity producing firms of the
same underlying commodity due to gradual information diffusion and market segmentation.
We find that the information contained in the commodity futures market, namely, the
1
positions of MM traders,1 who are sophisticated speculators, can predict the cross-section of
stock returns for commodity producers. In particular, if the DCOT reports an increase in
long position, or a decrease in short position, or an increase in net position, or an increase
in the ratio of long over short position of MM, then the stock price of producers of the
same commodity would increase in the following week. This finding is robust to a variety of
measures and weighting schemes. The fact that the commodity market is linked with other
financial markets has been documented by Hong and Yogo (2012), in which they find that
among other things, high commodity market activity, as measured by high open-interest
growth, predicts low bond returns. In other words, Hong and Yogo find that information
contained in the commodity futures market as measured by open interest growth, which is
an informative signal of future inflation, is priced into the bond market with a delay. We are
therefore not surprised that information contained in the positions of managed money in the
commodity futures market is priced in the equity market with delay, in the cross-section of
commodity producing firms.
We investigate why predictability arises by studying the relation of our results with
measures of information asymmetry. We find that our results are stronger in firms with higher
information asymmetry as measured by ex ante analyst dispersion and 60-day historical
stock volatility. In addition to informational friction, trading friction can also be regarded as
a form of market friction; thus we also investigate whether our predictability results arise
due to trading friction, and liquidity in particular, by double-sorting our commodity futures
market signal with Amihud’s illiquidity measure (Amihud, 2002). We find no evidence that
predictability is stronger (weaker) in firms with higher trading friction. Therefore, we argue
that informational, rather than trading friction, contribute to our predictability results.
In addition to finding abnormal return relative to the Carhart (1997) four-factor model
1 Our study is made possible with CFTC’s decision to publish disaggregated Commitments of
Traders (COT) reports for trades after June 13, 2006, which created the MM category of traders.
We will elaborate more on this point in Section 3.1.2. According to the CFTC website, the transition
from legacy COT reports to disaggregated COT reports was made because the new format provides
greater market transparency by including more delineated trader classification categories beyond
the two broad categories found in the legacy format.
2
and the Fama and French (2015) five-factor model, we show that abnormal return remains
even relative to the Stambaugh and Yuan (2016) model, which includes mispricing factors
constructed from a broad set of well-known anomalies including momentum, net stock issues,
asset growth, investment to assets, distress, O-score, and profitability, among others. However,
some return anomalies generate abnormal returns mainly from short-selling overpriced stocks
(Stambaugh et al., 2012), and such anomalies may be difficult to be arbitraged away due
to short-selling constraints. This is not the case in our particular finding, as the abnormal
return (alpha) in our long-short strategy is not solely due to the contribution of the short-leg
of the Long-Short portfolio. We also employ a number of empirical methodologies in our
analysis, including, single-sort, double-sort, and Fama-Macbeth regressions to confirm the
robustness of our baseline results.
To the best of our knowledge, this is the first paper that shows the cross-sectional
predictability of commodity producers’ stock returns based on the information contained in
the commodity futures market. Our contribution is mainly empirical rather than theoretical.
We contribute to the literature by documenting a robust return predictability phenomenon,
which are more pronounced in firms with higher information asymmetry, and by doing
so we provide more empirical evidence to the literature on investor specialization, market
segmentation, informational friction and gradual information diffusion. Our results are also
stronger in non-NBER recession periods. Furthermore, since we already know that the
Monday-Friday results are weaker than the Wednesday-Tuesday results, we thus do not
believe the predictability to be driven by the announcement of CFTC reports, and that
the predictability seems to be already present prior to the CFTC’s Friday release of trader
positions as of Tuesday.
The rest of the paper is organized as follows: Section 2 reviews the background and
related literature. Section 3 discusses how information on traders’ positions from the DCOT
reports is extracted, matched to the sample of commodity producers stocks, and used as
leading signal to form the Long-Short portfolio. We present our empirical results in Section 4,
3
including results on portfolio alpha, single sorting, Fama-Macbeth cross-sectional regressions,
and double-sorting. Finally, Section 5 concludes.
2. Background and Literature

Over the past 15 years, financial institutions have substantially increased their exposure to
the commodities. By studying the commodity-linked notes, Henderson et al. (2014) have
found that financial investor flows have significant impacts on commodity futures prices.
However, Sockin and Xiong (2015) have noticed that market participants of the commodity
market face severe informational frictions and developed a theoretical model that illustrates
the importance of feedback effects of commodity prices, in that a higher commodity price
signals stronger economic growth and may motivate goods producers to produce more goods
which leads to greater demand for the commodity as inputs. In addition, the paper also
argues that the futures price is not merely a shadow of the spot price and instead may contain
additional information.
By using monthly industry stock portfolio return data, Hong, Torous and Valkanov (2007)
found that industry portfolios including retail, services, commercial real estate, metal, and
petroleum, forecast the aggregate stock market by two months. The results support the
hypothesis that investors who specialize in trading the broad market index, receive information
originating from particular industries such as commercial real estate or commodities like
metals and petroleum only with a lag, and therefore the returns of industry portfolios that
are informative about macroeconomic fundamentals will lead the aggregate market. Our
study, however, involves information for a number of commodities in the futures market,
rather than the industry portfolios in stock market, in the construction of our leading signals.
Furthermore, we analyze predictability of firm-level stock return in contrast to aggregate
stock return. In addition, we study predictability at the weekly frequency rather than at the
monthly frequency, and we specifically find that information pertaining to the MM category
in the commodity futures market, as opposed to other categories of traders, have predictive
4
power at the weekly frequency.
Büyükşahin and Robe (2014) show that hedge funds’ activity matters for the linkage
between equity and commodity market. They also find that, after controlling for macroeco-
nomic and commodity-market fundamentals, comovements in commodity-equity markets
are positively related to greater commodity market participation by financial speculators
and especially hedge funds, although no such effect exists for other kinds of traders. Their
measure of commodity-equity market linkage is based on estimates of the dynamic conditional
correlations between the rates of return on two investable commodity indices (GSCI and
DJ-UBS), versus, the unlevered rate of return on the S&P 500 equity index. In addition,
Basak and Pavlova (2016) provides a theoretical model in which financialization of com-
modities would lead to increased correlation between the commodity and equity market
returns. Rather than using aggregate indices, we instead investigate the cross-section of
commodity-producing firms by using firm-level data, which allows us to employ specifically
cross-sectional methodologies such as Jensen’s alpha (abnormal return) analysis, single-sort,
double-sort, and Fama-Macbeth regressions. Furthermore, we are more concerned with the
lead-lag relationship between the two markets and whether the market linkage leads to stock
return predictability.
Utilizing COT data in the old format published at bi-weekly or monthly frequency, in
which the MM category is not established, Fernandez-Perez et al. (2017) find that the
information contained in the commodity futures market can be used to construct two risk-
factors useful for predicting long-horizon aggregate equity market return and for the business
cycle.
An extensive body of literature relates expected futures returns to the net (long minus
short) positions of hedgers in the futures market, known as hedging pressure. Markets in
which hedgers are net short (long) are found to have positive (negative) expected futures
returns per Carter et al. (1983), Bessembinder and Seguin (1992), and De Roon et al. (2000).
However, by exploring the COT reports in the energy market (including crude oil, unleaded
5
gasoline, heating oil, and natural gas futures contracts), Sanders, Boris and Manfredo (2004)
find that there is a positive contemporaneous correlation between commodity market returns
and positions held by noncommercial traders and a negative contemporaneous correlation
between commercial (commodity producers) positions and commodity market returns; yet,
they find that traders’ net positions, whether commercial or non-commercial, are generally
not useful in predicting weekly energy futures returns; and Sanders et al. (2009) generally
confirms the same result in the corn and live cattle futures market. On the other hand,
Buchanan et al. (2001) find that non-commercial positions do provide information on the
magnitude and direction of weekly price changes in the natural gas futures market. Wang
(2003) find that a sentiment index based on positions of commercial and non-commercial
traders are useful for predicting S&P 500 index future returns. In addition, by analyzing
CFTC’s COT reports of major currency futures, Tornell and Yuan (2012) find that the peaks
and troughs of commercial traders and non-commercial traders’ net positions are generally
useful predictors to the evolution of spot exchange rates, and a simple trading strategy based
on peaks/troughs proves to be quite profitable. Gorton, Hayashi and Rouwenhorst (2013)
find no evidence that participants’ positions in futures markets would predict risk premiums
on commodity futures. We note that our paper, however, is studying the predictability of
returns in the equity market rather than in the futures market.
Lastly, in the equity market, a number of studies have shown that investor sentiment
related measures are useful to predict the future development of stock returns, such as trading
volume, mutual fund flows, closed-end fund discounts, option implied volatility, and insider
trading information (see survey by Baker and Wurgler, 2007). Especially relevant is the
study by Han (2007), who find that a sentiment measure defined by the number of long
noncommercial contracts minus the number of short noncommercial contracts in S&P 500
futures (scaled by the total open interest), with data obtained from the COT reports, can
help explain the shape of the S&P 500 index option volatility smile and risk-neutral skewness.
Specifically, when investor sentiment is more bullish, the index risk-neutral skewness becomes
6
less negative and the index option volatility smile is flatter; and vice versa.
3. Data and Estimation
3.1. Data on Stocks and Commodity Futures
3.1.1. Matching Futures Market Positions with Producers’ Stocks
We utilize data from both the stock market and the commodity futures market. We use
daily data from the Center for Research in Security Prices (CRSP) for ordinary common stocks
(CRSP SHRCD= 10 or 11) traded on the New York Stock Exchange (NYSE), American
Stock Exchange (AMEX), and NASDAQ (CRSP EXCHCD= 1, 2, or 3).2 Stock returns are
based on the daily individual stock returns with dividends (CRSP RET) adjusted by the
delisting returns (CRSP DLRET).
To match commodity-producing firms with the commodities for which the CFTC is
collecting Commitments of Traders (COT) information, we follow the simple rule suggested
by Gorton and Rouwenhorst (2006). For each commodity that can be associated with
a four-digit U.S. Standard Industrial Classification (SIC) code,3 we associate with that
commodity all publicly-traded companies with that same four-digit SIC code. Ultimately, 10
commodities are considered: 2 industrial metals—copper and steel; 3 precious metals—gold,
silver and miscellaneous metals; 4 energy commodities—biofuel, oil & gas (crude oil and
natural gas), petroleum (unleaded gas) and coal; and one soft commodity—lumber. Reasons
2 We exclude firms with SHRCD 12 in our baseline analysis since these corporations are in-
corporated outside of the United States, i.e., U.S.-listed foreign firms. Thus, their equity prices
are subject to additional country-specific risk premia and country-specific shocks to asset prices
(Foerster and Karolyi, 1999; Karolyi and Stulz, 2003; Ammer, Vega and Wongswan, 2010). We also
exclude foreign firms trading on U.S. exchanges as ADRs (American depositary receipt, SHRCD 31
32). However, we will expand our scope in robustness checks and include SHRCD 12 in Table C.2.
3 An obvious complication arises when a commodity and a publicly traded commodity producer
may not be directly related. In the stock market, commodity-producing companies are not necessarily
only involved in the production and processing of a commodity but are also involved in a number of
sideline businesses. Hence, this methodology sacrifices precision for the simplicity necessary to build
commodity equity-based portfolios. When the SIC code does not offer reliable identification of firms
or yield a sufficient number of firms, we hand-pick firms by searching for U.S.-listed companies with
their main line of business being the production of a commodity. The hand-picked firms’ CRSP
PERMNO are listed in the second column of Table B.1
7
for not selecting more soft commodities are twofold. First, based on their SIC codes, too
few publicly-traded U.S. producers can be matched to a unique soft commodity, such as
coffee, sugar, cattle, cocoa or cotton. Second, most of the listed producers in the soft category
are highly diversified firms exposed to several agricultural goods simultaneously, rendering
further analysis and construction of the Long-Short portfolio difficult.4 Table B.1 in the
appendix provides the details on the matching of commodities with producers’ stocks.5 Our
panel sample consists of 127 firms, on average per week, from August 2006 to September
2018.
3.1.2. CFTC Data and Signal Measures
Our analysis is based on data from August 2006 to September 2018, due to data availability
on commodities’ disaggregated weekly reportable positions.6 We use position data and
the trader classification from the CFTC’s Commitments of Traders (COT) reports. The
publicly available Disaggregated Commitments of Traders (DCOT) report provides weekly
information on traders’ positions for 5 categories of market participants who are active in the
commodity futures markets. The 5 trader categories are defined in Table 1. In the legacy
format, however, the COT report divided reporting traders into the two broad categories of
‘‘commercial’’ and ‘‘non-commercial’’ traders. The ‘‘producer/merchant/processor/user’’ and
‘‘swap dealers’’ categories combined would equate to the commercial category in the legacy
4 In the previous version of this paper, we had a ‘Cattle’ commodity which we have removed in
this version because we think firms involved with cattle business are too diversified in other lines of
businesses.
5 Crude oil and natural gas are grouped together as firms engaged in crude oil production often
also engage in natural gas production. For these firms, the signals in the commodities futures
market are weighted yearly by the lagged U.S. oil segment and gas segment’s total revenue (data
retrieved from the U.S. Energy Information Administration). Miscellaneous metals comprise of
platinum and palladium which are grouped together as firms engaged in platinum production often
also engage in palladium production. In addition, price movements of these two rare metals exhibit
a very high degree of correlation in the post 2006 sample.
6 CFTC regulations require that clearing members, futures commission merchants, and foreign
brokers file reports with power granted to the Commission under the Commodity Exchange Act.
We use the Disaggregated Futures Only Reports. The CFTC also publishes reports combining
traders’ positions in futures and option markets. The CFTC began publishing the weekly Futures
Only COT reports in a disaggregated format on September 4th, 2009 and provided historical data
back to June of 2006 on October 20, 2009. Per CFTC: the COT reports provide a breakdown of
open interest for markets in which 20 or more traders hold positions equal to or above the reporting
levels established by the CFTC.
8
Table 1: CFTC Classification of Commodity Markets Participants from DCOT reports
Markets Participants Description

An entity that predominantly engages in the production,
Producers, Merchants, Proces- processing, packing or handling of a physical commodity and
sors, and Users (PM) uses the futures markets to manage or hedge risks associated
with those activities.
An entity that deals primarily in swaps for a commodity

and uses the futures markets to manage or hedge the risk
associated with those swaps transactions. The swap dealer
Swap Dealers (SW)
counterparties may be speculative traders, such as hedge
funds, or traditional commercial clients that are managing
risk arising from their dealings in the physical commodity.
Entities that manage and conduct organized futures trading

on behalf of their clients. This category includes registered
commodity trading advisers (CTAs), registered commodity
Money Managers (MM)
pool advisers (CPOs), and unregistered funds identified by
the CFTC. Hedge funds and large ETFs are part of this
category.
Other reportable traders who are not placed into one of the
Other Reporting traders (OR)
above three categories.
Non-Reporting traders (NR) Smaller traders who are not obliged to report their position.
report. Additionally, the ‘‘managed money’’ and ‘‘other reporting’’ groups combined would
equate the non-commercial (large speculator) category in the legacy report. DCOT reports
have normally been released every Friday at 3:30 p.m. Eastern time with the open interest
data as of the end-of-day on the Tuesday of the same week.
It should be noted that according to the CFTC Website: the actual trader category
or classification is based on the predominant business purpose self-reported by traders on
the CFTC Form 40 and is subject to review by CFTC staff for reasonableness, and failure
to answer the form truthfully is a violation of the Commodity Exchange Act and CFTC
regulations with violators subject to criminal or administrative sanctions. Furthermore, the
traders are able to report business purpose by commodity and, therefore, can have different
classifications in the COT reports for different commodities. Due to legal restraints however,
the CFTC does not publish information on how individual traders are classified in the COT
reports.
9
Traders’ positions for a given market are aggregated across all contract expiration months.
Managed money (MM), swap dealers (SW) and other reporting (OR) positions are divided
into long (l), short (s) and spreading (sp)—which measures the extent to which a trader
holds equal long and short futures positions;7 whereas producer/merchant/processor/user
(PM) and non-reporting (NR) positions are simply divided into long or short.8 The following
relation explains how the market’s total open interest (TOI, i.e., the number of futures
contracts outstanding) is disaggregated in the DCOT report, and the expressions above each
brace represent the contribution of open interest accountable to each of the trader categories:
   
   
 (PMl + PMs) + (SWl + SWs + 2SWsp) + (MMl + MMs + 2MMsp) + (ORl + ORs + 2ORsp) + NR s + NRl  = 2 × TOI
  
 | {z } | {z } | {z } | {z } | {z } | {z }
producer/merchants swap dealers managed money other reporting traders non-reporting total
traders open interest
processors/users
Figure B.I in the appendix shows the average market share held by each of the five
trader categories over the sample period and across all 10 commodities. On average, 24.3%
of the market share goes to MM, 30% to PM, 19% to SW, 17.7% to OR, and 9% to NR.
Table B.2 in the appendix provides a summary of the net positions for the different DCOT
trader categories in our sample of 10 commodities. For each commodity, we report the
position by each trader category as measured by the average weekly net long position (long
position - short position, and then scaled as a percentage of the open interest of that trader
category), its standard deviation, the percentage of the weeks the position is long, and the
first-order autocorrelation coefficient of the position. First, we observe that MM and NR
traders are on average net long in most markets, whereas PM positions are on average net
short, consistent with the notion that they act as hedgers for the most part. In contrast, SW
and OR positions are less clear-cut on average. MM traders can both long and short different
7 For example, if a non-commercial trader holds 2,000 long contracts and 1,500 short contracts,
500 contracts will appear in the ‘‘long’’ category and 1,500 contracts will appear in the ‘‘spreading’’
category.
8 In the legacy format, there was no spreading category for the commercial traders and there still
is not one for PM, since spreading is not considered a commercial activity (Turner, 2009).
10
commodities in a given week, and their average net long position across commodities is 25%.
In addition, the table shows both a large time series variability in net positions over time and
large cross-sectional differences across commodities. The average standard deviation of MM
net position across commodities is approximatively 33% per week. While MM traders are
almost equally likely to be long or short in copper, steel, coal and lumber, their positions are
long more than 90% of the weeks for the other commodities (except for oil & gas). Finally,
unsurprisingly for weekly positions, all traders’ positions across all commodities exhibit a
high degree of persistence. The first-order autocorrelation of MM positions ranges from 0.92
for biofuel to 0.99 for steel, miscellaneous metals and oil & gas.
In this paper, we mainly focus on the positions of MM traders.9 These traders take
speculative positions, invest others’ money on a discretionary basis, may make use of leverage
and usually do not intend on taking delivery of the underlying commodities they are trading.
We mainly focus on MM in our paper since they are the category of traders who have
the most incentive to seek out and process information related to changes in commodity
markets and they are believed to have the expertise. Indeed, the theoretical model by van
Nieuwerburgh and Veldkamp (2010) concludes that because information acquisition and
processing is costly, the optimal learning strategy for investors is to concentrate on one or
a small set of assets. In addition, MM may engage in a higher frequency of trading than
commercial hedgers and thus are more sensitive to information related to the short-term.
Later, we do find that MM positions in the commodity futures market can predict the
cross-section of stock returns whereas the signal extracted from PM positions are less robust.
To investigate the informativeness of MM positions in commodity futures markets, we form
four different, albeit closely related, measures. The results for all four measures usually
agree with each other. Prior studies have used CFTC open interest data to explore market
effects caused by trader categories, although they generally either focus on the returns within
the commodity futures market or on the equity market as measured by aggregate market
9 The CFTC reports use the terminology ”Managed Money”, and ”Money Managers” interchange-
ably.
11
indices rather than on a cross-section of firms.10 However, the disparity of results suggests
that multiple position indicators should be utilized to understand the robustness of the
results. The first signal measure we use is the ‘‘Net Change’’ in MM positions, defined as the
percentage change in long MM positions minus the percentage change in short MM positions
(M Ml +M Msp )t (M Ms +M Msp )t
( (M Ml +M Msp )t−1
− (M Ms +M Msp )t−1
); incidentally in this measure the spreading positions held
by long and short traders cancel out in the numerator. We also compute the signal measure:
‘‘Long Short Ratio Growth’’, which is defined as the growth in long positions of MM traders
M Ml +M Msp
divided by their short positions, (growth in M Ms +M Msp
). We also utilize the ‘‘Long Proportion
Growth’’, which is calculated as the growth rate in the long positions of MM traders divided
M Ml +M Msp
by the total positions held by MM: (growth in M Ml +M Ms +2M Msp
). Similarly, we also use
M Ms +M Msp
the signal measure ‘‘Short Proportion Growth’’ (growth in M Ml +M Ms +2M Msp
). In addition,
although not directly linked to our main hypothesis nor is it a measure of MM’s position per
se, we also examine the growth of total open interest (TOI) held by all trader categories, i.e.,
the total amount of futures contracts outstanding, as in Hong and Yogo (2012). Furthermore,
we also investigate the four analogous measures based on PM positions.
3.2. Data Manipulation and Portfolio Formation
3.2.1. Matching the Weekly Position Data with the Daily Stock Return Data
As previously described in Section 3.1.1, for each of the selected commodity, we identify the set
of firms with the corresponding industry code. This procedure generates 10 baskets of equity
portfolios, each based on a commodity. Each commodity-equity portfolio has a daily return
series, and is computed either as the value-weighted average of stocks’ returns belonging to
the same commodity (thereafter referred to as V-Weight) or as the equal-weighted average
(thereafter referred to as E-Weight). However, in either case, as will be described in Section
3.2.2, subsequently the 10 commodity-equity portfolios’ returns are then weighted equally,
or weighted according to the strength of the commodity signal, when calculating overall
10 See,for instance, Büyükşahin and Robe (2014); Sanders et al. (2004); Cheng et al. (2014);
Singleton (2013); Hamilton and Wu (2015)
12
portfolio returns; and the E-Weight versus V-Weight dichotomy only refers to the weighting
of firm returns within each commodity-equity portfolio. We will see depending on whether
or not subsequently the individual commodity portfolios are weighted by the strength of the
commodity signal measure, E-Weight will then give rise to the No-Weight and Degree-Weight
weighting schemes; whereas V-Weight will give rise to the Value-Weight and All-Weight
weighting schemes when calculating the returns of overall holdings.11
The DCOT reports are tabulated weekly from the beginning of trading on Wednesday to
Tuesday’s close (which is the compilation date). We match this time interval by computing
weekly returns of portfolios of commodity-producing firms from Wednesday through the next
compilation date. To be precise, the compilation date, which is usually a Tuesday unless
it’s a federal holiday, is considered the ‘‘signal generation date’’ for signals which we would
utilize a day later beginning on Wednesday in determining whether to long or short each
of the 10 commodity-equity portfolio. The same portfolio is held from the beginning of the
trading day right after the compilation date through the end of the next compilation date,
and this process is henceforth referred to as the Wednesday-to-Tuesday convention.12 The
weekly commodity-equity portfolio returns series are then further aggregated, with one of
the weighting schemes to be described in Section 3.2.2, to generate the overall return of our
Long-Short portfolio.
To summarize, and as will be made more precise with Equation 2 in Appendix A, here
we match weekly commodity position data with daily stock return data, and we begin by
building commodity-equity portfolios and calculating daily returns of the portfolios, and
then we compound the daily portfolio returns to obtain weekly returns for each of the
commodity-equity portfolios (which will then be further aggregated to yield weekly returns
11 Following Fama and French (1993), for V-Weight we use the market capitalization measured at
the end of December of the previous year as weights. While equal weighting will emphasize smaller
stocks more, our sample is less prone to this bias because commodity producers are usually large
firms (in terms of market capitalization). Nevertheless, one of the robustness checks we perform is
to drop all micro-cap stocks, which will be shown in Table C.4.
12 The CFTC is sometimes inconsistent with its handling of federal holidays, occasionally delaying
the compilation date by one or two days. We adjust the aggregation step accordingly using CFTC’s
actual compilation dates (directly provided in the CFTC data).
13
for our Long-Short portfolio). Our empirical results are also robust compared to another
method in which, instead of calculating daily returns for each of the commodity-equity
portfolios and then compounding them to yield weekly returns, we calculate weekly returns
for each of the stocks inside the portfolios and then construct and compute the weekly returns
for the commodity-equity portfolios.
Although the CFTC does not release traders’ positions as of Tuesday until Friday, we
nevertheless use the Wednesday-to-Tuesday convention in our baseline analysis since we are
interested in the study of market segmentation, specialization of market participants and slow
information diffusion following the line of research by (among others): Hong et al. (2007),
Menzly and Ozbas (2010), and van Nieuwerburgh and Veldkamp (2010) and ultimately
Grossman and Stiglitz (1980). Afterall, if MM are voting with their money about the certain
commodities’ prospects in the commodities futures market which reflect their current up-to-
date information regarding the commodities, why don’t they also trade simultaneously in the
equity market with the stocks of the commodity producers? It is due to the presence of some
uninformed (or unspecialized) investors in the equity market who, because of their limited
information-processing capabilities or lack of specialization in commodity-related information
due to costly information processing, fail to adjust their demand which results in a lead-lag
relationship in prices.
Nevertheless, we also study whether whether the time lag for incorporating information
contained in MM’s futures markets positions into producers’ equity prices is longer than what
was captured by the Wednesday-to-Tuesday convention and whether we our predictability
results remain, based on a Monday-to-Friday convention consistent with the actual release
schedule of DCOT reports by the CFTC. In other words, in this convention the report
release date, which is usually a Friday unless it’s a federal holiday, is considered the ‘‘signal
generation date’’ for signals which would determine the portfolio formation on the following
Monday. The results will be discussed in Section 4.4.1.
14
3.2.2. Long-Short Portfolio Formation
The Long-Short portfolio is constructed by sorting the commodity-equity portfolios weekly

according to their lagged MM position signals by going long (short) on stocks of commodity-
producing firms associated with a positive (negative) commodity signal if the signal we
choose to utilize is Long Proportion Growth, Net Change or Long Short Ratio Growth; and
vice versa if the signal is Short Proportion Growth. For example, if we are following the
Wednesday-to-Tuesday convention as we will in our baseline analysis, we would form our
weekly portfolio starting from Wednesday July 17th, 2013 utilizing the growth rate of MM’s
Long positions (or any one of the four signal measures) compiled by CFTC on Tuesday
July 16th over that value compiled on Tuesday July 9th. We will subsequently refer to the
timing of such signals as ‘‘lag1’’ or ‘‘1-week lag’’.13 As will be shown later in Table 3, we
also consider the case of utilizing J-week backward-looking moving averages of lagged signals
to dictate our portfolio formation, to be denoted with timing convention MA(J).
The return series for both the Long and Short portfolios are computed from the commodity-
level equity portfolios either according to the degree of the signals (the strength of the
commodity signal measure, i.e., their distance from zero), thereafter referred to as the
D-Weight as in Appendix A, or by averaging with equal-weight the returns of individual
commodity-level equity portfolio in the Long (or Short) portfolio, thereafter referred to as
the N-Weight. Combined with the two weighting schemes in the previous step as described
on page 12, we finally form a zero-investment Long-Short portfolio by taking the difference
between the Long and Short portfolio returns. Hence, we can then derive time-series of
weekly returns for each of the 4 types of Long-Short portfolios, representing the following: (1)
the Value-Weight, formed by applying V-Weight and then N-Weight; (2) the Degree-Weight,
formed by applying E-Weight and then D-Weight; (3) the All-Weight, formed by applying
V-Weight and then D-Weight; and (4) the No-Weight, formed by applying E-Weight and
then N-Weight. Appendix A summarizes the procedure to compute the Long-Short portfolio
13 For the Monday-to-Friday convention, we utilize an analogous procedure.
15
returns.
Figure B.II shows the total number stocks traded in the Long-Short portfolio each week
from August 2006 to September 2018. It should be noted that the minor dips and spikes
in the number of stocks traded are due to the fact that the commitments of traders data
corresponding to the steel commodity are often missing in the CFTC DCOT reports for
a number of weeks. The same problem occurs, although only occasionally, for coal prior
to August 2012, though they are continuously reported from August 2012 onwards. In
other words, two commodities sometimes experience data gaps due to the lack of positions
information in the CFTC reports. Thus, to ensure that our results are not driven by these
intermittent gaps in DCOT report, we also repeat our empirical analysis by removing steel
from the whole sample and by removing coal prior to August 2012. The results remain largely
the same and for sake of brevity, we do not report the results.
4. Empirical results
Our empirical analysis investigates the lead-lag relationship between the commodity futures
and stock markets by exploiting the joint dynamics of commodity producers’ equity price
changes and MM position changes in the commodity futures market. First, we present the
long-run performance of the Long-Short portfolios’ as described in 3.2.2. Second, we examine
the portfolios’ abnormal returns by using the standard time-series regression calculation of
Jensen’s alpha relative to the Fama and French three-factor model augmented with the
momentum factor, i.e., the Carhart four-factor model, and also other popular factor models as
robustness checks. Our results are then confronted to a single-sorting procedure to study the
price impacts of different signals and whether they exhibit a monotonic pattern according to
the strength of the signal. In addition, we examine return predictability with Fama-Macbeth
cross-sectional regressions across a range of time-lags. We also conduct a number of robustness
checks. Finally, we discuss the relationship between our results and market frictions that
could potentially contribute to our predictability results.
16
4.1. Portfolios’ Long-run Performance
Figure 1: Cumulative Gains from Investments, Signaled by the 1-week lagged Managed
Money Long-Short Ratio Growth (Value-Weight)
6 6
5 5
Value of $1 invested in a portfolio
4 4
3 3
2 2
1 1
0 0
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018
Long portfolio Short portfolio Long-Short portfolio
The cumulative return plots illustrate the cumulative weekly returns for investments in the
zero-investment portfolio (‘‘Long-Short portfolio’’) that takes a long position in stocks with
positive signal growth rates (‘‘Long portfolio’’) and a short position in equities with negative
growth (‘‘Short portfolio’’), unless the signal measure used is Short Proportion growth in
which case the long and short are flipped. Figure 1 presents the Value-Weight Long-Short
portfolios’ long-run performance signaled by the 1-week lag of MM Long Short Ratio Growth.
There is considerable difference between the Long and Short portfolios’ performance: an
investment of $1 in the Long portfolio in August 2006, when our data begins, would have
grown to $2.86 by September 2018; whereas a $1 investment in the Short portfolio would
have declined to $0.52 over the same period. A dollar invested in the Long-Short portfolio
17
Figure 2: Cumulative Gains from Investments, Signaled by the MA(2) Managed Money Net
Change (No-Weight)
32 32
28 28
24 24
20 20
16 16
12 12
8 8
4 4
0 0
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018
at start would have grown to $3.80 by September 2018, representing an annualized mean
excess return of 12.75%. The long-short strategies’ return pattern reveals a upward trend
over the entire sample period, with a sharp rise and heightened volatility from 2015 onwards.
The shaded area corresponds to the NBER recession periods. In Figure 2, the No-Weight
Long-Short portfolios’ cumulative weekly returns based on the 2-week backward moving
average of MM Net Changefollows a similar pattern.
The cause of the heightened volatility lately can be partly attributed to the financial
turmoil in China including a stock market crash and subsequent economic stimulus beginning
in 2015, which spillovered to the commodity market, as well as uncertainties about the future
prospect of China’s demand of commodities (Gross, 2017; Sada, 2016; Hutt, 2015). The
2015-2016 period saw large falls and subsequent recoveries in a number of commodity prices,
18
in line with large swings in actual and expected demand from China, the world’s largest
commodity consumer. As an asset class, commodities appear to have made a significant
bottom in late 2015, but the markets began to improve in the weeks and months that followed
beginning in early 2016. Taken together, Figure H.I and Figure H.II in the appendix indicate,
the large increase in the performance of the overall portfolio return beginning in 2015 as
observed in Figures 1 and 2, which were steadily growing even before this period, can not
be entirely attributed to any one commodity in particular. On closer inspection of the
contributions of each commodity equity-based portfolios, the performance of the Long-Short
portfolio in 2015 was mainly tied to returns from the Short portfolio—especially due to the
contributions of the Miscellaneous Metal and Silver portfolios in early 2015 and the Copper
and Coal portfolios in late 2015 according to Figure H.I. On the other hand, as the commodity
markets began to recover in early 2016, the performance of the Long-Short strategy was
driven by the outperformance of the Long portfolio—due to the contributions of commodities
such as Gold, Coal, Silver, Copper and Miscellaneous Metal according to Figure H.II.
Table 2 presents the return moments and summary statistics of the Long-Short portfolios
for each of the four weighting schemes as described in Section 3.2.2, that are constructed
either with the 1-week lagged or the 2-week backward moving average of the four MM signal
measures. Across all the different combinations of Long-Short portfolios and the signal
measures, the annualized mean excess returns are generally statistically significant with
t-statistics ranging from 1.87 to 3.72. The magnitude tends to be quite varied cross the
weighting schemes used, with the No- and Degree-Weights generally yielding higher values
than the Value- and All-Weights schemes. The risk-adjusted performances, as measured
by the annualized Sharpe ratios, vary from 0.504 to 1.035, which are measured by firstly
√
calculating weekly Sharpe ratios and then multiplied by 52. For brevity, we do not present
the results for other MA(J) signals beyond J=2.
19
Table 2: Long-Short Portfolio Characteristics
Notes: Table 2 presents the return moments and summary statistics of the Long-Short portfolios
for all four weighting schemes constructed either with the 1-week lagged or the 2-week backward
moving average of the four Money Managers’ (MM) signal variables. AW, DW, VW and NW stand
for All-Weight, Degree-Weight, Value-Weight and No-Weight, respectively, as defined in Section
3.2.2. The Long-Short portfolios are rebalanced weekly, and we follow a strategy of buying the
producers stocks with positive signal growth and selling short the stocks with negative signal growth
for the first three signal measures; and portfolios based on the Short Proportion Growth measure
are constructed conversely. We restrict our attention to ordinary common shares (CRSP share
codes 10 and 11) of U.S.-listed commodity producers as described in Section 3.1.1, and as usual in
our baseline analysis, we follow the Wednesday-to-Tuesday convention.
MA(2) 1-week lag
AW DW VW NW AW DW VW NW
MM Net Change
Ann. Mean excess return (in %) 16.66 28.65 15.52 28.7 19.84 32.16 12.54 22.98
Ann. Standard deviation (in %) 28.88 34.16 25.57 28.62 30.17 34.63 24.87 28.31
t-stat 2.12 3.01 2.24 3.6 2.39 3.33 1.88 2.94
Ann. Sharpe Ratio 0.577 0.839 0.607 1.003 0.657 0.929 0.504 0.812
Cumulative return in $ of $1 invested
5.48 18.88 5.2 23.29 7.5 27.18 3.67 11.57
(August 2006 - September 2018)
MM Long-Short Ratio Growth

t-stat 1.96 2.93 2.25 3.47 2.43 3.37 1.93 2.96
Ann. Sharpe Ratio 0.531 0.814 0.61 0.963 0.668 0.94 0.519 0.818
4.68 16.09 5.17 19.11 7.58 27.7 3.8 11.71
MM Long Proportion Growth

t-stat 1.97 3.02 2.2 3.04 2.31 3.27 1.92 3.1
Ann. Sharpe Ratio 0.535 0.841 0.597 0.841 0.633 0.911 0.515 0.858
4.68 19.33 4.95 13.31 6.68 24.25 3.63 13.26
MM Short Proportion Growth

t-stat 1.87 3.01 2.06 3.72 2.27 3.13 1.95 2.99
Ann. Sharpe Ratio 0.506 0.839 0.558 1.035 0.624 0.874 0.525 0.826
4.36 19.22 4.46 25.46 6.68 22.05 3.9 12.04
Sample period : August 2006 - September 2018
20
4.2. Jensen’s Alpha Analysis and Single Sorting
We have identified a significant mean return differential between the Long and Short stock
portfolios. However, differences in portfolio characteristics and different exposures to risk
factors could explain at least part of the return differential. To investigate the existence of
abnormal returns, we regress the weekly returns of the Long-Short portfolios according to the
Carhart (1997) four-factor model. The model consists of the momentum factor, which was
documented by Jegadeesh and Titman (1993), combined with the Fama and French three-
factor model. The α or Jensen’s alpha is then calculated as the intercept of the regression,
based on the White (1980) standard errors.
Table 3 presents the baseline results for the Carhart four-factor α, together with its
t-statistic. We investigate both the entire sample period (August 2006 to September 2018)
and non-NBER recession periods only (panel B).14 The table compiles results for the four MM
signals and for the total open interest growth, where the signal from the futures market is
either constructed as a 1-week lag or as a J-week moving average, where the look-back horizon
J is equal to 2, 3, 4, 8, or 12 weeks. We look beyond J = 1 since longer look-back horizons
could capture any potential momentum or trend-following effects within the positions taken
by MM traders. These effects have been documented to exist in a number of asset classes
including equity index, currency, commodity, and bond futures (Moskowitz et al., 2012).
In Table 3, for simplicity we follow a strategy of buying the commodity producers’ stocks
with positive signal growth and selling short the stocks with negative signal growth for all
of the four signals, including the Short Proportion Growth signal. Thus, we find that the
alphas are all significantly positive except for the MM Short Proportion Growth case, which
are significantly negative, as expected. Focusing on the J = 2 look-back horizon, across the
four MM sorting measures for the full sample period, the alphas are generally statistically
significant, for example, MM Net Change No-Weight has a t-statistic of 3.57. Specifically, the
statistical significance is robust for all of the four weighting schemes used for all of the four
14 For non-NBER recession periods, we remove the samples from December 2007 to June 2009.
21
MM signal measures used for J = 2. However, for longer look-back horizons, the statistical
significance varies more widely depending on the weighting scheme applied and the signal
measure used, and in some cases the abnormal return can become statistically insignificant.
In addition, the alpha’s magnitude and significance in non-NBER recession periods remain
consistent with the whole sample results. On the other hand, total open interest growth
generally does not contain valuable information for predicting commodity producers’ stock
returns.15
Table C.1 of the appendix replicates the results of Table 3 using the Stambaugh and Yuan
(2016) mispricing factor model that includes, in addition to the factors of market and size,
two composite mispricing factors constructed based upon a set of 11 prominent anomalies.
In particular, MGMT is a composite factor constructed on six anomaly proxies related to
investment and financing, while the second cluster, labeled PERF, is based on five anomaly
variables including return momentum and profitability. The sample period ends in December
2016 due to factors’ availability. This exercise yield estimates of alphas qualitatively and
quantitatively similar to our baseline findings.
In Table C.2 of the appendix, we find that the results presented in this section are robust
even if we expand our scope and include the stocks of U.S.-listed foreign incorporated com-
modity producers (CRSP SHRCD=12) in our analysis. However, the statistical significance
and economic magnitude decrease, consistent with the notion that the subsequent stock
returns of foreign incorporated commodity producers are subject to additional market risk
premium (Foerster and Karolyi, 1999; Karolyi and Stulz, 2003) and to country-specific foreign
shocks; therefore they are more noisy. Standard practice in financial research is to restrict
attention to SHRCD equal to 10 and 11, as in Lyon et al. (1999) and Cohen et al. (2008).
To ensure that our results are not driven by extreme observations, we also conduct a
15 Hong and Yogo (2012) construct a predictor of commodity futures market returns (and of
aggregate returns in currency, bond and stock markets) by taking a 12-month geometric average
in the time series of monthly open interest growth. In contrast, we focus on a much shorter time
horizon. The signal we use to predict the cross-section of commodity producers’ stock returns is
constructed at a weekly frequency and has a look-back horizon of, at most, 12 weeks.
22
Table 3: Baseline Carhart 4-Factor Alpha Results (%, per Week)
Notes: This table presents the Carhart (1997) four-factor alphas of the weekly returns for portfolios
of commodity producing firms sorted by various signal measures (as described in Section 3.1.2)
based on the positions of traders in the MM category (as defined in Table 1) in the commodity
futures market. We restrict our attention to ordinary common shares (CRSP share codes 10 and
11) of U.S.-listed commodity producers as described in Section 3.1.1. The Long-Short portfolios are
rebalanced weekly, and we follow a strategy of buying the producers stocks with positive signal
growth and selling short the stocks with negative signal growth. The signal from the futures market
is constructed as a 1-week lag or as a J-week moving average, where the look-back horizon J is
equal to 2, 3, 4, 8, or 12 weeks. The table report Jensen’s alphas together with t-statistics (based on
White standard errors) reported in parentheses. Panel A shows the results for the regressions on the
whole sample period, from August 2006 to September 2018, while Panel B focuses on non-NBER
recession periods only, without the January 2008-June 2009 period. AW, DW, VW and NW stand
for All-Weights, Degree-Weights, Value-Weights and No-Weight, respectively, as defined in Section
3.2.2. The weekly Jensen’s alphas have been multiplied by 100 so they can be interpreted as
percentages. *** p < 0.01, ** p < 0.05, *p < 0.1.
Money Managers’ Net Change
Panel A: Whole sample Panel B: Without Recession

J AW DW VW NW AW DW VW NW
0.422∗∗ 0.666∗∗∗ 0.284∗∗ 0.476∗∗∗ 0.463∗∗∗ 0.698∗∗∗ 0.281∗∗ 0.479∗∗∗
1
(2.5) (3.43) (2.04) (2.99) (3.01) (3.56) (2.25) (2.97)
0.355∗∗ 0.602∗∗∗ 0.325∗∗ 0.582∗∗∗ 0.403∗∗∗ 0.614∗∗∗ 0.373∗∗∗ 0.572∗∗∗
2
(2.21) (3.15) (2.25) (3.57) (2.61) (3.09) (2.73) (3.46)
0.419∗∗ 0.588∗∗∗ 0.295∗∗ 0.56∗∗∗ 0.428∗∗∗ 0.622∗∗∗ 0.343∗∗ 0.573∗∗∗
3
(2.56) (3.04) (2.09) (3.44) (2.73) (3.15) (2.55) (3.56)
0.36∗∗ 0.556∗∗∗ 0.16 0.33∗∗ 0.446∗∗∗ 0.717∗∗∗ 0.268∗∗ 0.476∗∗∗
4
(2.17) (2.96) (1.13) (2.02) (2.84) (3.7) (2.04) (2.87)
0.305∗ 0.444∗∗ 0.262∗ 0.357∗∗ 0.317∗∗ 0.448∗∗ 0.224∗ 0.284∗
8
(1.76) (2.27) (1.79) (2.02) (2.05) (2.42) (1.66) (1.67)
0.586∗∗∗ 0.733∗∗∗ 0.61∗∗∗ 0.671∗∗∗ 0.511∗∗∗ 0.569∗∗∗ 0.535∗∗∗ 0.478∗∗∗
12
(3.44) (3.66) (4.04) (3.82) (3.34) (3.01) (3.98) (3.03)
Money Managers’ Long Short Ratio Growth

∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗
0.422 0.664 0.288 0.476 0.443 0.683 0.254 0.45∗∗∗
1
(2.57) (3.49) (2.11) (3) (2.89) (3.48) (2.04) (2.79)
0.328∗∗ 0.556∗∗∗ 0.323∗∗ 0.539∗∗∗ 0.371∗∗ 0.614∗∗∗ 0.339∗∗ 0.548∗∗∗
2
(2.05) (3.03) (2.27) (3.43) (2.42) (3.21) (2.54) (3.4)
0.342∗∗ 0.513∗∗∗ 0.317∗∗ 0.551∗∗∗ 0.305∗∗ 0.556∗∗∗ 0.316∗∗ 0.575∗∗∗
3
(2.1) (2.77) (2.22) (3.4) (1.97) (2.88) (2.35) (3.46)
0.299∗ 0.54∗∗∗ 0.211 0.418∗∗∗ 0.344∗∗ 0.64∗∗∗ 0.283∗∗ 0.508∗∗∗
4
(1.79) (2.87) (1.5) (2.59) (2.16) (3.11) (2.12) (2.99)
0.146 0.372∗∗ 0.129 0.346∗∗ 0.272∗ 0.485∗∗ 0.201 0.401∗∗
8
(0.88) (2.02) (0.9) (2) (1.77) (2.52) (1.5) (2.2)
0.156 0.405∗∗ 0.113 0.377∗∗ 0.232 0.375∗ 0.221∗ 0.357∗∗
12
(0.94) (2.13) (0.78) (2.19) (1.58) (1.94) (1.7) (2.04)
23
Money Managers’ Long Proportion Growth

∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗
0.412 0.643 0.285 0.505 0.428 0.661 0.279 0.489∗∗∗
1
(2.43) (3.35) (2.04) (3.15) (2.81) (3.4) (2.26) (3.03)
0.336∗∗ 0.617∗∗∗ 0.317∗∗ 0.509∗∗∗ 0.345∗∗ 0.59∗∗∗ 0.333∗∗ 0.49∗∗∗
2
(2.04) (3.18) (2.23) (3.16) (2.18) (2.94) (2.51) (2.92)
0.389∗∗ 0.597∗∗∗ 0.244∗ 0.506∗∗∗ 0.401∗∗ 0.623∗∗∗ 0.283∗∗ 0.513∗∗∗
3
(2.31) (3.06) (1.75) (3.16) (2.51) (3.1) (2.17) (3.08)
0.335∗∗ 0.578∗∗∗ 0.196 0.366∗∗ 0.419∗∗∗ 0.705∗∗∗ 0.275∗∗ 0.502∗∗∗
4
(1.97) (3) (1.42) (2.33) (2.63) (3.49) (2.14) (3.05)
0.266 0.401∗∗ 0.226∗ 0.443∗∗∗ 0.373∗∗ 0.531∗∗∗ 0.309∗∗ 0.502∗∗∗
8
(1.58) (2.1) (1.66) (2.67) (2.41) (2.77) (2.36) (2.92)
0.433∗∗ 0.674∗∗∗ 0.299∗∗ 0.479∗∗∗ 0.396∗∗∗ 0.511∗∗∗ 0.333∗∗∗ 0.389∗∗
12
(2.46) (3.25) (2.02) (2.73) (2.7) (2.69) (2.69) (2.44)
Money Managers’ Short Proportion Growth

∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗
-0.401 -0.636 -0.292 -0.479 -0.45 -0.694 -0.289 -0.481∗∗∗
1
(-2.37) (-3.25) (-2.11) (-3.02) (-2.89) (-3.51) (-2.32) (-3)
-0.326∗∗ -0.615∗∗∗ -0.305∗∗ -0.596∗∗∗ -0.371∗∗ -0.627∗∗∗ -0.356∗∗∗ -0.59∗∗∗
2
(-1.99) (-3.2) (-2.11) (-3.69) (-2.4) (-3.16) (-2.61) (-3.61)
-0.355∗∗ -0.587∗∗∗ -0.263∗ -0.524∗∗∗ -0.343∗∗ -0.612∗∗∗ -0.33∗∗ -0.578∗∗∗
3
(-2.18) (-3.06) (-1.88) (-3.21) (-2.23) (-3.14) (-2.48) (-3.56)
-0.298∗ -0.573∗∗∗ -0.195 -0.377∗∗ -0.359∗∗ -0.734∗∗∗ -0.292∗∗ -0.522∗∗∗
4
(-1.81) (-2.98) (-1.39) (-2.31) (-2.28) (-3.65) (-2.23) (-3.16)
-0.269 -0.502∗∗ -0.295∗∗ -0.479∗∗∗ -0.26∗ -0.47∗∗ -0.268∗∗ -0.411∗∗
8
(-1.56) (-2.57) (-2.06) (-2.74) (-1.66) (-2.55) (-2.04) (-2.49)
-0.513∗∗∗ -0.748∗∗∗ -0.572∗∗∗ -0.694∗∗∗ -0.455∗∗∗ -0.627∗∗∗ -0.538∗∗∗ -0.604∗∗∗
12
(-2.98) (-3.77) (-3.8) (-4.1) (-2.87) (-3.34) (-4.02) (-4.02)
Open Interest Growth
0.119 0.062 0.064 -0.012 0.121 0.08 -0.016 -0.056
1
(0.83) (0.39) (0.51) (-0.08) (0.98) (0.52) (-0.15) (-0.39)
0.031 0.04 -0.014 -0.035 -0.011 0.038 -0.086 -0.06
2
(0.21) (0.24) (-0.11) (-0.24) (-0.09) (0.24) (-0.74) (-0.43)
-0.064 0.025 0.037 0.054 -0.077 -0.03 -0.017 -0.003
3
(-0.42) (0.14) (0.29) (0.36) (-0.53) (-0.17) (-0.15) (-0.02)
-0.222 -0.134 -0.19 -0.171 -0.277∗ -0.182 -0.252∗ -0.223
4
(-1.42) (-0.77) (-1.38) (-1.1) (-1.88) (-1.02) (-1.9) (-1.42)
-0.1 -0.147 -0.141 -0.26∗ -0.176 -0.142 -0.179 -0.197
8
(-0.66) (-0.85) (-1.05) (-1.71) (-1.27) (-0.82) (-1.47) (-1.29)
0.087 0.291 -0.015 0.118 -0.254∗ -0.062 -0.32∗∗ -0.206
12
(0.54) (1.59) (-0.1) (0.65) (-1.84) (-0.38) (-2.49) (-1.22)
24
robustness exercise in which we winsorize the individual stock’s weekly return at 1st and
99th percentiles, using annually updated cutoffs calculated from our matched sample. Formal
approaches to discriminate outliers include trimming and winsorizing data. For instance,
Knez and Ready (1997) employ a trimmed least squares estimation method and find that
outliers have a large influence on the cross-sectional regressions found in Fama and French
(1993). Specifically, the size factor seems to be due to a small proportion of the small firms
that do extraordinary well. We perform winsorization and present the results in Table C.3
in the appendix, which shows that outliers’ behaviors do not have large influences on our
estimated alphas, thus confirming our baseline findings reported in Table 3.
We next remove stocks in the bottom 25% of market capitalization using annually updated
cutoffs calculated from the CRSP universe. Table C.4 in the appendix shows that our baseline
results are not sensitive to dropping micro-cap stocks.
To summarize, these first findings show that economically large and statistically significant
alphas can be attributed to signals extracted from MM’s commodity futures market positions.
Most importantly, our results are robust to a variety of choices of signal measures, weighting
schemes and timing of lags. The results are also robust relative to the mispricing factor model,
to the removal of micro-cap stocks, and to the winsorizing of extreme return observations.
These results reasonably suggest that MM traders have valuable information that predicts
the returns of commodity-producers’ stocks which is not captured by commonly used factors
of equity returns.
To control for any potential non-linear relation between our signal measures and future
stock returns that is not captured in a regression framework, we also adopt a single-sorting
procedure as part of our analysis, in a way that differs in some respects from the method
previously outlined. Each week, all stocks are sorted into three bins based on the signal’s
value, with bin 1 representing the lowest tercile and bin 3 representing the highest tercile.16 In
16 Incontrast to single-sorting, the Jensen’s alpha estimation presented earlier was based on a
zero-investment strategy by going long (short) on stocks simply according to the sign of the signal
measure, and the strength of the signal measure was not ranked; although the Degree-Weight and
All-Weight weighting schemes do take the strength of the signal into consideration.
25
the next step, we compute the subsequent equal- or value-weighted returns for each portfolio
bin, and the long-short returns of going long on the highest tercile and going short on the
lowest tercile (‘‘3-1’’). The relevance of the signal variable is then assessed by evaluating
Jensen’s alpha relative to the Carhart four-factor model, as well as the Stambaugh and
Yuan (2016) mispricing factor model. Additionally, we use the Fama and French five-factor
model to control for the profitability and asset growth (i.e., investment) factors. With the
single-sorting procedure, we are able to investigate the price impacts of different signal bins
and to check if there is a monotonic relation between signal strength and subsequent portfolio
returns.
Table 4 shows the results for both the equal-weighted (Panel A) and value-weighted
(Panel B) returns for each portfolio bin. First, for all signal measures in both panels, the
mean column confirms the monotonicity of the raw returns over the signal-ranked bins.
Second, the return difference between the two extreme portfolios bins is economically large
and statistically significant at 1% to 5%. The results are qualitatively similar to the findings
reported in Table 3, although they are slightly different in numeric value because for the
single-sort analysis we rank the signals when constructing the portfolio bins, whereas for
the Jensen’s alpha analysis, we follow a simply strategy of buying the producers stocks with
positive signal growth and selling short the stocks with negative signal growth for all of the
four signals, including the Short Proportion Growth signal, for simplicity. The single-sort
results in Table 4 further confirm that our MM signals can predict the commodity producers’
stock returns in the one week following the DCOT report. Furthermore, notice that the sign
of the α for the Short Proportion Growth measure is negative while the sign is positive for
the other three signal measure, as expected.
4.3. Fama-MacBeth Analysis

We run Fama-Macbeth cross-sectional regression to investigate whether—conditional on
controls such as firm size, book-to-market ratio, short-term reversal, momentum and past
26
Table 4: Single Sorting Results
Notes: Each week, all producers’ stocks are sorted into three bins based on the signal value,
with bin 1 representing the lowest tercile and bin 3 representing the highest tercile. The signal
from the futures market is constructed as a 1-week lag. Subsequent equal-weighed (Panel A) or
value-weighted (Panel B) returns for each portfolio bin are computed. The long-short returns of
going long for the highest third portfolio and going short for the lowest third portfolio (‘‘3-1’’) are
also shown. From the long-short returns, we then calculate the abnormal return (α) relative to the
Carhart four-factor model (C4 α), to the Fama and French five-factor model (FF5 α) as well as to
the Stambaugh and Yuan mispricing factor model (SY4 α). The table reports average returns and
differences in returns, together with t-statistics (based on Newey-West adjusted standard errors)
reported in parentheses. The returns have been multiplied by 100 so they can be interpreted as
percentages. *** p < 0.01, ** p < 0.05, *p < 0.1.

Panel A: Equal-Weight Panel B: Value-Weight
rank Mean C4 α FF5 α SY4 α Mean C4 α FF5 α SY4 α
1 -0.198 (-0.88) -0.09 (-0.46)

2 0.165 (0.83) 0.193 (1.19)
3 0.492∗∗ (2.35) 0.422∗∗ (2.20)
0.690∗∗∗ 0.681∗∗∗ 0.714∗∗∗ 0.705∗∗∗ 0.512∗∗∗ 0.524∗∗∗ 0.467∗∗ 0.55∗∗∗
(3-1)
(3.89) (3.81) (3.55) (3.49) (3.16) (3.14) (2.53) (2.86)
Money Managers’ Long-Short Ratio Growth

1 -0.135 (-0.61) -0.065 (-0.33)

2 0.149 (0.73) 0.214 (1.33)
3 0.434∗∗ (2.06) 0.354∗ (1.84)
0.569∗∗∗ 0.566∗∗∗ 0.577∗∗∗ 0.57∗∗∗ 0.419∗∗∗ 0.433∗∗∗ 0.362∗∗ 0.439∗∗
(3-1)
(3.23) (3.19) (2.91) (2.84) (2.62) (2.63) (1.99) (2.31)

1 -0.172 (-0.79) -0.066 (-0.34)

2 0.167 (0.83) 0.205 (1.23)
3 0.473∗∗ (2.21) 0.415∗∗ (2.19)
0.645∗∗∗ 0.638∗∗∗ 0.638∗∗∗ 0.639∗∗∗ 0.482∗∗∗ 0.493∗∗∗ 0.419∗∗ 0.503∗∗∗
(3-1)
(3.75) (3.67) (3.28) (3.23) (3.07) (3.08) (2.35) (2.72)

1 0.457∗∗ (2.16) 0.383∗∗ (2.01)

2 0.196 (0.95) 0.226 (1.31)
3 -0.182 (-0.85) -0.06 (-0.32)
-0.638∗∗∗ -0.636∗∗∗ -0.647∗∗∗ -0.64∗∗∗ -0.443∗∗∗ -0.444∗∗∗ -0.401∗∗ -0.473∗∗
(3-1)
(-3.61) (-3.50) (-3.08) (-3.09) (-2.77) (-2.71) (-2.17) (-2.48)
27
change in commodity spot price—our signal can predict the stock returns of commodity
producers in the week following the DCOT report. The Fama-Macbeth regressions is
performed at the daily frequency, that is, for each of the 5 trading days (the first is a
Wednesday and the last is the Tuesday of next week, unless it is a trading holiday) following
a new DCOT report on a Tuesday, even though the value of the signal does not change for
these five subsequent trading days (although its value will change for the observations in the
next Wednesday-to-Tuesday cycle, as the signal is updated weekly). We then report average
slopes and Newey-West corrected t-statistics with five lags as there are five trading days in a
week in case one is concerned with auto-correlation in firm returns in a week. The signal
from the futures market is constructed as a 1-week lag or as a J-week backward moving
average, where J is equal to 2, 3, 4, 8, or 12. As usual in our baseline analysis, we restrict our
attention to ordinary common shares (CRSP share codes 10 and 11) of U.S.-listed commodity
producers.
The regression specification is as follows:
J
Ri,t+K = αt+K + βSignali,t + γXi,t + i,t+K , (1)
where K can be 1, 2, 3, 4 or 5 trading days ahead of the Tuesday this week which is denoted
time t, and Ri,t+K is the daily return of the stock of firm i in the five trading days (from the
Wednesday of this week to the Tuesday of next week) following a new signal generation of
J
Signali,t at the Tuesday of this week; and Xi,t represents control variables. The coefficient
β in Equation (1) is our focus.
Control variables include ret−1 , which is the stock return over the previous month;
ret−2,−12 , which is the stock return over the 11 months preceding the previous month;
ln(BE/ME) which denotes the log of the ratio of the book value of equity to the market value
of equity; and ln(ME), which denotes the log of the market value of equity. We also include
the change in spot price of the commodity in the previous week as an additional control.
28
Following Henderson et al. (2014), we use the prices of the front-month futures contracts
as proxies for spot prices. If our signals are significant in the presence of this control, then
it would indicate that the positions of MM do have predictive power in addition to the
information already contained in commodity prices.
We present the results using the MM Net Change signal measure in Table 5. We perform
the analysis for the full sample period (August 2006 to Spring 2018), and for the non-NBER
recession sample period. We find that a 1-week lag of our signal in the commodity futures
market can predict future stock returns with a t-statistic of 2.902, and an even higher
t-statistic of 3.676 if we focus only on non-NBER recession periods. We present the full
Fama-Macbeth regression results in Appendix F.
In addition to MM Net Change, in Appendix F we also use MM Long Short Ratio Growth,
MM Long Proportion Growth, and MM Short Proportion Growth as our signals. In each
table, the top panel shows the results using the 1-week lag and the 2- and 3-week moving
averages of our commodity futures market signals, while the lower panel presents the results
using the signals’ 4-, 8-, and 12-week moving averages. Columns (1), (3), (5), (7), (9) and
(11) show the results when the signal alone acts as the regressor. In terms of the values of the
adjusted R2 ’s of the Fama-Macbeth regressions, which are the averages of the cross-sectional
adjusted goodness-of-fit, the values at a little less than 7% are comparable in magnitude as
the values found in recent studies in the finance literature. For example, our R2 compare
favorably to those reported in table II of Tetlock et al. (2008). It should be noted that we
are running the Fama-Macbeth regressions at daily frequency using firm-level daily returns
for each of the subsequent five trading days as dependent variables on weekly-updated
commodity signals among other low-frequency controls, whereas higher R2 are expected for
returns at longer horizons (Fama, 2014).
We find that the coefficients on signals are all positive except in the case of MM short
proportion growth, which are negative as expected. For the 1-week lag and the 2-week
moving averages, the statistical significance of the signal variables varies from the 1% to the
29
Table 5: Fama-Macbeth Regressions: Managed Money Net Change
Notes: This table shows results from Fama-MacBeth cross-sectional regressions (average slopes, and
Newey-West adjusted t-statistics with five lags) of firms’ subsequent daily return on lagged signal
and other lagged controls for expected returns. The daily return of the firm occurs within 7 calendar
days (the first is always a Wednesday and the last is always a Tuesday unless they are postponed
due to holiday) following the newest CFTC Disaggregated Commitments of Traders report. We run
the Fama-Macbeth regression at daily frequency. The signal from the futures market is constructed
as a 1-week lag or as a J-week moving average, where J is equal to 2, 3, 4, 8, or 12. ret−1 is the
stock return over the previous month; ret−2,−12 is the stock return over the 11 months preceding
the previous month; ln(BE/ME) denotes the log of the ratio of book value of equity to market value
of equity; ln(ME) is the log of the market value of equity. We present t-statistics in parentheses.
***, **, and * indicate significance at the 1% 5%, and 10% level, respectively. Adj.R2 reports the
average of the cross-sectional adjusted R2 ’s. N –Companies is the number of unique firms, and
N –Observations is the number of firm-day return observations utilized in the regression.
All Observations lag1 MA(2) MA(3)
(1) (2) (3) (4) (5) (6)
Managed Money +0.002 +0.006∗∗∗ +0.005∗∗∗ +0.004∗∗ +0.005∗∗ +0.004∗
Net Change (+1.637) (+2.902) (+3.034) (+2.325) (+2.246) (+1.832)
ln(BE/M E) +0.000 +0.000 +0.000
(+0.248) (+0.169) (+0.101)
ln(M E) −0.000∗ −0.000∗ −0.000∗
(-1.823) (-1.900) (-1.905)
ret−1 +0.001 +0.001 +0.001
(+0.458) (+0.515) (+0.487)
ret−2,−12 +0.001 +0.001 +0.000
(+1.314) (+1.311) (+1.206)
∆CP rice−1 +0.005 +0.010 +0.009
(+0.753) (+1.569) (+1.384)
N –Companies 199 192 199 191 195 191
N –Observations 303387 288301 302141 287532 301271 286870
Adj.R2 0.06898 0.06915 0.06902
Non-Recession lag1 MA(2) MA(3)
Periods Only
(1) (2) (3) (4) (5) (6)
∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗
Managed Money +0.004 +0.005 +0.004 +0.004 +0.004 +0.005∗∗∗
Net Change (+3.330) (+3.676) (+2.622) (+2.809) (+2.462) (+2.876)
ln(BE/M E) +0.000 −0.000 −0.000
(+0.033) (-0.023) (-0.035)
ln(M E) −0.000 −0.000 −0.000
(-0.592) (-0.661) (-0.650)
ret−1 +0.001 +0.001 +0.001
(+0.590) (+0.656) (+0.614)
ret−2,−12 +0.001 +0.001 +0.001
(+1.610) (+1.595) (+1.491)
∆CP rice−1 +0.005 +0.012∗ +0.011
(+0.673) (+1.724) (+1.533)
N –Companies 198 191 198 190 194 190
N –Observations 259278 245096 258080 244327 257210 243665
Adj.R2 0.06809 0.06815 0.06821
30
5% level, according to columns (1) and (3). However, for the 3-, 4-, 8-, and 12-week moving
averages, our signal can become statistically insignificant. According to columns (2), and
(4), even in the presence of control variables, our signal’s coefficients remain statistically
significant, with statistical significance consistently at the 1% level. For example, during
non-recession periods, the MM long-short ratio growth is a statistically significant predictor
of future firm returns in the following 5 trading days with a t-statistic of 3.559, and the
signal’s coefficient is 0.005. The results from these multivariate regressions suggest that
signals in the commodity futures market can indeed predict subsequent stock returns, and
this effect cannot be absorbed by a set of existing determinants of stock returns.
Furthermore, we have also conducted the Fama-Macbeth regressions while excluding
the previous week’s change in commodity spot price as an control variable. The results are
robust though we do not present these results for brevity.
As another robustness test, we have also conducted the Fama-Macbeth regressions on
samples that first underwent winsorization of stock returns at the 1st and 99th percentile
cutoffs based on the distribution of stock returns in the cross-section in the CRSP universe
in a given year. The results remain qualitatively the same and we again omit the results in
presentation.
4.4. Further Discussions
4.4.1. Jensen’s Alpha Analysis for the Monday-to-Friday Convention
Our baseline results are based on the Wednesday-to-Tuesday convention, which uses the
CFTC report compilation date as the signal generation date and the portfolio formations are
made starting at the beginning of the next trading day (on a Wednesday).
We now apply a different method to aggregate the daily commodity-equity based portfolios’
return series into a weekly frequency, by computing the weekly return from one trading
day after the actual release date through the end of the next release date, referred as the
31
Monday-to-Friday convention.17
Our main objectives with this new convention is to study whether the time lag for
incorporating information contained in MM’s futures markets positions into producers’ equity
prices is longer than what was captured by the Wednesday-to-Tuesday convention and
whether we our predictability results remain even in the Monday-to-Friday convention which
corresponds to actual release schedules although it forgoes the Wednesday, Thursday and
Friday immediately following the compilation of DCOT reports on Tuesday. Figure B.III in
the appendix plots the portfolios’ cumulative weekly returns, signaled by the lag-1 MM Long
Short Ratio Growth, which documents an overall upward trend in the cumulative return of
the Long-Short portfolio over the sample period.
The results shown in Table D indicate that although the strategy based on the CFTC’s
information release schedule can still generate statistically significant returns, the findings
are however more sensitive to the choices of signal measures, weighting schemes, timing
of lags and sample periods. For example, focusing on the ‘‘Without Recession’’ sample, in
which restrict our attention to the non-NBER recession periods by removing the January
2008-June 2009 period, a trading strategy based on the 4-week moving average MM Long
Proportion Growth measure can generate alphas with t-statistics from 2.51 to 2.89 across the
four weighting schemes.
4.4.2. Additional Results
In addition, we investigate whether MM traders are the only smart money in the commodity
futures market by studying whether we can also construct a profitable trading strategy based
on the commodity futures market positions of the PM category, which includes commodity
producers, merchants, processors and users, who are entities that predominantly engages
in the production, processing, packing or handling of a physical commodity and uses the
17 The DCOT reports are usually released on Friday, but federal holidays may delay release by
one or two days. Similarly to our treatment of the compilation dates in Section 3.2.1, we adjust the
convention accordingly by using CFTC’s actual release dates (which is available on CFTC website
for only the last 12 months, but older actual historical release dates can be obtained using archived
cache of the website).
32
futures markets to manage or hedge risks associated with those activities per CFTC. As
Table E.1 indicates, utilizing PM signals instead of MM signals generally does not yield
significant Jensen’s alphas for lag1 and moving averages of the immediate past, however, the
signals calculated using moving averages of the past 8- or 12-weeks do lead to marginally
significant predictability, although the results are less robust than that of the MM category of
traders. Notice that as commercial hedgers, naturally, the signs on these Jensen’s alphas are
flipped when comparing with the results of speculators (MM) in Table 3. To the extent that
PM’s MA(8) signal is statistically significant in the whole sample but not in the non-NBER
recession sample, this can be rationalized by the notion that as hedgers, PM activity in
the futures market, especially to the extent that the PM category captures the activity of
commodity producers who are worried about commodity price drops in recessions, are perhaps
more potent predictors during the recession periods but much less so during normal times.
Overall, the results are consistent with the notion that MM traders are sophisticated
speculators, who are often levered, and who have the most incentive to process and acquire
information related to movements in the fundamentals of the commodity market; whereas
PM are not necessarily as short-term focused as the MM traders. In addition, Cheng and
Xiong (2014) have found that the PM category traders have significantly engaged in trades
for reasons other than hedging, since they frequently change their futures positions for reasons
unrelated to output fluctuations, possibly due to speculation. As hedging versus speculation
trades mandate opposite (long versus short) positions to be taken, this limits the usefulness
of the PM data as a signal. The fact that users, processors, merchants and producers are
grouped together in the DCOT reports further limits the usefulness of the signal as they are
a diverse group with varying incentives to trade in the commodity futures market. All in all,
the positions on the PM category of traders, as a whole, do not contain information that is
robustly conducive towards predictability of equity returns in the short-term.
We omit here the results for the Wednesday-to-Tuesday convention in which we remove
one commodity at a time from our analysis as robustness checks. In general, removing any
33
one commodity do not change our results qualitatively for our baseline analysis. However, we
note here we should not remove too many commodities from our sample as it is a well-known
fact in the empirical literature of commodities, that an empirical phenomenon which may
hold for a diversified portfolio of commodities does not necessarily hold when one restricts
the sample size to only one particular commodity.18
4.5. Double-Sorting Results and Relationship with Market Fric-
tions
Our analysis revealed the existence of sizable and significant abnormal returns for commodity
producing stocks in the equity market based on information extracted from the commodity
futures market. In this section, we examine why this predictability arises. Due to legal and
privacy concerns, the CFTC does not publish information on how individual traders are
classified in the COT reports and information are always aggregated to protect the identity
of any individual reportable trader; though we can still use the double-sort methodology to
focus our investigation on two sets of market frictions that could potentially contribute to
our predictability results, namely, informational frictions and trading frictions.
Informational friction is at the center of a number of consistent findings in the literature
of return anomalies. Firms with higher analyst dispersion have been found by Sadka and
Scherbina (2007) to earn lower subsequent returns, which is believed to be because these
firms have higher information asymmetry. We proxy informational friction based on two
measures, ex ante analyst dispersion and 60-day historical stock volatility.19
18 For example, to quote Fabozzi et al. (2008): ‘‘Thus, while individual (commodity futures)
markets did not appear to provide consistent risk premiums, a portfolio of futures did appear to
provide significant premiums. One explanation for this finding (Fama and French, 1987) is that
diversification across commodities reduces portfolio risk without reducing return.’’ See also, Erb
and Harvey (2005).
19 One can draw an analogy between these two measures and two similar stock volatility measures,
namely, historical volatility and option implied volatility of stocks. These two measures do not
necessarily coincide with each other (Bollerslev et al., 2009). The former is a backward-looking
measure, calculated ex post, based on realized historical data. On the other hand, the later is a
forward-looking measure based on market participants’ expectations of future return variation, like
analyst dispersion.
34
We proxy trading friction with the illiquidity measure proposed by Amihud (2002).
Liquidity captures the ease of trading a security. Exogenous transaction costs, demand
pressure, inventory and search friction risk are all possible sources of illiquidity.
Ultimately, we aim to explore the relationship between market friction and our stock
return predictability results. To do so, we utilize double-sorting as our main methodology.
Specifically, each week, all producers stocks are first sorted into three friction portfolios using
one of the three aforementioned firm-level proxies for market friction, with the requirement
that each commodity appears across those three portfolios. Then, two signal portfolios are
formed dependently within each friction portfolio based on one of the four MM signal variables.
The 3-by-2 double-sorting method produces six portfolios. Finally, we calculate equal- and
value-weighted returns of the sorted portfolios in a similar fashion as the single-sorted
portfolios presented in Section 4.2.
Table 6 presents the results for the returns utilizing the (lag1) MM Net Change signal
measure and double-sorted with the three friction proxies.20 We pay special attention to the
Long-Short portfolio returns and whether the 4- or 5-factor alphas arise in the difference
between the high- and low-friction bins. By double-sorting our commodity futures market
signal with Amihud’s illiquidity measure (LIQ), we find no evidence that predictability is
stronger (or weaker) in firms with higher trading friction. However, we find that our results
are significantly stronger in firms with higher information asymmetry as measured by 60-day
historical stock volatility (VOL) and ex ante analyst dispersion (AD) as compared to the
firms with lower information asymmetry. For instance, if we focus on the Value-Weight
column, the difference in the Long-Short portfolio’s FF5 α between the highest tercile and
lowest tercile of 60-day historical stock volatility (VOL) is in itself a significant difference
with t-statistic of 3.27, and the statistical significance of the differences remain if we use
alternative factor model such as the Carhart’s four-factor model (C4) or the mispricing factor
20 As a technical detail, unlike the single-sorting exercise in which we applied the method of first
build the commodity-equity portfolios from stocks and then compound the daily portfolios returns
into weekly returns for each of the commodity-equity portfolios, in the double-sorting exercise we
first compute weekly returns for each stock and then build the commodity-equity portfolios.
35
Table 6: Double Sorting: Money Managers’ Net Change (%, per Week)
Notes: This table presents results from our double-sorted cross-sectional exercise. Specifically, each
week, all the producer stocks are first sorted into three friction portfolios using one of the three
firm-level proxies of friction (LIQ, VOL, and AD), with the requirement that each commodity
appears across those three portfolios. LIQ, VOL and AD stand for the Amihud’s illiquidity measure,
the 60-day historical stock volatility and the ex ante analyst dispersion, respectively. Then, the
two signal portfolios are formed dependently within each of the three friction portfolios, based
on the MM Net Change signal. The signal from the futures market is constructed with a 1-week
lag, and the same is true for the three proxies of market friction. The results are based on the
Wednesday-to-Tuesday convention, which uses the CFTC report compilation date as the signal
generation date. This 3-by-2 double-sorting procedure produces six portfolios.Finally, we both
equally weight (Panel A) or value-weight (Panel B) the sorts. The significance of the sorting variable
is assessed by calculating the Jensen’s alpha relative to the Carhart (1997) four-factor model (C4 α)
and the Fama and French (2015) five-factor model (FF5 α). The table reports average returns,
together with t-statistics reported in parentheses, and the t-statistics for the Jensen’s alphas are
based on Newey-West adjusted standard errors. The returns and Jensen’s alphas, per week, have
been multiplied by 100. *** p < 0.01, ** p < 0.05, *p < 0.1.

AD 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.269 -0.017 -0.38 -0.649∗∗∗ 0.276 -0.031 -0.425 -0.701∗∗∗
1
(1.40) (-0.09) (-1.40) (-3.38) (1.50) (-0.16) (-1.60) (-3.63)
0.313∗ 0.451∗∗ 0.409 0.096 0.277 0.455∗∗ 0.432∗ 0.156
2
(1.71) (2.27) (1.53) (0.48) (1.65) (2.37) (1.65) (0.79)
0.039 0.465∗∗∗ 0.784∗∗∗ 0.744∗∗∗ 0.001 0.482∗∗∗ 0.878∗∗∗ 0.877∗∗∗
(2-1)
(0.23) (2.97) (3.15) (2.92) (0.01) (3.09) (3.59) (3.51)
0.05 0.443∗∗∗ 0.843∗∗∗ 0.793∗∗∗ 0.046 0.471∗∗∗ 0.936∗∗∗ 0.891∗∗∗
C4 α
(0.29) (2.82) (3.15) (2.89) (0.26) (2.98) (3.56) (3.29)
0.069 0.437∗∗ 0.862∗∗∗ 0.793∗∗∗ 0.037 0.452∗∗∗ 0.96∗∗∗ 0.923∗∗∗
FF5 α
(0.37) (2.55) (2.88) (2.61) (0.20) (2.64) (3.29) (3.09)
0.134 0.48∗∗∗ 0.871∗∗∗ 0.737∗∗ 0.129 0.489∗∗∗ 0.98∗∗∗ 0.851∗∗∗
SY4 α
(0.71) (2.79) (2.96) (2.44) (0.66) (2.85) (3.38) (2.82)
VOL 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.135 0.037 -0.349 -0.483∗∗∗ 0.054 0.038 -0.341 -0.395∗∗
1
(0.87) (0.17) (-1.31) (-2.67) (0.36) (0.17) (-1.27) (-2.15)
0.292∗ 0.347∗ 0.459∗ 0.167 0.258∗ 0.35∗ 0.468∗ 0.21
2
(1.84) (1.73) (1.66) (0.79) (1.67) (1.67) (1.73) (1.03)
0.158 0.31∗∗ 0.808∗∗∗ 0.65∗∗∗ 0.204 0.311∗∗ 0.809∗∗∗ 0.605∗∗
(2-1)
(1.20) (2.00) (3.42) (2.75) (1.61) (2.02) (3.39) (2.58)
0.165 0.329∗∗ 0.809∗∗∗ 0.645∗∗∗ 0.22∗ 0.325∗∗ 0.806∗∗∗ 0.586∗∗
C4 α
(1.22) (2.10) (3.33) (2.64) (1.67) (2.10) (3.31) (2.45)
0.127 0.291∗ 0.984∗∗∗ 0.857∗∗∗ 0.174 0.295∗ 1.061∗∗∗ 0.887∗∗∗
FF5 α
(0.84) (1.68) (3.52) (3.07) (1.20) (1.68) (3.84) (3.27)
0.158 0.323∗ 0.909∗∗∗ 0.751∗∗∗ 0.237 0.344∗ 0.978∗∗∗ 0.741∗∗∗
SY4 α
(1.02) (1.79) (3.30) (2.71) (1.59) (1.90) (3.55) (2.72)
36
LIQ 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
-0.03 -0.036 -0.117 -0.087 -0.041 0.051 -0.183 -0.142
1
(-0.16) (-0.16) (-0.50) (-0.50) (-0.24) (0.23) (-0.77) (-0.79)
0.288 0.391∗ 0.452∗ 0.164 0.281∗ 0.461∗∗ 0.41∗ 0.129
2
(1.62) (1.87) (1.85) (0.93) (1.70) (2.22) (1.67) (0.74)
0.318∗∗ 0.428∗∗ 0.569∗∗ 0.251 0.323∗∗ 0.41∗∗ 0.593∗∗ 0.271
(2-1)
(2.29) (2.44) (2.52) (1.05) (2.34) (2.47) (2.58) (1.13)
0.337∗∗ 0.446∗∗ 0.553∗∗ 0.216 0.355∗∗ 0.427∗∗ 0.573∗∗ 0.218
C4 α
(2.38) (2.55) (2.37) (0.88) (2.53) (2.56) (2.41) (0.88)
0.285∗ 0.491∗∗ 0.634∗∗ 0.349 0.312∗∗ 0.454∗∗ 0.684∗∗∗ 0.372
FF5 α
(1.76) (2.43) (2.45) (1.29) (1.97) (2.35) (2.60) (1.34)
0.329∗∗ 0.486∗∗ 0.6∗∗ 0.271 0.367∗∗ 0.442∗∗ 0.661∗∗ 0.293
SY4 α
(2.06) (2.38) (2.29) (0.98) (2.29) (2.28) (2.44) (1.04)
model (SY4).
Similar results are observed when we use alternative signal measures—MM Long-Short
Ratio Growth, MM Long Proportion Growth, and MM Short Proportion Growth, which are
presented in Appendix G. Based on the results, we posit that the lead-lag relationship is due
to informational friction rather than trading friction such as stock liquidity.
Because the results suggest that predictability is more pronounced in firms with higher
information asymmetry, our results suggest that MM traders may view trading on the
commodity futures market as a first choice instead of dealing with a non-transparent firm in
the equity market, contributing to the time lag of pricing information in the equity market.
In addition to the usual limits-to-arbitrage arguments such as the fact that MM cannot
infinitely lever-up, we can rationalize our finding through a number of angles, all of which
involve asset pricing under imperfect information.
We first note that van Nieuwerburgh and Veldkamp (2010) have developed a model
assuming information processing and information acquisition in the financial market is costly,
as at a minimum it costs time; and their results suggest it’s optimal for agents to be under-
diversified, and to hold one well-diversified portfolio plus one specialized portfolio which
contains only one or very few (specialized) assets, and that agents cannot learn everything
37
about all risks which would be too costly. In other words, we posit that MM may find
devoting resources to learn about the company fundamentals of the commodity producing
firms, many of which can influence company stock price in addition to the general prospects
of the commodity, of each of stocks in the commodity-equity based portfolio to be too costly
(especially those stocks with higher information frictions); and thus MM do not trade on these
stocks in the equity market, which leads to the stronger predictability results in subsequent
stock returns especially amongst the stocks that have higher information asymmetry, leading
to price inefficiency in these stocks at least in the short-term.
In addition, our empirical results are in line with the theoretical model developed by
Sockin and Xiong (2015), who find that market participants of the commodity market face
severe informational frictions (i.e., market participants do not have perfect information on
supply and demand shocks affecting a commodity) and emphasized the informational role
and feedback effect of commodity prices, in that a higher commodity price signals stronger
economic growth and may motivate goods producers to produce more goods which leads to
greater demand for the commodity as inputs. Thus, in our context, if the market participants
believe MM’s increased positions in a certain commodity signal stronger growth in economic
demand requiring that commodity, that is good news for the commodity producers.
The lead-lag results in the equity and futures market can also be rationalized by a model
involving Bayesian learning, through noisy signals, of correlated fundamentals that are
not directly observable (Ho, 2013). Specifically, the fundamental of a commodity and the
fundamental of a commodity-producing firm are clearly correlated, though the fundamental
of the commodity-producing firm has components other than commodity prices as at a
minimum it involves its capital structure, local cost of production, CEO ability among other
things. If an equity investor sees an increased MM position, which is a noisy signal but
nevertheless contains some information about the fundamental of a commodity, she will now
have a more optimistic posterior probability on the fundamental of the firm, whether or not
the signal is really caused by a true change in the fundamental of the commodity. In other
38
words, there can be a contagion effect from the futures market to the equity market, and
Bayesian learning can generate comovements beyond what is mandated by true innovation in
fundamentals (Ho, 2013). To put differently, through learning, the equity investor will have
beliefs correlated with that of MM, similar to the mechanism in Simonovska et al. (2016).
5. Conclusion
Using the CFTC Disaggregated Commitments of Traders (DCOT) reports from August 2006
to September 2018, we document a new empirical phenomenon in which the positions of
managed money (MM) in the commodity futures market can predict the stock returns of
commodity producers in subsequent weeks. This predictability is established by a number
of methodologies including single-sort and Fama-Macbeth cross-sectional regressions across
a range of measures and time-lags. Specifically, if the DCOT reports an increase in long
position, or a decrease in short position, or an increase in net position, or an increase in
the ratio of long over short position of MM, then the stock price of producers of the same
commodity would increase in the following week. Across the different measures, 1-week lag
and 2-week moving averages usually produce the best predictability results.
Since money managers, speculators and hedge funds can trade on the stock market as well,
why do we find predictability results? We argue that predictability arises due to informational
frictions. Our double-sorting exercises indicate that our results are more pronounced with
firms with higher ex ante analyst dispersion and higher historical stock volatility. We do not
find predictability arises as a result of trading friction, as indicated in our double-sorting
results with regards to Amihud’s illiquidity measure. Our findings provide more empirical
evidence to the literature on investor specialization, market segmentation, informational
friction and gradual information diffusion.
Following the discussant’s insights,21 as a next step, we will further disentangle the
information environment and conduct analysis separately on the Wednesday-Thursday-
21 We thank Brian Henderson (discussant) for the comments as well as the participants at AFA
2019.
39
Friday interval and the Monday-and-Tuesday interval, for contrast. Since we already know
that the Monday-to-Friday results are weaker than the Wednesday-to-Tuesday results, we
thus do not believe the predictability to be driven by the announcement of CFTC reports,
and that the predictability seems to be already present prior to the CFTC’s Friday release of
trader positions as of Tuesday.
40
References
Amihud, Yakov. (2002). ‘Illiquidity and stock returns: cross-section and time-series effects’,
Journal of Financial Markets 5(1), 31--56.
Ammer, John, Vega, Clara and Wongswan, Jon. (2010). ‘International transmission of
US monetary policy shocks: Evidence from stock prices’, Journal of Money, Credit and
Banking 42(s1), 179--198.
Baker, Malcolm and Wurgler, Jeffrey. (2007). ‘Investor sentiment in the stock market’,
Journal of Economic Perspectives 21(2), 129--152.
Basak, Suleyman and Pavlova, Anna. (2016). ‘A model of financialization of commodities’,

Journal of Finance 71(4), 1511--1556.
Bessembinder, Hendrik and Seguin, Paul J. (1992). ‘Futures-trading activity and stock price
volatility’, Journal of Finance 47(5), 2015--2034.
Bollerslev, Tim, Tauchen, George and Zhou, Hao. (2009). ‘Expected stock returns and
variance risk premia’, Review of Financial Studies 22(11), 4463--4492.
Buchanan, William K, Hodges, Paul and Theis, John. (2001). ‘Which way the natural gas
price: an attempt to predict the direction of natural gas spot price movements using trader
positions’, Energy economics 23(3), 279--293.
Büyükşahin, Bahattin and Robe, Michel A. (2014). ‘Speculators, commodities and cross-
market linkages’, Journal of International Money and Finance 42, 38--70.
Carhart, Mark M. (1997). ‘On persistence in mutual fund performance’, Journal of Finance
52(1), 57--82.
Carter, Colin A, Rausser, Gordon C and Schmitz, Andrew. (1983). ‘Efficient asset portfolios
and the theory of normal backwardation’, Journal of Political Economy 91(2), 319--331.
41
Cheng, Ing-Haw, Kirilenko, Andrei and Xiong, Wei. (2014). ‘Convective risk flows in
commodity futures markets’, Review of Finance 19(5), 1733--1781.
Cheng, Ing-Haw and Xiong, Wei. (2014). ‘Financialization of commodity markets’, Annual
Review of Financial Economics 6(1), 419--441.
Cohen, Lauren, Frazzini, Andrea and Malloy, Christopher. (2008). ‘The small world of
investing: Board connections and mutual fund returns’, Journal of Political Economy
116(5), 951--979.
De Roon, Frans A, Nijman, Theo E and Veld, Chris. (2000). ‘Hedging pressure effects in
futures markets’, Journal of Finance 55(3), 1437--1456.
Erb, Claude B and Harvey, Campbell R. (2005), The tactical and strategic value of commodity
futures, Technical report, National Bureau of Economic Research.
Fabozzi, Frank J, Fuss, Roland and Kaiser, Dieter G. (2008), The handbook of commodity
investing, Vol. 156, John Wiley & Sons.
Fama, Eugene F. (2014). ‘Two pillars of asset pricing’, American Economic Review
104(6), 1467--85.
Fama, Eugene F and French, Kenneth R. (1987). ‘Commodity Futures Prices: Some
Evidence On Forecast Power, Premiums, and the Theory of Storage’, The Journal of
Business 60(1), 55--73.
Fama, Eugene F and French, Kenneth R. (1993). ‘Common risk factors in the returns on
stocks and bonds’, Journal of Financial Economics 33(1), 3--56.
Fama, Eugene F. and French, Kenneth R. (2015). ‘A five-factor asset pricing model’, Journal
of Financial Economics 116(1), 1--22.
Fernandez-Perez, Adrian, Fuertes, Ana-Maria and Miffre, Joelle. (2017). ‘Commodity Markets,
Long-Run Predictability, and Intertemporal Pricing’, Review of Finance 21(3), 1159--1188.
42
Foerster, Stephen R and Karolyi, G Andrew. (1999). ‘The effects of market segmentation
and investor recognition on asset prices: Evidence from foreign stocks listing in the United
States’, Journal of Finance 54(3), 981--1013.
Gorton, Gary B, Hayashi, Fumio and Rouwenhorst, K Geert. (2013). ‘The fundamentals of
commodity futures returns’, Review of Finance pp. 35--105.
Gorton, Gary and Rouwenhorst, K Geert. (2006). ‘Facts and fantasies about commodity
futures’, Financial Analysts Journal 62(2), 47--68.
Gross, Christian. (2017). ‘Examining the Common Dynamics of Commodity Futures Prices’,
CQE Working Papers pp. Center for Quantitative Economics (CQE), University of Muen-
ster.
Grossman, Sanford J and Stiglitz, Joseph E. (1980). ‘On the impossibility of informationally
efficient markets’, American Economic Review 70(3), 393--408.
Hamilton, James D and Wu, Jing Cynthia. (2015). ‘Effects of Index-Fund Investing On
Commodity Futures Prices’, International Economic Review 56(1), 187--205.
Han, Bing. (2007). ‘Investor sentiment and option prices’, Review of Financial Studies
21(1), 387--414.
Henderson, Brian J, Pearson, Neil D and Wang, Li. (2014). ‘New evidence on the financial-
ization of commodity markets’, Review of Financial Studies 28(5), 1285--1311.
Ho, Steven Wei. (2013). ‘Learning in International Markets and a Rational Expectation
Approach to the Contagion Puzzle’, Working Paper .
Hong, Harrison and Stein, Jeremy C. (1999). ‘A unified theory of underreaction, momentum
trading, and overreaction in asset markets’, Journal of Finance 54(6), 2143--2184.
Hong, Harrison, Torous, Walter and Valkanov, Rossen. (2007). ‘Do industries lead stock
markets?’, Journal of Financial Economics 83(2), 367--396.
43
Hong, Harrison and Yogo, Motohiro. (2012). ‘What does futures market interest tell us about
the macroeconomy and asset prices?’, Journal of Financial Economics 105(3), 473--490.
Hutt, Rosamond. (2015). ‘Why have commodities crashed?’, World Economic Forum .
Jegadeesh, Narasimhan and Titman, Sheridan. (1993). ‘Returns to buying winners and selling
losers: Implications for stock market efficiency’, Journal of Finance 48(1), 65--91.
Karolyi, G Andrew and Stulz, René M. (2003). ‘Are financial assets priced locally or globally?’,
Handbook of the Economics of Finance 1, 975--1020.
Knez, Peter J and Ready, Mark J. (1997). ‘On the robustness of size and book-to-market in
cross-sectional regressions’, Journal of Finance 52(4), 1355--1382.
Lyon, John D, Barber, Brad M and Tsai, Chih-Ling. (1999). ‘Improved methods for tests of
long-run abnormal stock returns’, Journal of Finance 54(1), 165--201.
Menzly, Lior and Ozbas, Oguzhan. (2010). ‘Market segmentation and cross-predictability of
returns’, Journal of Finance 65(4), 1555--1580.
Moskowitz, Tobias J, Ooi, Yao Hua and Pedersen, Lasse Heje. (2012). ‘Time series momen-
tum’, Journal of Financial Economics 104(2), 228--250.
Sada, Maria. (2016). ‘Ripple Effects of China’s Slowdown on Latin America’, Harvard
International Review 37(2), 12.
Sadka, Ronnie and Scherbina, Anna. (2007). ‘Analyst disagreement, mispricing, and liquidity’,
Sanders, Dwight R, Boris, Keith and Manfredo, Mark. (2004). ‘Hedgers, funds, and small
speculators in the energy futures markets: an analysis of the CFTC’s Commitments of
Traders reports’, Energy Economics 26(3), 425--445.
44
Sanders, Dwight R, Irwin, Scott H and Merrin, Robert P. (2009). ‘Smart money: The
forecasting ability of CFTC large traders in agricultural futures markets’, Journal of
Agricultural and Resource Economics pp. 276--296.
Simonovska, Ina, David, Joel et al. (2016), Correlated Beliefs, Returns, and Stock Market
Volatility, in ‘2016 Meeting Papers’, number 130, Society for Economic Dynamics.
Singleton, Kenneth J. (2013). ‘Investor flows and the 2008 boom/bust in oil prices’, Manage-
ment Science 60(2), 300--318.
Sockin, Michael and Xiong, Wei. (2015). ‘Informational frictions and commodity markets’,
Stambaugh, Robert F, Yu, Jianfeng and Yuan, Yu. (2012). ‘The short of it: Investor
sentiment and anomalies’, Journal of Financial Economics 104(2), 288--302.
Stambaugh, Robert F and Yuan, Yu. (2016). ‘Mispricing factors’, Review of Financial Studies
30(4), 1270--1315.
Tetlock, Paul C, Saar-Tsechansky, Maytal and Macskassy, Sofus. (2008). ‘More than words:
Quantifying language to measure firms’ fundamentals’, Journal of Finance 63(3), 1437--
1467.
Tornell, Aaron and Yuan, Chunming. (2012). ‘Speculation and hedging in the currency
futures markets: Are they informative to the spot exchange rates’, Journal of Futures
Markets 32(2), 122--151.
Turner, Matthew. (2009). ‘The Commitment of Traders Report and its usefulness’, The
London Bullion Market Association .
van Nieuwerburgh, Stijn and Veldkamp, Laura. (2010). ‘Information acquisition and under-
diversification’, Review of Economic Studies 77(2), 779--805.
45
Wang, Changyun. (2003). ‘Investor sentiment, market timing, and futures returns’, Applied
Financial Economics 13(12), 891--898.
White, Halbert. (1980). ‘A heteroskedasticity-consistent covariance matrix estimator and a

direct test for heteroskedasticity’, Econometrica pp. 817--838.
46
Appendices
A. Procedure to compute the Long-Short portfolio returns
This appendix supplements the description of the procedure to compute the Long-Short
portfolio returns in Section 3.2.
First, we compute the daily stock returns for each of the 10 commodity-equity portfolios:
1 X
rtC = P V
WiV ritC (2)
i∈C Wi i∈C
where ritC is the stock return at day t of producer i belonging to commodity C, and
(
marketcapi,year−1 if V-Weight is applied
WiV =
1 if E-Weight is applied
Then, we compound rtC to obtain weekly returns Rw C

, where the week w in the Wednesday-
Tuesday convention runs from the beginning-of-day of Wednesday (t = 1) till the end-of-day
of next Tuesday (t = T ):
YT
C
Rw = (1 + rtC ) − 1 (3)
t=1
For a signal with a J-week look-back horizon, J ≥ 1,Pthe weekly signal s of the futures
market are aggregated over the look-back horizon as: J1 Jk=1 sC w−k . If the signal is positive,
then the commodity-equity portfolio C belongs to the Long portfolio (L) in week w. If the
signal is negative, then it belongs to the Short portfolio (S).
L S LS
Finally, we compute the Long (Rw ), Short (Rw ) and Long-Short (Rw ) portfolio returns
at week w, as follows:
L 1 X
Rw =P D
WCD RwC
(4)
C∈L W C C∈L
S 1 X
Rw =P D
WCD Rw
C
(5)
C∈S WC C∈S
LS L S
Rw = Rw − Rw (6)
where,
( P
| 1 J sC | if D-Weight is applied
WCD = J k=1 w−k
1 if N-Weight is applied
In other words, when applying D-Weight, we take into account the strength of the signal.
47
B. Data
Table B.1: Sample of Commodity Producers

Notes: The number of companies and the four-digit SIC codes (plus, whenever we hand-pick a
firm in addition to SIC code matching, its five-digit permno) reported in the table correspond
to the cases where we select ordinary common stocks with CRSP SHRCD=(10 or 11) or CRSP
SHRCD=(10, 11 or 12). 1 Oil & Gas comprises crude oil [067651] and natural gas [23651], weighted
yearly by the lagged U.S. oil and gas industry’s total revenue (data retrieved from the U.S. Energy
Information Administration) 2 Petroleum Refining comprises unleaded gas [111659]. 3 Miscellaneous
metals comprises platinum [076651], and palladium [075651], weighted equally.
Number of stocks where SHRCD:

(10 or 11) (10, 11 or 12)
Commodity Matching SIC codes SIC description Avg. per Total Avg. per Total
[Contract Code] (and permno) week week
Copper 1020; 1021; 3331 Copper Ores; 3 4 7 12

[85692] (91418) Primary Smelting
and Refining of
Copper
Oil & Gas1 1310; 1311 Crude Petroleum 68 122 83 149
and Gas Extraction
Petroleum 2910; 2911 Petroleum refining 11 16 13 19
Refining2 Extraction
Gold 1041, 1040 Gold ores; Gold and 5 9 39 72
[88691] silver ores
Silver 1044; 1040 Silver Ores 3 4 10 14
[84691]
Biofuel 2869; 2860 Industrial organic 10 16 11 19
[25601] chemicals
Steel 3312 (37402, 80375) Steel works, blast 11 19 13 22
[192651] furnaces and rolling
mills
Lumber (21186, 39917, 56143, n.a. 4 6 4 6
[58643] 56223, 79878, 84365)
MiscMetal3 1099 Miscellaneous Metal 4 5 5 8
Ores
Coal 1220; 1221; 1222 Bituminous Coal 8 16 8 16
[24651,24658] and Lignite mining
Total = 127 = 217 = 193 = 337
Sample period : August 2006 - September 2018
48
Table B.2: Summary of positions held by DCOT’s traders categories, August 2006-September 2018
Notes: The table summarizes the positions of traders in commodity futures markets according to the classification employed in the
Disaggregated Commitments of Traders (DCOT) reports. For each category of traders, as described in Table 1, positions are measured as
Net Long (long position - short position)and scaled by the open interest of that trader category. The columns report the sample average
position, the standard deviation of the position, the fraction of the weeks the position is long, and the first order autocorrelation (ρ) of the
position. The end of the sample period is September 2018 for all commodities. The starting date of the sample period is indicated in
parenthesis below each commodity. While the first month of the sample period for coal is June 2007, the series shows many missing gaps,
but becomes continuous from August 2012 onwards. Open interests extracted from CFTC on steel show similar time gaps. The average
and standard deviation of the position have been multiplied by 100 so that they can be interpreted as percentages.
Net Long position of traders as percent of open interest

Producers,
Money Managers Merchants, Swap Dealers Other Reporting Non-Reporting
Processors & Users
Commodity Std. Long Std. Long Std. Long Std. Long Std. Long
Avg ρ Avg ρ Avg ρ Avg ρ Avg ρ
(Start) dev. (%) dev. (%) dev. (%) dev. (%) dev. (%)
Copper
5.5 27.8 55.4 0.96 -47.1 22.3 0.6 0.96 57.8 14.1 100 0.96 -17 17.4 16.1 0.94 -6.1 13.3 33.8 0.92
(2006m8)
49
Steel
5.4 66 53.7 0.98 1.9 49.2 49.1 0.98 85.5 30.5 95.6 0.98 -19.7 12.5 7.7 0.96 47.7 55.2 82.8 0.85
(2012m12)
Gold
45.3 24.7 95.7 0.97 -53.2 19 1.4 0.97 -23.5 15.3 7.7 0.96 25.3 12.4 96.2 0.93 30.8 19 93.4 0.95
(2006m8)
Silver
30.6 21.2 91 0.95 -59.8 14.4 0 0.97 -3.9 22 43.1 0.98 24 14.7 98.4 0.94 35.5 12.1 100 0.93
(2006m8)
MiscMetal
57.8 25.6 96.5 0.98 -74 13 0 0.97 -30.6 21.2 9.9 0.95 44.4 14.9 98.7 0.89 41.6 16.6 99.8 0.93
(2006m8)
Biofuel
51 37.4 91.8 0.92 -16.6 24.3 22.6 0.97 -20 40.8 29.4 0.90 15 30.7 67.9 0.92 20.8 18.2 86.6 0.79
(2009m11)
OilGas
9.4 9.7 83.9 0.98 -18 11.2 5.5 0.99 -3.1 18.7 50.6 1 0.3 8.3 59.1 0.97 12.9 6.6 98.9 0.86
(2006m8)
Petroleum
39 19.2 97.3 0.96 -35.2 6.3 0 0.89 49.7 14.5 100 0.97 17.7 15.8 85.2 0.95 15.1 12.6 87.1 0.89
(2006m8)
Coal
0.6 56.5 48.2 0.96 -9.6 33.1 35.2 0.99 -38.9 36.7 15 0.97 30.1 45.7 66.4 0.99 5.5 28.2 58.2 0.90
(2007m6)
Lumber
7.8 40.5 52.5 0.97 -43.1 37.9 12.8 0.96 97.5 5.6 100 0.86 -2. 18 41.8 0.92 4.1 12 62 0.90
(2006m8)
Figure B.I: Breakdown of Total Open Interest (TOI) by Trader Category
Notes: This figure shows the average market share, for the commodities considered in our analysis,
held by each of the five trader categories in the commodity futures market over the whole sample
period from August 2006 to September 2018. The definitions of the trader categories are described
in Table 1.
1
.8
Share of open interest by categories
.6
.4
.2
0
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018
Money Managers Producers, Merchants, Processors and Users Swap Dealers

Other Reporting Non-Reporting
Figure B.II: Total number of stocks traded each week, baseline analysis
Notes: This figure shows the total number stocks traded in each week from August 2006 to September
2018, summed over all commodities with valid data (i.e., those reporting non-missing positions data)
in CFTC reports. In our baseline analysis, we restrict our attention to ordinary common shares
(CRSP share codes 10 and 11) of U.S.-listed commodity producers as described in Section 3.1.1.
180
160
140
Total Number of Companies
120
100
80
60
40
20
0
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018
50
Figure B.III: Cumulative Gains from Investments, Signaled by the 1-week lagged Managed
Money Long Short Ratio Growth (Monday-to-Friday convention, No-Weight). The shaded
area corresponds to the NBER recession periods.
12 12
8 8
4 4
0 0
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018
51
C. Robustness of Table 3
Table C.1: Stambaugh and Yu’s 4-Factor Alpha Results (%, per Week)
Notes: We estimate the Stambaugh and Yuan (2016) four-factor alphas of the weekly returns for
portfolios of commodity producing firms sorted by various signal measures (as described in Section
3.1.2) based on the positions of traders in the MM category (as defined in Table 1) in the commodity
11) of U.S.-listed commodity producers as described in Section 3.1.1. The Long-Short portfolios are
rebalanced weekly, and we follow a strategy of buying the producers stocks with positive signal
growth and selling short the stocks with negative signal growth. The results are based on the
Wednesday-to-Tuesday convention, which uses the CFTC report compilation date as the signal
generation date. The signal from the futures market is constructed as a 1-week lag or as a J-week
moving average, where the look-back horizon J is equal to 2, 3, 4, 8, or 12 weeks. The table report
Jensen’s alphas together with t-statistics (based on White standard errors) reported in parentheses.
Panel A shows the results for the regressions on the whole sample period, which extends from
August 2006 to December 2016 due to availability of the factor data, while Panel B focuses on
non-NBER recession periods only, without the January 2008-June 2009 period. AW, DW, VW
and NW stand for All-Weights, Degree-Weights, Value-Weights and No-Weight, respectively, as
defined in Section 3.2.2. The weekly Jensen’s alphas have been multiplied by 100 so they can be
interpreted as percentages. *** p < 0.01, ** p < 0.05, *p < 0.1.

∗∗ ∗∗∗ ∗ ∗∗ ∗∗∗ ∗∗∗ ∗
0.445 0.628 0.297 0.441 0.451 0.678 0.266 0.457∗∗
1
(2.28) (2.89) (1.83) (2.46) (2.62) (3.05) (1.9) (2.49)
0.431∗∗ 0.642∗∗∗ 0.383∗∗ 0.637∗∗∗ 0.484∗∗∗ 0.706∗∗∗ 0.434∗∗∗ 0.648∗∗∗
2
(2.35) (3.01) (2.34) (3.53) (2.83) (3.14) (2.92) (3.56)
0.484∗∗∗ 0.631∗∗∗ 0.327∗∗ 0.608∗∗∗ 0.491∗∗∗ 0.724∗∗∗ 0.377∗∗ 0.666∗∗∗
3
(2.6) (2.89) (2.04) (3.29) (2.85) (3.26) (2.57) (3.67)
0.423∗∗ 0.593∗∗∗ 0.174 0.343∗ 0.537∗∗∗ 0.827∗∗∗ 0.312∗∗ 0.564∗∗∗
4
(2.25) (2.8) (1.1) (1.88) (3.04) (3.74) (2.17) (2.96)
0.24 0.378∗ 0.207 0.335∗ 0.273 0.436∗∗ 0.178 0.301
8
(1.26) (1.78) (1.34) (1.78) (1.59) (2.06) (1.25) (1.58)
0.55∗∗∗ 0.708∗∗∗ 0.555∗∗∗ 0.645∗∗∗ 0.537∗∗∗ 0.617∗∗∗ 0.512∗∗∗ 0.505∗∗∗
12
(2.95) (3.23) (3.37) (3.36) (3.17) (2.86) (3.48) (2.82)

0.446∗∗ 0.624∗∗∗ 0.308∗ 0.444∗∗ 0.438∗∗ 0.665∗∗∗ 0.246∗ 0.432∗∗
1
(2.35) (2.92) (1.92) (2.48) (2.56) (3) (1.76) (2.35)
0.407∗∗ 0.602∗∗∗ 0.388∗∗ 0.582∗∗∗ 0.457∗∗∗ 0.715∗∗∗ 0.402∗∗∗ 0.613∗∗∗
2
(2.23) (2.94) (2.41) (3.33) (2.66) (3.33) (2.76) (3.46)
0.405∗∗ 0.557∗∗∗ 0.359∗∗ 0.595∗∗∗ 0.362∗∗ 0.66∗∗∗ 0.351∗∗ 0.661∗∗∗
3
(2.19) (2.66) (2.22) (3.25) (2.11) (3.04) (2.38) (3.57)
0.362∗ 0.627∗∗∗ 0.236 0.438∗∗ 0.427∗∗ 0.781∗∗∗ 0.344∗∗ 0.59∗∗∗
4
(1.93) (2.94) (1.52) (2.42) (2.43) (3.33) (2.34) (3.02)
0.08 0.348∗ 0.091 0.356∗ 0.229 0.456∗∗ 0.184 0.435∗∗
8
(0.42) (1.66) (0.57) (1.8) (1.32) (2.05) (1.22) (2)
0.123 0.456∗∗ 0.094 0.422∗∗ 0.229 0.402∗ 0.233
12 0.4∗∗ (2)
(0.65) (2.11) (0.57) (2.15) (1.37) (1.83) (1.59)
52
∗∗ ∗∗∗ ∗ ∗∗ ∗∗ ∗∗∗ ∗
0.44 0.612 0.291 0.466 0.428 0.642 0.246 0.445∗∗
1
(2.25) (2.85) (1.77) (2.54) (2.53) (2.93) (1.78) (2.42)
0.443∗∗ 0.707∗∗∗ 0.379∗∗ 0.561∗∗∗ 0.456∗∗∗ 0.716∗∗∗ 0.395∗∗∗ 0.564∗∗∗
2
(2.37) (3.28) (2.38) (3.17) (2.6) (3.18) (2.75) (3.08)
0.46∗∗ 0.674∗∗∗ 0.277∗ 0.582∗∗∗ 0.479∗∗∗ 0.756∗∗∗ 0.323∗∗ 0.623∗∗∗
3
(2.41) (3.1) (1.76) (3.31) (2.71) (3.35) (2.25) (3.39)
0.418∗∗ 0.691∗∗∗ 0.258∗ 0.444∗∗ 0.527∗∗∗ 0.862∗∗∗ 0.366∗∗ 0.635∗∗∗
4
(2.16) (3.18) (1.65) (2.52) (2.94) (3.68) (2.55) (3.35)
0.227 0.405∗ 0.181 0.465∗∗ 0.342∗∗ 0.531∗∗ 0.264∗ 0.528∗∗∗
8
(1.19) (1.85) (1.21) (2.44) (1.97) (2.4) (1.85) (2.62)
0.4∗∗ 0.722∗∗∗ 0.253 0.501∗∗ 0.391∗∗ 0.531∗∗ 0.319∗∗ 0.408∗∗
12
(2.09) (3.13) (1.56) (2.55) (2.34) (2.48) (2.3) (2.26)

-0.42∗∗ -0.585∗∗∗ -0.308∗ -0.449∗∗ -0.431∗∗ -0.668∗∗∗ -0.277∗∗ -0.469∗∗
1
(-2.14) (-2.65) (-1.9) (-2.51) (-2.46) (-2.96) (-1.99) (-2.57)
-0.385∗∗ -0.631∗∗∗ -0.362∗∗ -0.627∗∗∗ -0.431∗∗ -0.698∗∗∗ -0.417∗∗∗ -0.64∗∗∗
2
(-2.06) (-2.91) (-2.2) (-3.47) (-2.51) (-3.09) (-2.81) (-3.51)
-0.403∗∗ -0.604∗∗∗ -0.291∗ -0.55∗∗∗ -0.39∗∗ -0.69∗∗∗ -0.367∗∗ -0.65∗∗∗
3
(-2.17) (-2.77) (-1.83) (-2.94) (-2.29) (-3.12) (-2.52) (-3.53)
-0.347∗ -0.598∗∗∗ -0.204 -0.362∗∗ -0.437∗∗ -0.848∗∗∗ -0.332∗∗ -0.587∗∗∗
4
(-1.87) (-2.77) (-1.31) (-1.99) (-2.48) (-3.75) (-2.32) (-3.13)
-0.193 -0.459∗∗ -0.224 -0.448∗∗ -0.233 -0.485∗∗ -0.223 -0.43∗∗
8
(-1.05) (-2.17) (-1.53) (-2.44) (-1.37) (-2.32) (-1.62) (-2.33)
-0.433∗∗ -0.707∗∗∗ -0.493∗∗∗ -0.672∗∗∗ -0.418∗∗ -0.668∗∗∗ -0.493∗∗∗ -0.655∗∗∗
12
(-2.33) (-3.23) (-3) (-3.64) (-2.46) (-3.11) (-3.35) (-3.82)

0.106 -0.025 0.076 -0.079 0.104 -0.033 -0.027 -0.148
1
(0.65) (-0.14) (0.54) (-0.48) (0.75) (-0.19) (-0.22) (-0.91)
-0.012 -0.042 -0.023 -0.085 -0.023 -0.016 -0.064 -0.066
2
(-0.07) (-0.22) (-0.15) (-0.49) (-0.16) (-0.09) (-0.5) (-0.42)
-0.11 -0.016 0.026 0.054 -0.125 -0.109 -0.043 -0.04
3
(-0.65) (-0.08) (0.18) (0.31) (-0.81) (-0.56) (-0.35) (-0.25)
-0.241 -0.198 -0.168 -0.168 -0.289∗ -0.256 -0.219 -0.231
4
(-1.41) (-1.02) (-1.13) (-0.99) (-1.86) (-1.29) (-1.59) (-1.36)
-0.061 -0.196 -0.09 -0.291∗ -0.143 -0.237 -0.122 -0.241
8
(-0.35) (-0.98) (-0.58) (-1.69) (-0.9) (-1.16) (-0.89) (-1.38)
0.144 0.385∗ 0.015 0.21 -0.251 -0.066 -0.328∗∗ -0.182
12
(0.79) (1.85) (0.09) (1.03) (-1.64) (-0.35) (-2.3) (-0.95)
53
Table C.2: Carhart 4-Factor Alpha Results with CRSP Share Codes 10, 11 & 12 (%, per
Week)
futures market. We include in our sample ordinary common shares (CRSP share codes 10 and
11), as well as U.S.-listed foreign incorporated commodity producers (CRSP share code 12). The
Long-Short portfolios are rebalanced weekly, and we follow a strategy of buying the producers stocks
with positive signal growth and selling short the stocks with negative signal growth. The results are
based on the Wednesday-to-Tuesday convention, which uses the CFTC report compilation date as
the signal generation date. The signal from the futures market is constructed as a 1-week lag or as
a J-week moving average, where the look-back horizon J is equal to 2, 3, 4, 8, or 12 weeks. The
table reports Jensen’s alphas together with t-statistics (based on White standard errors) reported in
parentheses. Panel A shows the results for the regressions on the entire sample period, i.e., from
August 2006 to September 2018, while Panel B focuses on non-NBER recession periods only, i.e.,
without the January 2008 to June 2009 period. AW, DW, VW and NW stand for All-Weights,
Degree-Weights, Value-Weights and No-Weight, respectively, as defined in Section 3.2.2. The
weekly Jensen’s alphas have been multiplied by 100 so they can be interpreted as percentages. ***
p < 0.01, ** p < 0.05, *p < 0.1.

∗∗ ∗∗∗ ∗∗ ∗∗ ∗∗ ∗∗∗ ∗
0.339 0.487 0.245 0.337 0.34 0.54 0.217 0.375∗∗
1
(2.17) (2.75) (1.96) (2.33) (2.22) (2.9) (1.82) (2.5)
0.29∗ 0.512∗∗∗ 0.305∗∗ 0.52∗∗∗ 0.274∗ 0.489∗∗∗ 0.304∗∗ 0.484∗∗∗
2
(1.92) (2.93) (2.26) (3.47) (1.84) (2.65) (2.37) (3.19)
0.331∗∗ 0.549∗∗∗ 0.196 0.445∗∗∗ 0.355∗∗ 0.565∗∗∗ 0.257∗∗ 0.503∗∗∗
3
(2.13) (3.07) (1.48) (2.94) (2.33) (3.06) (2) (3.33)
0.242 0.484∗∗∗ 0.082 0.276∗ 0.337∗∗ 0.61∗∗∗ 0.199 0.41∗∗∗
4
(1.53) (2.71) (0.63) (1.79) (2.21) (3.26) (1.61) (2.61)
0.163 0.316∗ 0.16 0.262 0.208 0.318∗ 0.146 0.203
8
(0.98) (1.74) (1.12) (1.59) (1.41) (1.83) (1.13) (1.31)
0.446∗∗∗ 0.514∗∗∗ 0.476∗∗∗ 0.52∗∗∗ 0.437∗∗∗ 0.427∗∗ 0.456∗∗∗ 0.376∗∗∗
12
(2.71) (2.77) (3.25) (3.15) (2.93) (2.43) (3.57) (2.63)

0.33∗∗ 0.489∗∗∗ 0.239∗ 0.338∗∗ 0.32∗∗ 0.531∗∗∗ 0.191 0.349∗∗
1
(2.14) (2.79) (1.92) (2.34) (2.09) (2.86) (1.6) (2.32)
0.257∗ 0.487∗∗∗ 0.308∗∗ 0.506∗∗∗ 0.247∗ 0.509∗∗∗ 0.306∗∗ 0.507∗∗∗
2
(1.71) (2.85) (2.33) (3.45) (1.66) (2.82) (2.43) (3.34)
0.235 0.499∗∗∗ 0.214 0.441∗∗∗ 0.226 0.531∗∗∗ 0.237∗ 0.506∗∗∗
3
(1.55) (2.9) (1.62) (2.97) (1.52) (2.95) (1.86) (3.34)
0.163 0.475∗∗∗ 0.107 0.324∗∗ 0.226 0.584∗∗∗ 0.204 0.433∗∗∗
4
(1.02) (2.66) (0.81) (2.14) (1.45) (2.98) (1.6) (2.7)
-0.034 0.234 0.009 0.272∗ 0.169 0.361∗∗ 0.16 0.333∗∗
8
(-0.22) (1.41) (0.06) (1.73) (1.14) (2.06) (1.26) (2.04)
0.08 0.224 0.061 0.213 0.172 0.271 0.186 0.253
12
(0.53) (1.31) (0.44) (1.32) (1.21) (1.54) (1.5) (1.61)
54
∗∗ ∗∗∗ ∗∗ ∗∗ ∗∗ ∗∗∗ ∗
0.35 0.475 0.244 0.346 0.331 0.519 0.22 0.371∗∗
1
(2.21) (2.69) (1.96) (2.42) (2.19) (2.81) (1.86) (2.47)
0.306∗ 0.551∗∗∗ 0.3∗∗ 0.481∗∗∗ 0.241 0.487∗∗∗ 0.286∗∗ 0.453∗∗∗
2
(1.94) (3.1) (2.26) (3.31) (1.58) (2.63) (2.29) (2.96)
0.306∗ 0.564∗∗∗ 0.147 0.377∗∗∗ 0.329∗∗ 0.583∗∗∗ 0.212∗ 0.459∗∗∗
3
(1.95) (3.18) (1.14) (2.64) (2.13) (3.13) (1.7) (3)
0.215 0.535∗∗∗ 0.088 0.293∗∗ 0.292∗ 0.648∗∗∗ 0.196 0.428∗∗∗
4
(1.32) (2.93) (0.7) (2) (1.89) (3.37) (1.62) (2.77)
0.106 0.259 0.082 0.323∗∗ 0.247∗ 0.386∗∗ 0.217∗ 0.383∗∗
8
(0.67) (1.5) (0.65) (2.12) (1.67) (2.19) (1.77) (2.46)
0.337∗∗ 0.475∗∗ 0.215 0.363∗∗ 0.341∗∗ 0.393∗∗ 0.279∗∗ 0.307∗∗
12
(2.02) (2.56) (1.53) (2.26) (2.43) (2.3) (2.38) (2.13)

-0.301∗ -0.468∗∗∗ -0.245∗∗ -0.327∗∗ -0.313∗∗ -0.533∗∗∗ -0.218∗ -0.366∗∗
1
(-1.95) (-2.62) (-1.96) (-2.27) (-2.04) (-2.85) (-1.83) (-2.45)
-0.239 -0.507∗∗∗ -0.286∗∗ -0.515∗∗∗ -0.232 -0.484∗∗∗ -0.294∗∗ -0.483∗∗∗
2
(-1.58) (-2.88) (-2.13) (-3.43) (-1.56) (-2.6) (-2.29) (-3.18)
-0.25 -0.528∗∗∗ -0.165 -0.434∗∗∗ -0.26∗ -0.543∗∗∗ -0.237∗ -0.502∗∗∗
3
(-1.62) (-2.93) (-1.26) (-2.83) (-1.74) (-2.94) (-1.85) (-3.28)
-0.164 -0.495∗∗∗ -0.109 -0.318∗∗ -0.242 -0.628∗∗∗ -0.217∗ -0.452∗∗∗
4
(-1.04) (-2.71) (-0.83) (-2.06) (-1.58) (-3.25) (-1.74) (-2.88)
-0.119 -0.341∗ -0.181 -0.342∗∗ -0.136 -0.321∗ -0.182 -0.3∗∗
8
(-0.72) (-1.88) (-1.29) (-2.06) (-0.91) (-1.87) (-1.46) (-2.01)
-0.37∗∗ -0.53∗∗∗ -0.47∗∗∗ -0.527∗∗∗ -0.376∗∗ -0.472∗∗∗ -0.47∗∗∗ -0.456∗∗∗
12
(-2.27) (-2.9) (-3.27) (-3.31) (-2.51) (-2.73) (-3.79) (-3.34)

0.04 -0.069 0.036 -0.119 0.041 -0.038 -0.042 -0.131
1
(0.29) (-0.48) (0.32) (-0.92) (0.36) (-0.26) (-0.42) (-1)
-0.14 -0.196 -0.147 -0.186 -0.139 -0.17 -0.179∗ -0.192
2
(-1.04) (-1.34) (-1.19) (-1.42) (-1.12) (-1.2) (-1.68) (-1.57)
-0.174 -0.155 -0.074 -0.075 -0.151 -0.151 -0.084 -0.079
3
(-1.23) (-1) (-0.61) (-0.56) (-1.12) (-0.96) (-0.8) (-0.61)
-0.334∗∗ -0.321∗∗ -0.315∗∗ -0.319∗∗ -0.331∗∗ -0.306∗ -0.302∗∗ -0.308∗∗
4
(-2.29) (-2.04) (-2.46) (-2.27) (-2.38) (-1.87) (-2.45) (-2.13)
-0.293∗∗ -0.219 -0.294∗∗ -0.296∗∗ -0.239∗ -0.129 -0.213∗ -0.175
8
(-2.03) (-1.37) (-2.27) (-2.12) (-1.79) (-0.79) (-1.85) (-1.27)
-0.113 0.042 -0.137 -0.044 -0.334∗∗ -0.126 -0.355∗∗∗ -0.2
12
(-0.73) (0.24) (-0.97) (-0.26) (-2.55) (-0.79) (-2.94) (-1.24)
55
Table C.3: Carhart 4-Factor Alpha Results, after winsorizing the yearly distribution of stocks’
weekly returns at 1st and 99th percentiles (%, per Week)
11) of U.S.-listed commodity producers as described in Section 3.1.1. The commodity producers’
weekly stock returns are winsorized at the 1st and 99th percentiles to remove outliers present
in the distribution of weekly stock returns, using annually updated cutoffs calculated from our
matched sample. The Long-Short portfolios are rebalanced weekly, and we follow a strategy of
buying the producers stocks with positive signal growth and selling short the stocks with negative
signal growth. The results are based on the Wednesday-to-Tuesday convention, which uses the
CFTC report compilation date as the signal generation date. The signal from the futures market is
constructed as a 1-week lag or as a J-week moving average, where the look-back horizon J is equal
to 2, 3, 4, 8, or 12 weeks. The table report Jensen’s alphas together with t-statistics (based on
Newey-West adjusted standard errors) reported in parentheses. Panel A shows the results for the
regressions on the whole sample period, from August 2006 to September 2018, while Panel B focuses
on non-NBER recession periods only, without the January 2008-June 2009 period. AW, DW, VW
and NW stand for All-Weights, Degree-Weights, Value-Weights and No-Weight, respectively, as
defined in Section 3.2.2. The weekly Jensen’s alphas have been multiplied by 100 so they can be
interpreted as percentages. *** p < 0.01, ** p < 0.05, *p < 0.1.

∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗
0.41 0.645 0.27 0.462 0.439 0.647 0.265 0.453∗∗∗
1
(2.56) (3.58) (2.04) (3.11) (3.04) (3.61) (2.22) (3.1)
0.345∗∗ 0.573∗∗∗ 0.311∗∗ 0.535∗∗∗ 0.383∗∗∗ 0.558∗∗∗ 0.34∗∗∗ 0.499∗∗∗
2
(2.23) (3.25) (2.27) (3.56) (2.63) (3.12) (2.65) (3.36)
0.402∗∗∗ 0.579∗∗∗ 0.272∗∗ 0.524∗∗∗ 0.413∗∗∗ 0.592∗∗∗ 0.324∗∗ 0.53∗∗∗
3
(2.59) (3.23) (2.01) (3.4) (2.82) (3.29) (2.55) (3.53)
0.338∗∗ 0.524∗∗∗ 0.147 0.312∗∗ 0.42∗∗∗ 0.661∗∗∗ 0.251∗∗ 0.447∗∗∗
4
(2.13) (2.96) (1.07) (2.03) (2.82) (3.71) (1.99) (2.91)
0.291∗ 0.426∗∗ 0.232 0.329∗ 0.311∗∗ 0.434∗∗ 0.203 0.273∗
8
(1.72) (2.26) (1.63) (1.94) (2.07) (2.45) (1.56) (1.7)
0.558∗∗∗ 0.723∗∗∗ 0.571∗∗∗ 0.659∗∗∗ 0.492∗∗∗ 0.555∗∗∗ 0.501∗∗∗ 0.463∗∗∗
12
(3.38) (3.83) (3.97) (3.95) (3.34) (3.13) (3.98) (3.09)

0.41∗∗∗ 0.65∗∗∗ 0.275∗∗ 0.462∗∗∗ 0.419∗∗∗ 0.64∗∗∗ 0.239∗∗ 0.425∗∗∗
1
(2.65) (3.67) (2.11) (3.14) (2.91) (3.56) (2.01) (2.91)
0.323∗∗ 0.525∗∗∗ 0.315∗∗ 0.49∗∗∗ 0.355∗∗ 0.555∗∗∗ 0.314∗∗ 0.473∗∗∗
2
(2.11) (3.06) (2.33) (3.37) (2.46) (3.16) (2.5) (3.24)
0.322∗∗ 0.487∗∗∗ 0.289∗∗ 0.509∗∗∗ 0.286∗∗ 0.506∗∗∗ 0.291∗∗ 0.523∗∗∗
3
(2.09) (2.84) (2.11) (3.36) (1.99) (2.89) (2.29) (3.44)
0.268∗ 0.482∗∗∗ 0.198 0.385∗∗ 0.308∗∗ 0.555∗∗∗ 0.266∗∗ 0.46∗∗∗
4
(1.69) (2.78) (1.46) (2.56) (2.07) (3.02) (2.08) (2.96)
0.115 0.305∗ 0.105 0.285∗ 0.243∗ 0.422∗∗ 0.188 0.349∗∗
8
(0.7) (1.75) (0.76) (1.77) (1.66) (2.36) (1.46) (2.11)
0.127 0.357∗∗ 0.086 0.313∗ 0.214 0.349∗ 0.203 0.29∗
12
(0.79) (1.98) (0.61) (1.95) (1.5) (1.91) (1.63) (1.83)
56
∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗
0.404 0.627 0.274 0.491 0.41 0.621 0.264 0.463∗∗∗
1
(2.53) (3.5) (2.05) (3.29) (2.89) (3.48) (2.24) (3.17)
0.334∗∗ 0.6∗∗∗ 0.302∗∗ 0.492∗∗∗ 0.333∗∗ 0.542∗∗∗ 0.304∗∗ 0.435∗∗∗
2
(2.13) (3.38) (2.24) (3.36) (2.26) (3.02) (2.45) (2.93)
0.367∗∗ 0.593∗∗∗ 0.227∗ 0.479∗∗∗ 0.382∗∗∗ 0.601∗∗∗ 0.269∗∗ 0.477∗∗∗
3
(2.32) (3.31) (1.69) (3.24) (2.59) (3.31) (2.17) (3.15)
0.306∗ 0.53∗∗∗ 0.181 0.345∗∗ 0.389∗∗∗ 0.638∗∗∗ 0.258∗∗ 0.463∗∗∗
4
(1.9) (2.97) (1.35) (2.34) (2.61) (3.49) (2.08) (3.05)
0.232 0.35∗ 0.201 0.386∗∗ 0.346∗∗ 0.494∗∗∗ 0.293∗∗ 0.464∗∗∗
8
(1.42) (1.92) (1.51) (2.44) (2.35) (2.75) (2.33) (2.88)
0.405∗∗ 0.642∗∗∗ 0.28∗ 0.449∗∗∗ 0.38∗∗∗ 0.514∗∗∗ 0.325∗∗∗ 0.382∗∗
12
(2.38) (3.34) (1.94) (2.74) (2.69) (2.88) (2.73) (2.57)

-0.39∗∗ -0.624∗∗∗ -0.278∗∗ -0.464∗∗∗ -0.425∗∗∗ -0.648∗∗∗ -0.271∗∗ -0.455∗∗∗
1
(-2.41) (-3.43) (-2.1) (-3.14) (-2.88) (-3.59) (-2.28) (-3.13)
-0.314∗∗ -0.577∗∗∗ -0.294∗∗ -0.547∗∗∗ -0.348∗∗ -0.561∗∗∗ -0.325∗∗ -0.513∗∗∗
2
(-1.98) (-3.21) (-2.12) (-3.63) (-2.35) (-3.09) (-2.53) (-3.46)
-0.336∗∗ -0.564∗∗∗ -0.239∗ -0.484∗∗∗ -0.326∗∗ -0.566∗∗∗ -0.308∗∗ -0.532∗∗∗
3
(-2.15) (-3.12) (-1.79) (-3.13) (-2.24) (-3.14) (-2.46) (-3.51)
-0.277∗ -0.528∗∗∗ -0.18 -0.345∗∗ -0.335∗∗ -0.659∗∗∗ -0.275∗∗ -0.478∗∗∗
4
(-1.73) (-2.91) (-1.32) (-2.23) (-2.21) (-3.57) (-2.18) (-3.1)
-0.258 -0.469∗∗ -0.271∗ -0.442∗∗∗ -0.255∗ -0.445∗∗ -0.254∗∗ -0.39∗∗
8
(-1.52) (-2.49) (-1.94) (-2.62) (-1.68) (-2.51) (-2) (-2.49)
-0.49∗∗∗ -0.738∗∗∗ -0.544∗∗∗ -0.676∗∗∗ -0.441∗∗∗ -0.608∗∗∗ -0.517∗∗∗ -0.581∗∗∗
12
(-2.93) (-3.95) (-3.74) (-4.18) (-2.9) (-3.45) (-4.05) (-4.05)

0.114 0.106 0.059 0.014 0.085 0.088 -0.033 -0.055
1
(0.84) (0.73) (0.49) (0.11) (0.72) (0.62) (-0.31) (-0.43)
0.011 0.087 -0.003 0.018 -0.05 0.073 -0.103 -0.032
2
(0.08) (0.58) (-0.03) (0.13) (-0.4) (0.5) (-0.93) (-0.25)
-0.073 0.07 0.04 0.08 -0.106 0.032 -0.036 0.024
3
(-0.5) (0.44) (0.32) (0.58) (-0.78) (0.2) (-0.33) (0.18)
-0.213 -0.085 -0.176 -0.122 -0.28∗∗ -0.132 -0.258∗∗ -0.181
4
(-1.43) (-0.52) (-1.34) (-0.85) (-1.99) (-0.8) (-2.03) (-1.24)
-0.1 -0.151 -0.14 -0.252∗ -0.173 -0.124 -0.172 -0.174
8
(-0.68) (-0.93) (-1.06) (-1.73) (-1.31) (-0.76) (-1.47) (-1.22)
0.07 0.249 -0.036 0.088 -0.267∗∗ -0.059 -0.338∗∗∗ -0.188
12
(0.45) (1.46) (-0.24) (0.54) (-2.05) (-0.39) (-2.79) (-1.28)
57
Table C.4: Carhart 4-Factor Alpha Results, after removing micro-cap stocks (%, per Week)
of commodity producing firms sorted by various signal measures (as described in Section 3.1.2) based
on the positions of traders in the MM category (as defined in Table 1) in the commodity futures
market. We restrict our attention to ordinary common shares (CRSP share codes 10 and 11) of
U.S.-listed commodity producers as described in Section 3.1.1. We remove stocks in the bottom 25%
of market capitalization using annually updated cutoffs calculated from the CRSP universe. The
Long-Short portfolios are rebalanced weekly, and we follow a strategy of buying the producers stocks
with positive signal growth and selling short the stocks with negative signal growth. The results are
based on the Wednesday-to-Tuesday convention, which uses the CFTC report compilation date as
the signal generation date. The signal from the futures market is constructed as a 1-week lag or as
a J-week moving average, where the look-back horizon J is equal to 2, 3, 4, 8, or 12 weeks. The
table report Jensen’s alphas together with t-statistics (based on White standard errors) reported in
parentheses. Panel A shows the results for the regressions on the whole sample period, from August
2006 to September 2018, while Panel B focuses on non-NBER recession periods only, without the
January 2008-June 2009 period. AW, DW, VW and NW stand for All-Weights, Degree-Weights,
Value-Weights and No-Weight, respectively, as defined in Section 3.2.2. The weekly Jensen’s alphas
have been multiplied by 100 so they can be interpreted as percentages. *** p < 0.01, ** p < 0.05,
*p < 0.1.

0.414∗∗ 0.52∗∗∗ 0.276∗∗ 0.32∗∗ 0.453∗∗∗ 0.531∗∗∗ 0.272∗∗ 0.305∗∗
1
(2.45) (2.77) (1.99) (2.05) (2.95) (2.87) (2.18) (2)
0.345∗∗ 0.455∗∗∗ 0.316∗∗ 0.447∗∗∗ 0.391∗∗ 0.443∗∗ 0.361∗∗∗ 0.421∗∗∗
2
(2.15) (2.58) (2.19) (2.92) (2.54) (2.48) (2.65) (2.78)
0.413∗∗ 0.477∗∗∗ 0.29∗∗ 0.458∗∗∗ 0.42∗∗∗ 0.485∗∗∗ 0.336∗∗ 0.447∗∗∗
3
(2.51) (2.61) (2.05) (2.91) (2.66) (2.68) (2.49) (2.91)
0.347∗∗ 0.398∗∗ 0.151 0.215 0.43∗∗∗ 0.524∗∗∗ 0.256∗ 0.336∗∗
4
(2.09) (2.24) (1.07) (1.39) (2.73) (2.95) (1.95) (2.18)
0.3∗ 0.412∗∗ 0.253∗ 0.347∗∗ 0.311∗∗ 0.409∗∗ 0.213 0.275∗
8
(1.73) (2.17) (1.72) (2.02) (2) (2.32) (1.55) (1.7)
0.601∗∗∗ 0.785∗∗∗ 0.626∗∗∗ 0.78∗∗∗ 0.528∗∗∗ 0.634∗∗∗ 0.553∗∗∗ 0.618∗∗∗
12
(3.5) (4.11) (4.11) (4.46) (3.42) (3.64) (4.07) (3.88)

∗∗ ∗∗∗ ∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗
0.413 0.507 0.281 0.319 0.432 0.505 0.246 0.275∗
1
(2.53) (2.75) (2.05) (2.06) (2.82) (2.73) (1.97) (1.81)
0.317∗∗ 0.406∗∗ 0.314∗∗ 0.394∗∗∗ 0.356∗∗ 0.441∗∗ 0.328∗∗ 0.386∗∗∗
2
(1.98) (2.36) (2.22) (2.66) (2.32) (2.5) (2.47) (2.61)
0.332∗∗ 0.349∗∗ 0.31∗∗ 0.393∗∗ 0.291∗ 0.358∗∗ 0.306∗∗ 0.385∗∗
3
(2.02) (2) (2.16) (2.55) (1.87) (2.02) (2.27) (2.49)
0.279∗ 0.319∗ 0.197 0.27∗ 0.32∗∗ 0.372∗∗ 0.266∗∗ 0.335∗∗
4
(1.67) (1.87) (1.4) (1.77) (2) (2.09) (1.98) (2.15)
0.143 0.274 0.121 0.259 0.268∗ 0.362∗∗ 0.191 0.298∗
8
(0.85) (1.57) (0.84) (1.63) (1.73) (2.04) (1.41) (1.86)
0.158 0.391∗∗ 0.12 0.391∗∗ 0.234 0.356∗∗ 0.229∗ 0.376∗∗
12
(0.95) (2.21) (0.83) (2.33) (1.59) (2.06) (1.76) (2.26)
58
∗∗ ∗∗ ∗∗ ∗∗ ∗∗∗ ∗∗ ∗∗
0.399 0.447 0.278 0.372 0.412 0.436 0.27 0.335∗∗
1
(2.36) (2.45) (1.99) (2.39) (2.72) (2.44) (2.18) (2.21)
0.321∗ 0.43∗∗ 0.309∗∗ 0.386∗∗∗ 0.327∗∗ 0.368∗∗ 0.322∗∗ 0.347∗∗
2
(1.94) (2.43) (2.18) (2.58) (2.05) (2.08) (2.44) (2.28)
0.376∗∗ 0.446∗∗ 0.24∗ 0.405∗∗∗ 0.386∗∗ 0.441∗∗ 0.277∗∗ 0.391∗∗
3
(2.22) (2.49) (1.72) (2.69) (2.39) (2.47) (2.11) (2.54)
0.317∗ 0.387∗∗ 0.185 0.274∗ 0.396∗∗ 0.471∗∗∗ 0.261∗∗ 0.391∗∗∗
4
(1.86) (2.23) (1.34) (1.85) (2.47) (2.72) (2.03) (2.58)
0.26 0.344∗ 0.212 0.364∗∗ 0.365∗∗ 0.456∗∗∗ 0.293∗∗ 0.417∗∗∗
8
(1.54) (1.92) (1.55) (2.34) (2.34) (2.61) (2.22) (2.66)
0.445∗∗ 0.728∗∗∗ 0.308∗∗ 0.549∗∗∗ 0.41∗∗∗ 0.58∗∗∗ 0.345∗∗∗ 0.478∗∗∗
12
(2.53) (3.79) (2.08) (3.21) (2.78) (3.53) (2.77) (3.14)

-0.396∗∗ -0.496∗∗∗ -0.285∗∗ -0.314∗∗ -0.443∗∗∗ -0.537∗∗∗ -0.28∗∗ -0.3∗∗
1
(-2.34) (-2.59) (-2.05) (-2.03) (-2.86) (-2.83) (-2.25) (-1.99)
-0.316∗ -0.439∗∗ -0.295∗∗ -0.451∗∗∗ -0.359∗∗ -0.426∗∗ -0.344∗∗ -0.428∗∗∗
2
(-1.94) (-2.43) (-2.05) (-2.95) (-2.33) (-2.34) (-2.53) (-2.84)
-0.351∗∗ -0.436∗∗ -0.257∗ -0.38∗∗ -0.336∗∗ -0.428∗∗ -0.322∗∗ -0.404∗∗∗
3
(-2.15) (-2.36) (-1.84) (-2.39) (-2.18) (-2.34) (-2.41) (-2.59)
-0.284∗ -0.346∗ -0.186 -0.208 -0.341∗∗ -0.458∗∗ -0.28∗∗ -0.322∗∗
4
(-1.72) (-1.9) (-1.32) (-1.33) (-2.16) (-2.48) (-2.14) (-2.1)
-0.264 -0.402∗∗ -0.284∗∗ -0.412∗∗ -0.253 -0.349∗ -0.254∗ -0.336∗∗
8
(-1.52) (-2.09) (-1.97) (-2.4) (-1.61) (-1.94) (-1.9) (-2.1)
-0.522∗∗∗ -0.767∗∗∗ -0.586∗∗∗ -0.802∗∗∗ -0.465∗∗∗ -0.651∗∗∗ -0.554∗∗∗ -0.74∗∗∗
12
(-3.01) (-3.94) (-3.89) (-4.7) (-2.91) (-3.56) (-4.13) (-4.8)

0.118 0.076 0.072 0.015 0.12 0.074 -0.008 -0.066
1
(0.82) (0.49) (0.57) (0.11) (0.97) (0.5) (-0.08) (-0.47)
0.024 0.052 -0.02 -0.056 -0.019 0.04 -0.094 -0.107
2
(0.17) (0.33) (-0.15) (-0.39) (-0.14) (0.26) (-0.81) (-0.78)
-0.071 0.067 0.034 0.107 -0.084 0.027 -0.02 0.055
3
(-0.47) (0.39) (0.26) (0.73) (-0.59) (0.16) (-0.18) (0.41)
-0.237 -0.135 -0.2 -0.146 -0.294∗∗ -0.179 -0.261∗∗ -0.198
4
(-1.51) (-0.79) (-1.44) (-0.96) (-1.97) (-1.03) (-1.96) (-1.29)
-0.119 -0.225 -0.159 -0.296∗∗ -0.199 -0.223 -0.2 -0.246∗
8
(-0.79) (-1.39) (-1.17) (-2.03) (-1.44) (-1.43) (-1.63) (-1.77)
0.061 0.192 -0.034 0.101 -0.286∗∗ -0.175 -0.341∗∗∗ -0.215
12
(0.38) (1.06) (-0.22) (0.58) (-2.08) (-1.11) (-2.68) (-1.4)
59
D. Carhart 4-Factor Alpha for the Monday-to-Friday Convention
Table D: Carhart 4-Factor Alpha for the Monday-to-Friday Convention (%, per Week)
on the positions of traders in the MM category (as defined in Table 1) in the commodity futures
market. The results are based on the Monday-to-Friday convention, which uses the CFTC report
release date as the signal generation date. We restrict our attention to ordinary common shares
The Long-Short portfolios are rebalanced weekly, and we follow a strategy of buying the producers
stocks with positive signal growth and selling short the stocks with negative signal growth. The
signal from the futures market is constructed as a 1-week lag or as a J-week moving average, where
the look-back horizon J is equal to 2, 3, 4, 8, or 12 weeks. The table report Jensen’s alphas together
with t-statistics (based on White standard errors) reported in parentheses. Panel A shows the
results for the regressions on the whole sample period, from August 2006 to September 2018, while
Panel B focuses on non-NBER recession periods only, without the January 2008-June 2009 period.
AW, DW, VW and NW stand for All-Weights, Degree-Weights, Value-Weights and No-Weight,
respectively, as defined in Section 3.2.2. The weekly Jensen’s alphas have been multiplied by 100 so
they can be interpreted as percentages. *** p < 0.01, ** p < 0.05, *p < 0.1.
∗∗∗ ∗∗∗ ∗ ∗∗∗ ∗ ∗
0.423 0.519 0.248 0.456 0.291 0.315 0.156 0.334∗∗
1
(2.59) (2.9) (1.84) (3.03) (1.73) (1.71) (1.15) (2.17)
0.27 0.457∗∗ 0.228 0.422∗∗ 0.294∗ 0.422∗∗ 0.29∗∗ 0.463∗∗∗
2
(1.57) (2.35) (1.58) (2.52) (1.68) (2.06) (1.97) (2.66)
0.287∗ 0.355∗ 0.23 0.306∗ 0.365∗∗ 0.434∗∗ 0.256∗ 0.315∗
3
(1.73) (1.77) (1.62) (1.86) (2.19) (2.05) (1.81) (1.84)
0.356∗∗ 0.584∗∗∗ 0.332∗∗ 0.442∗∗∗ 0.443∗∗ 0.672∗∗∗ 0.401∗∗∗ 0.479∗∗∗
4
(2.04) (2.71) (2.26) (2.74) (2.51) (2.86) (2.8) (2.89)
0.312∗ 0.425∗∗ 0.258∗ 0.387∗∗ 0.322∗∗ 0.43∗∗ 0.252∗ 0.333∗∗
8
(1.85) (2.26) (1.88) (2.51) (1.99) (2.2) (1.94) (2.09)
0.263 0.508∗∗ 0.176 0.387∗∗ 0.15 0.272 0.136 0.23
12
(1.46) (2.42) (1.14) (2.17) (0.98) (1.39) (1.04) (1.41)

-0.215 -0.447∗∗ -0.146 -0.352∗∗ -0.236 -0.381∗∗ -0.097 -0.267∗
1
(-1.33) (-2.54) (-1.09) (-2.37) (-1.49) (-2.1) (-0.72) (-1.74)
-0.168 -0.44∗∗ -0.246∗ -0.486∗∗∗ -0.187 -0.43∗∗ -0.308∗∗ -0.534∗∗∗
2
(-1.01) (-2.34) (-1.73) (-2.99) (-1.1) (-2.14) (-2.11) (-3.12)
-0.212 -0.421∗∗ -0.243∗ -0.397∗∗ -0.215 -0.484∗∗ -0.311∗∗ -0.451∗∗∗
3
(-1.27) (-2.23) (-1.66) (-2.51) (-1.27) (-2.41) (-2.09) (-2.73)
-0.247 -0.558∗∗∗ -0.258∗ -0.385∗∗ -0.271 -0.631∗∗∗ -0.308∗∗ -0.445∗∗∗
4
(-1.41) (-2.82) (-1.82) (-2.37) (-1.59) (-2.99) (-2.16) (-2.65)
-0.368∗∗ -0.535∗∗∗ -0.296∗∗ -0.464∗∗∗ -0.23 -0.399∗∗ -0.156 -0.305∗
8
(-2.02) (-2.76) (-1.99) (-2.7) (-1.36) (-2.1) (-1.13) (-1.8)
-0.454∗∗∗ -0.647∗∗∗ -0.471∗∗∗ -0.589∗∗∗ -0.277∗ -0.418∗∗ -0.315∗∗ -0.386∗∗
12
(-2.66) (-3.38) (-3.19) (-3.48) (-1.8) (-2.23) (-2.42) (-2.48)
60
∗∗ ∗∗∗ ∗∗ ∗∗ ∗∗
0.324 0.491 0.183 0.358 0.316 0.392 0.144 0.284∗
1
(2.04) (2.82) (1.39) (2.41) (1.96) (2.17) (1.08) (1.85)
0.249 0.429∗∗ 0.249∗ 0.473∗∗∗ 0.299∗ 0.439∗∗ 0.322∗∗ 0.52∗∗∗
2
(1.5) (2.34) (1.73) (2.87) (1.75) (2.24) (2.18) (3)
0.292∗ 0.378∗∗ 0.275∗ 0.396∗∗ 0.328∗∗ 0.46∗∗ 0.309∗∗ 0.435∗∗∗
3
(1.78) (2.04) (1.9) (2.47) (1.98) (2.36) (2.08) (2.61)
0.299∗ 0.507∗∗∗ 0.287∗∗ 0.392∗∗ 0.359∗∗ 0.599∗∗∗ 0.315∗∗ 0.422∗∗
4
(1.74) (2.63) (2.01) (2.39) (2.1) (2.91) (2.21) (2.5)
0.403∗∗ 0.501∗∗∗ 0.259∗ 0.333∗∗ 0.307∗ 0.37∗∗ 0.146 0.168
8
(2.25) (2.6) (1.72) (1.97) (1.85) (1.98) (1.06) (1.01)
0.484∗∗∗ 0.597∗∗∗ 0.478∗∗∗ 0.509∗∗∗ 0.34∗∗ 0.363∗ 0.35∗∗∗ 0.312∗
12
(2.77) (3.04) (3.17) (2.85) (2.14) (1.91) (2.59) (1.84)
0.291∗ 0.455∗∗∗ 0.207 0.389∗∗∗ 0.303∗ 0.386∗∗ 0.169 0.334∗∗

1
(1.86) (2.68) (1.57) (2.63) (1.9) (2.16) (1.26) (2.17)
0.144 0.346∗ 0.202 0.392∗∗ 0.171 0.369∗ 0.256∗ 0.453∗∗
2
(0.87) (1.89) (1.42) (2.35) (1.01) (1.88) (1.75) (2.56)
0.18 0.317∗ 0.254∗ 0.327∗ 0.205 0.415∗∗ 0.277∗ 0.368∗∗
3
(1.1) (1.65) (1.74) (1.95) (1.24) (2.02) (1.85) (2.08)
0.24 0.547∗∗ 0.308∗∗ 0.468∗∗∗ 0.271 0.607∗∗ 0.354∗∗ 0.491∗∗∗
4
(1.35) (2.55) (2.08) (2.82) (1.51) (2.57) (2.46) (2.87)
0.199 0.336∗ 0.136 0.267∗ 0.204 0.322∗ 0.131 0.236
8
(1.2) (1.88) (0.99) (1.7) (1.24) (1.67) (0.98) (1.41)
-0.035 0.204 -0.029 0.177 0.004 0.159 0.089 0.186
12
(-0.21) (1.04) (-0.19) (1.04) (0.02) (0.81) (0.65) (1.09)
61
E. Signals of the Commercial Hedger (PM) Category
Table E.1: Carhart 4-Factor Alpha Results for the Monday-to-Friday Convention based on
PM Signals (%, per Week)
on the positions of traders in the PM category (as defined in Table 1) in the commodity futures
market. The results are based on the Monday-to-Friday convention, which uses the CFTC report
release date as the signal generation date. We restrict our attention to ordinary common shares
The Long-Short portfolios are rebalanced weekly, and we follow a strategy of buying the producers
stocks with positive signal growth and selling short the stocks with negative signal growth. The
signal from the futures market is constructed as a 1-week lag or as a J-week moving average, where
the look-back horizon J is equal to 2, 3, 4, 8, or 12 weeks. The table report Jensen’s alphas together
with t-statistics (based on White standard errors) reported in parentheses. Panel A shows the
results for the regressions on the whole sample period, from August 2006 to September 2018, while
Panel B focuses on non-NBER recession periods only, without the January 2008-June 2009 period.
AW, DW, VW and NW stand for All-Weights, Degree-Weights, Value-Weights and No-Weight,
respectively, as defined in Section 3.2.2. The weekly Jensen’s alphas have been multiplied by 100 so
they can be interpreted as percentages. *** p < 0.01, ** p < 0.05, *p < 0.1.
PM Traders’ Net Change
-0.043 -0.094 0.048 -0.037 -0.052 -0.134 -0.007 -0.14
1
(-0.28) (-0.57) (0.37) (-0.25) (-0.34) (-0.78) (-0.06) (-0.93)
-0.19 -0.296∗ -0.1 -0.228 -0.089 -0.283 -0.014 -0.206
2
(-1.19) (-1.67) (-0.77) (-1.55) (-0.59) (-1.56) (-0.11) (-1.36)
-0.209 -0.259 -0.159 -0.238∗ -0.145 -0.257 -0.161 -0.308∗∗
3
(-1.25) (-1.46) (-1.2) (-1.65) (-0.88) (-1.42) (-1.29) (-2.08)
-0.223 -0.332∗ -0.228∗ -0.206 -0.062 -0.212 -0.109 -0.111
4
(-1.29) (-1.79) (-1.67) (-1.41) (-0.37) (-1.09) (-0.87) (-0.75)
-0.408∗∗ -0.519∗∗ -0.407∗∗∗ -0.457∗∗ -0.114 -0.289 -0.154 -0.256
8
(-2.3) (-2.48) (-2.86) (-2.53) (-0.67) (-1.28) (-1.17) (-1.3)
-0.398∗∗ -0.545∗∗ -0.311∗∗ -0.383∗∗ -0.132 -0.27 -0.2 -0.244
12
(-2.3) (-2.44) (-2.14) (-2.02) (-0.83) (-1.14) (-1.55) (-1.27)
PM Traders’ Long Short Ratio Growth
-0.039 -0.085 0.046 -0.051 -0.047 -0.124 -0.01 -0.157

1
(-0.25) (-0.52) (0.35) (-0.35) (-0.31) (-0.73) (-0.08) (-1.04)
-0.209 -0.306∗ -0.145 -0.267∗ -0.11 -0.298 -0.081 -0.269∗
2
(-1.3) (-1.73) (-1.12) (-1.82) (-0.72) (-1.64) (-0.65) (-1.76)
-0.198 -0.24 -0.15 -0.22 -0.104 -0.212 -0.118 -0.249∗
3
(-1.17) (-1.36) (-1.1) (-1.51) (-0.62) (-1.16) (-0.92) (-1.65)
-0.227 -0.335∗ -0.223 -0.246∗ -0.061 -0.217 -0.116 -0.165
4
(-1.29) (-1.83) (-1.64) (-1.71) (-0.36) (-1.14) (-0.93) (-1.16)
-0.424∗∗ -0.504∗∗ -0.375∗∗∗ -0.392∗∗ -0.135 -0.292 -0.147 -0.237
8
(-2.39) (-2.32) (-2.65) (-2.14) (-0.79) (-1.23) (-1.13) (-1.18)
-0.404∗∗ -0.526∗∗ -0.382∗∗∗ -0.43∗∗ -0.132 -0.278 -0.255∗∗ -0.298
12
(-2.38) (-2.41) (-2.66) (-2.33) (-0.83) (-1.19) (-1.96) (-1.57)
62
PM Traders’ Long Proportion Growth
0 -0.051 0.075 0.005 -0.024 -0.108 -0.002 -0.108
1
(0) (-0.3) (0.58) (0.04) (-0.16) (-0.61) (-0.02) (-0.72)
-0.188 -0.276 -0.136 -0.279∗ -0.092 -0.268 -0.087 -0.29∗
2
(-1.14) (-1.51) (-1.05) (-1.88) (-0.59) (-1.43) (-0.7) (-1.9)
-0.226 -0.244 -0.179 -0.227 -0.159 -0.236 -0.146 -0.275∗
3
(-1.32) (-1.36) (-1.34) (-1.59) (-0.94) (-1.28) (-1.16) (-1.86)
-0.242 -0.322∗ -0.257∗ -0.26∗ -0.07 -0.202 -0.119 -0.152
4
(-1.37) (-1.73) (-1.87) (-1.82) (-0.41) (-1.04) (-0.96) (-1.06)
-0.444∗∗ -0.523∗∗ -0.406∗∗∗ -0.466∗∗ -0.137 -0.303 -0.145 -0.311
8
(-2.52) (-2.45) (-2.87) (-2.54) (-0.81) (-1.29) (-1.1) (-1.54)
-0.371∗∗ -0.437∗ -0.312∗∗ -0.364∗∗ -0.143 -0.255 -0.221∗ -0.295
12
(-2.13) (-1.94) (-2.22) (-1.98) (-0.9) (-1.07) (-1.77) (-1.56)
PM Traders’ Short Proportion Growth

0.159 0.225 -0.016 0.089 0.136 0.244 0.033 0.187
1
(1.07) (1.38) (-0.12) (0.61) (0.9) (1.41) (0.26) (1.24)
0.243 0.351∗∗ 0.053 0.167 0.176 0.399∗∗ 0.066 0.243
2
(1.59) (2) (0.41) (1.14) (1.17) (2.17) (0.54) (1.61)
0.143 0.242 0.112 0.2 0.084 0.26 0.159 0.306∗∗
3
(0.89) (1.37) (0.86) (1.41) (0.52) (1.41) (1.25) (2.07)
0.254 0.387∗∗ 0.212 0.234 0.116 0.294 0.136 0.169
4
(1.53) (2.14) (1.58) (1.64) (0.72) (1.53) (1.1) (1.17)
0.37∗∗ 0.444∗∗ 0.248∗ 0.239 0.235 0.376∗ 0.155 0.219
8
(2.19) (2.2) (1.76) (1.54) (1.45) (1.69) (1.22) (1.42)
0.395∗∗ 0.409∗∗ 0.327∗∗ 0.31∗ 0.142 0.124 0.159 0.112
12
(2.37) (2.08) (2.21) (1.74) (0.97) (0.66) (1.28) (0.66)
63
F. Additional Fama-Macbeth Regression Results for Section 4.3
Table F.1: Fama-Macbeth Regressions: Managed Money Long Short Ratio Growth, Non-Recession
Periods Only
Notes: This table shows results from Fama-MacBeth cross-sectional regressions (average slopes, and Newey-
West adjusted t-statistics with five lags) of firms’ subsequent daily return (subtracted by risk-free rate) on
lagged signal and other lagged controls for expected returns. The daily return of the firm occurs within 7
calendar days (the first is always a Wednesday and the last is always a Tuesday unless they are postponed
by one day due to holidays) following the newest CFTC Disaggregated Commitments of Traders report. We
run the Fama-Macbeth regression at daily frequency. The signal from the futures market is constructed as a
1-week lag or as a J-week moving average, where J is equal to 2, 3, 4, 8, or 12. ret−1 is the stock return
over the previous month; ret−2,−12 is the stock return over the 11 months preceding the previous month;
ln(BE/ME) denotes the log of the ratio of book value of equity to market value of equity; ln(ME) is the log of
the market value of equity. ∆CP rice−1 is the change in commodiy pice over the previous week. We present
t-statistics in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% level, respectively.
Adj.R2 reports the average of the cross-sectional adjusted R2 ’s. N –Companies is the number of unique firms,
and N –Observations is the number of firm-day return observations utilized in the regression.
Non-Recession Periods Only
lag1 MA(2) MA(3)
(1) (2) (3) (4) (5) (6)
Managed Money +0.004∗∗∗ +0.005∗∗∗ +0.003∗∗ +0.004∗∗∗ +0.003∗ +0.004∗∗
Long Short Ratio Growth (+3.071) (+3.559) (+2.103) (+2.585) (+1.795) (+2.367)
ln(BE/M E) +0.000 +0.000 +0.000
(+0.074) (+0.041) (+0.040)
ln(M E) −0.000 −0.000 −0.000
(-0.627) (-0.711) (-0.708)
ret−1 +0.001 +0.001 +0.001
(+0.596) (+0.660) (+0.582)
ret−2,−12 +0.001 +0.001 +0.001
(+1.544) (+1.540) (+1.429)
∆CP rice−1 +0.006 +0.013∗ +0.008
(+0.823) (+1.870) (+1.185)
N –Companies 198 191 198 190 194 190
N –Observations 258910 244791 257839 244094 257014 243477
Adj.R2 0.06850 0.06823 0.06849
MA(4) MA(8) MA(12)
(7) (8) (9) (10) (11) (12)
Managed Money +0.006∗∗∗ +0.006∗∗∗ +0.005∗ +0.005∗∗ +0.005 +0.005
ln(BE/M E) +0.000 −0.000 −0.000
(+0.048) (-0.123) (-0.054)
ln(M E) −0.000 −0.000 −0.000
(-0.766) (-0.817) (-0.799)
ret−1 +0.001 +0.001 +0.000
(+0.638) (+0.466) (+0.306)
ret−2,−12 +0.001 +0.001 +0.001
(+1.391) (+1.367) (+1.302)
∆CP rice−1 +0.009 +0.015∗∗ +0.015∗∗
(+1.280) (+2.121) (+2.040)
N –Companies 194 189 192 189 192 189
N –Observations 256321 242901 254081 241094 250365 238626
Adj.R2 0.06925 0.06841 0.06792
64
Table F.2: Fama-Macbeth Regressions: Managed Money Long Short Ratio Growth, All Observations
All Observations
lag1 MA(2) MA(3)
(1) (2) (3) (4) (5) (6)
Managed Money +0.002∗ +0.006∗∗∗ +0.004∗∗ +0.004∗∗ +0.004∗ +0.004∗
ln(BE/M E) +0.000 +0.000 +0.000
(+0.310) (+0.235) (+0.190)
ln(M E) −0.000∗ −0.000∗ −0.000∗
(-1.845) (-1.910) (-1.921)
ret−1 +0.001 +0.001 +0.001
(+0.467) (+0.527) (+0.473)
ret−2,−12 +0.000 +0.000 +0.000
(+1.266) (+1.270) (+1.162)
∆CP rice−1 +0.006 +0.010 +0.006
(+0.843) (+1.595) (+0.974)
N –Companies 199 192 199 191 195 191
N –Observations 302971 287996 301900 287299 301075 286682
Adj.R2 0.06935 0.06920 0.06923
MA(4) MA(8) MA(12)
(7) (8) (9) (10) (11) (12)
∗∗ ∗∗ ∗ ∗
Managed Money +0.006 +0.006 +0.006 +0.007 +0.005 +0.010∗
ln(BE/M E) +0.000 +0.000 +0.000
(+0.226) (+0.107) (+0.161)
ln(M E) −0.000∗∗ −0.000∗∗ −0.000∗
(-2.009) (-2.018) (-1.921)
ret−1 +0.001 +0.000 +0.000
(+0.505) (+0.355) (+0.138)
ret−2,−12 +0.000 +0.000 +0.000
(+1.173) (+1.132) (+1.088)
∆CP rice−1 +0.009 +0.013∗∗ +0.015∗∗
(+1.348) (+2.041) (+2.289)
N –Companies 195 190 193 190 193 190
N –Observations 300382 286106 298142 284299 294426 281831
Adj.R2 0.06996 0.06926 0.06886
65
Table F.3: Fama-Macbeth Regressions: Managed Money Net Change, Non-Recession Periods Only
lag1 MA(2) MA(3)
(1) (2) (3) (4) (5) (6)
Managed Money +0.004∗∗∗ +0.005∗∗∗ +0.004∗∗∗ +0.004∗∗∗ +0.004∗∗ +0.005∗∗∗
Net Change (+3.330) (+3.676) (+2.622) (+2.809) (+2.462) (+2.876)
ln(BE/M E) +0.000 −0.000 −0.000
(+0.033) (-0.023) (-0.035)
ln(M E) −0.000 −0.000 −0.000
(-0.592) (-0.661) (-0.650)
ret−1 +0.001 +0.001 +0.001
(+0.590) (+0.656) (+0.614)
ret−2,−12 +0.001 +0.001 +0.001
(+1.610) (+1.595) (+1.491)
∆CP rice−1 +0.005 +0.012∗ +0.011
(+0.673) (+1.724) (+1.533)
N –Companies 198 191 198 190 194 190
N –Observations 259278 245096 258080 244327 257210 243665
Adj.R2 0.06809 0.06815 0.06821
MA(4) MA(8) MA(12)
(7) (8) (9) (10) (11) (12)
∗∗∗ ∗∗∗ ∗
Managed Money +0.006 +0.007 +0.001 +0.004 +0.007 +0.012∗∗∗
Net Change (+2.975) (+3.172) (+0.329) (+1.208) (+1.935) (+2.951)
ln(BE/M E) −0.000 −0.000 −0.000
(-0.052) (-0.052) (-0.008)
ln(M E) −0.000 −0.000 −0.000
(-0.666) (-0.712) (-0.749)
ret−1 +0.001 +0.001 +0.001
(+0.630) (+0.563) (+0.381)
ret−2,−12 +0.001 +0.001 +0.001
(+1.428) (+1.436) (+1.489)
∆CP rice−1 +0.009 +0.015∗∗ +0.019∗∗
(+1.316) (+2.185) (+2.535)
N –Companies 194 189 192 189 192 189
N –Observations 256485 243057 254121 241134 250405 238666
Adj.R2 0.06883 0.06854 0.06794
66
Table F.4: Fama-Macbeth Regressions: Managed Money Net Change, All Observations
All Observations
lag1 MA(2) MA(3)
(1) (2) (3) (4) (5) (6)
Managed Money +0.002 +0.006∗∗∗ +0.005∗∗∗ +0.004∗∗ +0.005∗∗ +0.004∗
Net Change (+1.637) (+2.902) (+3.034) (+2.325) (+2.246) (+1.832)
ln(BE/M E) +0.000 +0.000 +0.000
(+0.248) (+0.169) (+0.101)
ln(M E) −0.000∗ −0.000∗ −0.000∗
(-1.823) (-1.900) (-1.905)
ret−1 +0.001 +0.001 +0.001
(+0.458) (+0.515) (+0.487)
ret−2,−12 +0.001 +0.001 +0.000
(+1.314) (+1.311) (+1.206)
∆CP rice−1 +0.005 +0.010 +0.009
(+0.753) (+1.569) (+1.384)
N –Companies 199 192 199 191 195 191
N –Observations 303387 288301 302141 287532 301271 286870
Adj.R2 0.06898 0.06915 0.06902
MA(4) MA(8) MA(12)
(7) (8) (9) (10) (11) (12)
∗∗ ∗∗ ∗ ∗
Managed Money +0.006 +0.006 +0.005 +0.006 +0.009 +0.013∗∗∗
Net Change (+2.095) (+2.078) (+1.256) (+1.650) (+1.933) (+2.831)
ln(BE/M E) +0.000 +0.000 +0.000
(+0.142) (+0.145) (+0.205)
ln(M E) −0.000∗ −0.000∗∗ −0.000∗∗
(-1.931) (-1.964) (-1.967)
ret−1 +0.001 +0.000 +0.000
(+0.489) (+0.415) (+0.246)
ret−2,−12 +0.000 +0.000 +0.000
(+1.237) (+1.221) (+1.236)
∆CP rice−1 +0.009 +0.014∗∗ +0.017∗∗
(+1.364) (+2.068) (+2.502)
N –Companies 195 190 193 190 193 190
N –Observations 300546 286262 298182 284339 294466 281871
Adj.R2 0.06967 0.06955 0.06897
67
Table F.5: Fama-Macbeth Regressions: Managed Money Long Proportion Growth, Non-Recession
Periods Only
lag1 MA(2) MA(3)
(1) (2) (3) (4) (5) (6)
Managed Money +0.012∗∗∗ +0.013∗∗∗ +0.012∗∗ +0.009∗ +0.014∗∗ +0.009
Long Proportion Growth (+3.171) (+2.818) (+2.463) (+1.847) (+2.521) (+1.452)
ln(BE/M E) −0.000 −0.000 −0.000
(-0.027) (-0.079) (-0.110)
ln(M E) −0.000 −0.000 −0.000
(-0.513) (-0.611) (-0.598)
ret−1 +0.001 +0.001 +0.001
(+0.908) (+0.905) (+0.860)
ret−2,−12 +0.001 +0.001 +0.001
(+1.399) (+1.583) (+1.461)
∆CP rice−1 +0.008 +0.014∗∗ +0.014∗∗
(+1.123) (+2.072) (+2.018)
N –Companies 198 191 198 190 198 190
N –Observations 263022 248205 262106 247718 261394 247277
Adj.R2 0.06686 0.06651 0.06693
MA(4) MA(8) MA(12)
(7) (8) (9) (10) (11) (12)
∗ ∗ ∗ ∗∗ ∗
Managed Money +0.012 +0.011 +0.016 +0.018 +0.021 +0.032∗∗
ln(BE/M E) −0.000 −0.000 −0.000
(-0.092) (-0.106) (-0.027)
ln(M E) −0.000 −0.000 −0.000
(-0.651) (-0.830) (-0.830)
ret−1 +0.001 +0.001 +0.001
(+0.852) (+0.745) (+0.716)
ret−2,−12 +0.001 +0.001 +0.001
(+1.323) (+1.325) (+1.420)
∆CP rice−1 +0.014∗∗ +0.016∗∗ +0.015∗∗
(+2.113) (+2.510) (+2.268)
N –Companies 198 189 192 189 192 189
N –Observations 260727 246826 258553 245184 254810 242726
Adj.R2 0.06733 0.06726 0.06647
68
Table F.6: Fama-Macbeth Regressions: Managed Money Long Proportion Growth, All Observations
All Observations
lag1 MA(2) MA(3)
(1) (2) (3) (4) (5) (6)
Managed Money +0.009∗ +0.010∗ +0.015∗∗∗ +0.006 +0.015∗∗ +0.002
ln(BE/M E) +0.000 +0.000 +0.000
(+0.177) (+0.122) (+0.060)
ln(M E) −0.000∗ −0.000∗ −0.000∗
(-1.741) (-1.865) (-1.820)
ret−1 +0.001 +0.001 +0.001
(+0.777) (+0.740) (+0.707)
ret−2,−12 +0.000 +0.001 +0.000
(+1.122) (+1.286) (+1.190)
∆CP rice−1 +0.009 +0.012∗∗ +0.013∗
(+1.230) (+1.967) (+1.956)
N –Companies 199 192 199 191 199 191
N –Observations 307311 291410 306395 290923 305683 290482
Adj.R2 0.06797 0.06756 0.06771
MA(4) MA(8) MA(12)
(7) (8) (9) (10) (11) (12)
∗ ∗ ∗∗
Managed Money +0.010 +0.002 +0.021 +0.023 +0.029 +0.049∗∗∗
ln(BE/M E) +0.000 +0.000 +0.000
(+0.082) (+0.116) (+0.169)
ln(M E) −0.000∗ −0.000∗∗ −0.000∗∗
(-1.947) (-2.095) (-1.993)
ret−1 +0.001 +0.001 +0.001
(+0.714) (+0.606) (+0.525)
ret−2,−12 +0.000 +0.000 +0.000
(+1.140) (+1.139) (+1.213)
∆CP rice−1 +0.014∗∗ +0.014∗∗ +0.015∗∗
(+2.167) (+2.166) (+2.375)
N –Companies 199 190 199 190 193 190
N –Observations 305016 290031 302674 288389 298871 285931
Adj.R2 0.06815 0.06821 0.06739
69
Table F.7: Fama-Macbeth Regressions: Managed Money Short Proportion Growth, Non-Recession
Periods Only
lag1 MA(2) MA(3)
(1) (2) (3) (4) (5) (6)
Managed Money −0.005∗∗∗ −0.008∗∗∗ −0.004∗∗ −0.007∗∗∗ −0.004∗ −0.006∗∗
Short Proportion Growth (-2.914) (-3.436) (-2.121) (-3.156) (-1.686) (-2.172)
ln(BE/M E) +0.000 −0.000 +0.000
(+0.028) (-0.019) (+0.020)
ln(M E) −0.000 −0.000 −0.000
(-0.570) (-0.627) (-0.616)
ret−1 +0.001 +0.001 +0.001
(+0.609) (+0.647) (+0.628)
ret−2,−12 +0.001 +0.001 +0.001
(+1.586) (+1.498) (+1.422)
∆CP rice−1 +0.005 +0.010 +0.011
(+0.644) (+1.523) (+1.549)
N –Companies 198 191 198 190 197 190
N –Observations 262285 247136 261224 246432 260374 245835
Adj.R2 0.06818 0.06825 0.06849
MA(4) MA(8) MA(12)
(7) (8) (9) (10) (11) (12)
∗∗∗ ∗∗∗ ∗
Managed Money −0.008 −0.010 +0.001 −0.003 −0.008 −0.012∗∗
Short Proportion Growth (-2.621) (-3.091) (+0.137) (-0.716) (-1.717) (-2.378)
ln(BE/M E) −0.000 −0.000 −0.000
(-0.024) (-0.012) (-0.021)
ln(M E) −0.000 −0.000 −0.000
(-0.624) (-0.631) (-0.656)
ret−1 +0.001 +0.001 +0.001
(+0.633) (+0.586) (+0.408)
ret−2,−12 +0.001 +0.001 +0.001
(+1.364) (+1.456) (+1.446)
∆CP rice−1 +0.009 +0.016∗∗ +0.019∗∗∗
(+1.307) (+2.362) (+2.703)
N –Companies 197 189 197 189 194 189
N –Observations 259665 245289 257381 243616 253750 241395
Adj.R2 0.06904 0.06863 0.06772
70
Table F.8: Fama-Macbeth Regressions: Managed Money Short Proportion Growth, All Observations
All Observations
lag1 MA(2) MA(3)
(1) (2) (3) (4) (5) (6)
Managed Money −0.003 −0.007∗∗∗ −0.007∗∗∗ −0.008∗∗∗ −0.007∗∗ −0.006
ln(BE/M E) +0.000 +0.000 +0.000
(+0.249) (+0.167) (+0.152)
ln(M E) −0.000∗ −0.000∗ −0.000∗
(-1.803) (-1.862) (-1.896)
ret−1 +0.001 +0.001 +0.001
(+0.465) (+0.501) (+0.499)
ret−2,−12 +0.001 +0.000 +0.000
(+1.303) (+1.233) (+1.152)
∆CP rice−1 +0.005 +0.009 +0.009
(+0.753) (+1.339) (+1.409)
N –Companies 199 192 199 191 198 191
N –Observations 306394 290341 305285 289637 304435 289040
Adj.R2 0.06898 0.06921 0.06925
MA(4) MA(8) MA(12)
(7) (8) (9) (10) (11) (12)
∗∗ ∗∗ ∗
Managed Money −0.009 −0.010 −0.006 −0.008 −0.011 −0.014∗∗
ln(BE/M E) +0.000 +0.000 +0.000
(+0.160) (+0.187) (+0.211)
ln(M E) −0.000∗ −0.000∗ −0.000∗
(-1.897) (-1.869) (-1.863)
ret−1 +0.001 +0.001 +0.000
(+0.479) (+0.425) (+0.262)
ret−2,−12 +0.000 +0.000 +0.000
(+1.178) (+1.233) (+1.185)
∆CP rice−1 +0.009 +0.015∗∗ +0.017∗∗
(+1.427) (+2.285) (+2.550)
N –Companies 198 190 198 190 195 190
N –Observations 303726 288494 301442 286821 297811 284600
Adj.R2 0.06983 0.06972 0.06902
71
G. Additional Double-Sorting Results for Section 4.5
Table G.1: Double Sorting: Money Managers’ Long-Short Ratio Growth (%, per Week)
two signal portfolios are formed dependently within each of the three friction portfolios, based on
the MM Long Short Ratio Growth signal. The signal from the futures market is constructed with
a 1-week lag, and the same is true for the three proxies of market friction. The results are based
on the Wednesday-to-Tuesday convention, which uses the CFTC report compilation date as the
signal generation date. This 3-by-2 double-sorting procedure produces six portfolios. Finally, we
both equally weight (Panel A) or value-weight (Panel B) the sorts. The significance of the sorting
variable is assessed by calculating the Jensen’s alpha relative to the Carhart (1997) four-factor
model (C4 α) and the Fama and French (2015) five-factor model (FF5 α). The table reports average
returns, together with t-statistics reported in parentheses, and the t-statistics for the Jensen’s alphas
are based on Newey-West adjusted standard errors. The returns and Jensen’s alphas, per week,
have been multiplied by 100. *** p < 0.01, ** p < 0.05, *p < 0.1.

AD 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.249 0.004 -0.385 -0.634∗∗∗ 0.254 -0.003 -0.432 -0.685∗∗∗
1
(1.30) (0.02) (-1.42) (-3.30) (1.38) (-0.01) (-1.63) (-3.54)
0.333∗ 0.427∗∗ 0.427 0.094 0.3∗ 0.423∗∗ 0.447∗ 0.146
2
(1.82) (2.15) (1.60) (0.47) (1.78) (2.21) (1.71) (0.74)
0.08 0.419∗∗∗ 0.808∗∗∗ 0.727∗∗∗ 0.048 0.421∗∗∗ 0.9∗∗∗ 0.852∗∗∗
(2-1)
(0.47) (2.67) (3.23) (2.81) (0.28) (2.71) (3.66) (3.36)
0.092 0.399∗∗ 0.867∗∗∗ 0.775∗∗∗ 0.092 0.413∗∗∗ 0.958∗∗∗ 0.866∗∗∗
C4 α
(0.53) (2.53) (3.22) (2.79) (0.52) (2.62) (3.62) (3.16)
0.113 0.399∗∗ 0.877∗∗∗ 0.764∗∗ 0.079 0.408∗∗ 0.968∗∗∗ 0.888∗∗∗
FF5 α
(0.61) (2.32) (2.91) (2.48) (0.43) (2.39) (3.29) (2.93)
0.177 0.439∗∗ 0.884∗∗∗ 0.708∗∗ 0.169 0.444∗∗∗ 0.984∗∗∗ 0.815∗∗∗
SY4 α
(0.93) (2.55) (2.98) (2.31) (0.87) (2.59) (3.37) (2.66)
VOL 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.15 0.059 -0.355 -0.505∗∗∗ 0.071 0.062 -0.341 -0.412∗∗
1
(0.98) (0.27) (-1.33) (-2.80) (0.48) (0.29) (-1.27) (-2.24)
0.276∗ 0.328 0.47∗ 0.194 0.242 0.328 0.466∗ 0.224
2
(1.73) (1.64) (1.70) (0.92) (1.55) (1.57) (1.71) (1.09)
0.125 0.27∗ 0.825∗∗∗ 0.7∗∗∗ 0.171 0.265∗ 0.807∗∗∗ 0.635∗∗∗
(2-1)
(0.95) (1.75) (3.49) (2.94) (1.33) (1.72) (3.35) (2.67)
0.132 0.287∗ 0.83∗∗∗ 0.698∗∗∗ 0.187 0.278∗ 0.804∗∗∗ 0.617∗∗
C4 α
(0.97) (1.84) (3.39) (2.84) (1.41) (1.80) (3.27) (2.54)
0.097 0.236 0.976∗∗∗ 0.878∗∗∗ 0.148 0.236 1.03∗∗∗ 0.881∗∗∗
FF5 α
(0.64) (1.37) (3.47) (3.12) (1.01) (1.35) (3.68) (3.19)
0.127 0.269 0.895∗∗∗ 0.767∗∗∗ 0.209 0.288 0.943∗∗∗ 0.734∗∗∗
SY4 α
(0.82) (1.49) (3.23) (2.75) (1.38) (1.59) (3.38) (2.65)
72
LIQ 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.006 -0.024 -0.121 -0.128 -0.002 0.065 -0.192 -0.19
1
(0.04) (-0.10) (-0.52) (-0.73) (-0.01) (0.29) (-0.81) (-1.05)
0.252 0.38∗ 0.462∗ 0.21 0.244 0.446∗∗ 0.416∗ 0.172
2
(1.41) (1.82) (1.89) (1.18) (1.47) (2.14) (1.69) (0.97)
0.246∗ 0.403∗∗ 0.583∗∗ 0.337 0.246∗ 0.381∗∗ 0.608∗∗∗ 0.362
(2-1)
(1.75) (2.30) (2.57) (1.39) (1.77) (2.29) (2.61) (1.47)
0.267∗ 0.422∗∗ 0.565∗∗ 0.298 0.28∗∗ 0.396∗∗ 0.584∗∗ 0.303
C4 α
(1.86) (2.40) (2.40) (1.19) (1.98) (2.38) (2.43) (1.20)
0.216 0.45∗∗ 0.619∗∗ 0.403 0.24 0.414∗∗ 0.666∗∗ 0.426
FF5 α
(1.31) (2.22) (2.37) (1.46) (1.50) (2.14) (2.50) (1.51)
0.256 0.447∗∗ 0.578∗∗ 0.323 0.29∗ 0.407∗∗ 0.636∗∗ 0.346
SY4 α
(1.58) (2.18) (2.19) (1.15) (1.80) (2.10) (2.32) (1.21)
73
Table G.2: Double Sorting: Money Managers’ Long Proportion Growth (%, per Week)
the MM Long Proportion Growth signal. The signal from the futures market is constructed with

AD 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.208 -0.004 -0.416 -0.623∗∗∗ 0.234 -0.013 -0.426 -0.66∗∗∗
1
(1.07) (-0.02) (-1.51) (-3.15) (1.28) (-0.07) (-1.57) (-3.31)
0.366∗∗ 0.451∗∗ 0.435∗ 0.069 0.314∗ 0.455∗∗ 0.426∗ 0.112
2
(2.00) (2.28) (1.65) (0.35) (1.87) (2.40) (1.66) (0.58)
0.154 0.451∗∗∗ 0.846∗∗∗ 0.691∗∗∗ 0.081 0.465∗∗∗ 0.873∗∗∗ 0.792∗∗∗
(2-1)
(0.90) (3.00) (3.34) (2.65) (0.48) (3.13) (3.50) (3.07)
0.159 0.425∗∗∗ 0.909∗∗∗ 0.75∗∗∗ 0.116 0.452∗∗∗ 0.931∗∗∗ 0.815∗∗∗
C4 α
(0.92) (2.81) (3.35) (2.67) (0.67) (3.00) (3.48) (2.92)
0.144 0.414∗∗ 0.891∗∗∗ 0.746∗∗ 0.104 0.44∗∗∗ 0.928∗∗∗ 0.824∗∗∗
FF5 α
(0.78) (2.51) (2.94) (2.40) (0.57) (2.68) (3.12) (2.66)
0.189 0.478∗∗∗ 0.902∗∗∗ 0.713∗∗ 0.165 0.476∗∗∗ 0.938∗∗∗ 0.774∗∗
SY4 α
(1.00) (2.91) (2.99) (2.30) (0.86) (2.93) (3.14) (2.48)
VOL 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.184 0.084 -0.415 -0.599∗∗∗ 0.097 0.062 -0.452∗ -0.549∗∗∗
1
(1.20) (0.39) (-1.57) (-3.29) (0.66) (0.28) (-1.72) (-3.00)
0.248 0.341∗ 0.5∗ 0.252 0.235 0.36∗ 0.556∗∗ 0.321
2
(1.58) (1.70) (1.79) (1.20) (1.51) (1.72) (1.99) (1.52)
0.064 0.257∗ 0.915∗∗∗ 0.851∗∗∗ 0.138 0.298∗ 1.008∗∗∗ 0.87∗∗∗
(2-1)
(0.50) (1.66) (3.81) (3.56) (1.11) (1.93) (4.01) (3.51)
0.068 0.263∗ 0.909∗∗∗ 0.841∗∗∗ 0.147 0.304∗ 0.991∗∗∗ 0.844∗∗∗
C4 α
(0.52) (1.68) (3.66) (3.38) (1.14) (1.96) (3.88) (3.32)
0.028 0.193 1.049∗∗∗ 1.021∗∗∗ 0.137 0.27 1.179∗∗∗ 1.042∗∗∗
FF5 α
(0.20) (1.12) (3.68) (3.65) (0.97) (1.55) (4.04) (3.65)
0.067 0.246 0.997∗∗∗ 0.93∗∗∗ 0.177 0.325∗ 1.137∗∗∗ 0.961∗∗∗
SY4 α
(0.45) (1.36) (3.49) (3.28) (1.22) (1.81) (3.82) (3.27)
74
LIQ 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
-0.025 -0.025 -0.079 -0.054 -0.024 0.076 -0.182 -0.157
1
(-0.14) (-0.11) (-0.34) (-0.31) (-0.15) (0.34) (-0.78) (-0.88)
0.308∗ 0.392∗ 0.411∗ 0.103 0.289∗ 0.449∗∗ 0.41 0.122
2
(1.70) (1.87) (1.67) (0.58) (1.73) (2.16) (1.64) (0.67)
0.333∗∗ 0.417∗∗ 0.49∗∗ 0.157 0.313∗∗ 0.374∗∗ 0.592∗∗ 0.279
(2-1)
(2.40) (2.39) (2.16) (0.65) (2.27) (2.25) (2.52) (1.11)
0.345∗∗ 0.425∗∗ 0.466∗∗ 0.12 0.338∗∗ 0.381∗∗ 0.562∗∗ 0.223
C4 α
(2.45) (2.44) (1.99) (0.48) (2.39) (2.28) (2.33) (0.86)
0.284∗ 0.425∗∗ 0.527∗∗ 0.242 0.303∗ 0.367∗ 0.643∗∗ 0.34
FF5 α
(1.78) (2.11) (2.02) (0.88) (1.92) (1.90) (2.40) (1.18)
0.341∗∗ 0.462∗∗ 0.496∗ 0.154 0.344∗∗ 0.399∗∗ 0.647∗∗ 0.303
SY4 α
(2.12) (2.27) (1.87) (0.55) (2.16) (2.06) (2.33) (1.03)
75
Table G.3: Double Sorting: Money Managers’ Short Proportion Growth (%, per Week)
the MM Short Proportion Growth signal. The signal from the futures market is constructed with

AD 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.409∗∗ 0.441∗∗ 0.546∗∗ 0.137 0.357∗ 0.439∗∗ 0.569∗∗ 0.212
1
(2.05) (2.15) (2.02) (0.66) (1.93) (2.18) (2.14) (1.01)
0.125 0.062 -0.37 -0.495∗∗∗ 0.15 0.056 -0.407 -0.556∗∗∗
2
(0.70) (0.32) (-1.39) (-2.74) (0.90) (0.30) (-1.58) (-3.06)
-0.29∗ -0.386∗∗ -0.927∗∗∗ -0.636∗∗ -0.207 -0.389∗∗ -0.96∗∗∗ -0.752∗∗∗
(2-1)
(-1.71) (-2.41) (-3.74) (-2.51) (-1.19) (-2.49) (-3.99) (-2.99)
-0.295∗ -0.371∗∗ -0.983∗∗∗ -0.688∗∗ -0.239 -0.374∗∗ -1.013∗∗∗ -0.775∗∗∗
C4 α
(-1.73) (-2.30) (-3.72) (-2.55) (-1.35) (-2.36) (-3.93) (-2.87)
-0.24 -0.333∗ -1.001∗∗∗ -0.761∗∗ -0.183 -0.337∗ -1.035∗∗∗ -0.852∗∗∗
FF5 α
(-1.32) (-1.89) (-3.39) (-2.57) (-1.00) (-1.94) (-3.62) (-2.88)
-0.3 -0.404∗∗ -0.952∗∗∗ -0.652∗∗ -0.26 -0.409∗∗ -1.003∗∗∗ -0.743∗∗
SY4 α
(-1.60) (-2.29) (-3.26) (-2.19) (-1.33) (-2.36) (-3.53) (-2.49)
VOL 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.25 0.386∗ 0.467 0.217 0.233 0.397∗ 0.491∗ 0.258
1
(1.57) (1.85) (1.62) (0.98) (1.48) (1.82) (1.73) (1.18)
0.192 0.022 -0.227 -0.418∗∗ 0.1 0.027 -0.255 -0.356∗∗
2
(1.26) (0.11) (-0.88) (-2.44) (0.69) (0.13) (-0.98) (-2.02)
-0.058 -0.364∗∗ -0.693∗∗∗ -0.635∗∗∗ -0.133 -0.369∗∗ -0.746∗∗∗ -0.613∗∗
(2-1)
(-0.45) (-2.44) (-2.94) (-2.68) (-1.06) (-2.41) (-3.10) (-2.55)
-0.061 -0.379∗∗ -0.672∗∗∗ -0.611∗∗ -0.137 -0.379∗∗ -0.729∗∗∗ -0.592∗∗
C4 α
(-0.46) (-2.52) (-2.76) (-2.50) (-1.06) (-2.46) (-2.99) (-2.41)
0.007 -0.335∗∗ -0.778∗∗∗ -0.785∗∗∗ -0.074 -0.353∗∗ -0.923∗∗∗ -0.849∗∗∗
FF5 α
(0.05) (-1.97) (-2.76) (-2.81) (-0.52) (-2.01) (-3.28) (-3.05)
-0.026 -0.388∗∗ -0.698∗∗ -0.673∗∗ -0.129 -0.417∗∗ -0.854∗∗∗ -0.725∗∗
SY4 α
(-0.17) (-2.21) (-2.51) (-2.40) (-0.88) (-2.27) (-3.05) (-2.57)
76
LIQ 1 2 3 (3 − 1) 1 2 3 (3 − 1)
Signal
0.275 0.402∗ 0.442∗ 0.167 0.262 0.462∗∗ 0.433∗ 0.171
1
(1.53) (1.85) (1.74) (0.90) (1.57) (2.14) (1.66) (0.90)
0.01 -0.003 -0.022 -0.032 -0.005 0.088 -0.12 -0.115
2
(0.06) (-0.01) (-0.10) (-0.20) (-0.03) (0.42) (-0.52) (-0.69)
-0.265∗ -0.405∗∗ -0.464∗∗ -0.199 -0.268∗∗ -0.374∗∗ -0.554∗∗ -0.286
(2-1)
(-1.93) (-2.49) (-2.05) (-0.84) (-1.99) (-2.41) (-2.34) (-1.19)
-0.28∗∗ -0.414∗∗ -0.433∗ -0.153 -0.293∗∗ -0.382∗∗ -0.521∗∗ -0.228
C4 α
(-1.99) (-2.52) (-1.87) (-0.63) (-2.13) (-2.43) (-2.16) (-0.93)
-0.211 -0.444∗∗ -0.383 -0.172 -0.245 -0.396∗∗ -0.521∗ -0.276
FF5 α
(-1.32) (-2.34) (-1.46) (-0.64) (-1.58) (-2.16) (-1.92) (-1.01)
-0.258 -0.468∗∗ -0.363 -0.105 -0.3∗ -0.42∗∗ -0.539∗ -0.239
SY4 α
(-1.64) (-2.47) (-1.37) (-0.38) (-1.93) (-2.31) (-1.91) (-0.85)
77
H. Contributions of Portfolio Return by Commodity
Figure H.I: Contributions of the returns of the No-Weight Short portfolio by commodity, in
the years 2015-2016, utilizing the signal of MA(2) Managed Money Net Change
Notes: This figure focus on the years 2015 and 2016 only. The right axis of the figure plots the
Short portfolio’s cumulative weekly returns in dollar amount assuming $1 was invested in August
2006, with returns equal-weighted across firms within a commodity equity-based portfolio, then
equal-weighted across all commodity equity-based portfolios traded, signaled by the MA(2) MM Net
Change measure. The portfolio are constructed per Section 3.2 following the Wednesday-to-Tuesday
convention. On the left axis, this figure decomposes the weekly contributions in % of each commodity
equity-based portfolio to the return of the Short portfolio. For instance, the Short Portfolio realizes,
in the week 31 of 2015, a weekly return of -0.53%. During that week, 7 commodity-equity portfolios
were signaled short and traded as part of the Short portfolio (while coal, petroleum and steel were
signaled long). Among the shorted commodity-equity portfolios, 4 commodities realized negative
returns, on average -2.52% (silver, oil & gas, biofuel and copper contributed to 38.1%, 36.3%, 17.5%,
8.1% of this negative average, respectively) and 3 commodities demonstrated positive returns, on
average 1.99% (gold, lumber and miscmetal contributed to 62.2%, 33.9% and 3.9% of this positive
average, respectively). Together these 7 commodities yield a weekly return of -0.53% for the Short
portfolio.
20% 0.6
15%
Cumulative return in $, of 1$ invested in August 2006

0.5
10%
5% 0.4
Return in %
0%
0.3
-5%
-10% 0.2
-15%
0.1
-20%
-25% 0
2015w7
2015w1
2015w4
2015w40
2016w3
2016w6
2016w9
2016w27
2015w10
2015w13
2015w16
2015w19
2015w22
2015w25
2015w28
2015w31
2015w34
2015w37
2015w43
2015w46
2015w49
2015w52
2016w12
2016w15
2016w18
2016w21
2016w24
2016w30
2016w33
2016w36
2016w39
2016w42
2016w45
2016w48
2016w51
Biofuel Coal Copper Gold

Lumber MiscMetal Oil & Gas Petroleum
Silver Steel Short Portfolio (right axis)
78
Figure H.II: Contributions to the returns of the No-Weight Long portfolio by commodity, in
the years 2015-2016, utilizing the signal of MA(2) Managed Money Net Change
Notes: This figure focus on the years 2015 and 2016 only. The right axis of the figure plots the
Long portfolio’s cumulative weekly returns in dollar amount assuming $1 was invested in August
2006, with returns equal-weighted across firms within a commodity equity-based portfolio, then
equal-weighted across all commodity equity-based portfolios traded, signaled by the MA(2) MM Net
Change measure. The portfolio are constructed per Section 3.2 following the Wednesday-to-Tuesday
convention. On the left axis, this figure decomposes the weekly contributions in % of each commodity
equity-based portfolio to the return of the Long portfolio. For instance, the Long Portfolio realizes,
in the week 45 of 2016, a weekly return of 1.46%. During that week, 6 commodity-equity portfolios
were signaled long and traded as part of the Long portfolio (the other 4 commodities were signaled
short). Among the longed commodity-equity portfolios, 4 commodities realized positive returns, on
average 2.48% (silver, copper, biofuel and steel contributed to 45.7%, 30.9%, 15.6% and 7.8% of
this positive average, respectively) and 2 commodities demonstrated negative returns, on average
-1.02% (petroleum and gold contributed to 54.4% and 45.6% of this negative average, respectively),
and together the 6 commodities yield a weekly return of 1.46% for the Long portfolio.
20% 7
15%
Cumulative return in $, of 1$ invested in August 2006

6
10%
5
5%
Return in %
0% 4
-5% 3
-10%
2
-15%
1
-20%
-25% 0
2015w1
2015w4
2016w3
2015w7
2015w10
2015w13
2015w16
2015w19
2015w22
2015w25
2015w28
2015w31
2015w34
2015w37
2015w40
2015w43
2015w46
2015w49
2015w52
2016w6
2016w9
2016w12
2016w15
2016w18
2016w21
2016w24
2016w27
2016w30
2016w33
2016w36
2016w39
2016w42
2016w45
2016w48
2016w51
Biofuel Coal Copper Gold

Lumber MiscMetal Oil & Gas Petroleum
Silver Steel Long Portfolio (right axis)
79

Paper KkfN9BAN

Uploaded by

Copyright:

Available Formats

Paper KkfN9BAN

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Paper KkfN9BAN

Uploaded by

Copyright:

Available Formats

Is there smart money?

How information in the futures market is

Steven Wei Ho∗ Alexandre R. Lauwers§

First draft: November 17th, 2017

JEL codes: G11, G14

Keywords: Return Predictability, Limits to Arbitrage, Commodity, Market Seg-

2. Background and Literature

3. Data and Estimation

3.1. Data on Stocks and Commodity Futures

3.1.1. Matching Futures Market Positions with Producers’ Stocks

Markets Participants Description

An entity that deals primarily in swaps for a commodity

Entities that manage and conduct organized futures trading

3.2. Data Manipulation and Portfolio Formation

The Long-Short portfolio is constructed by sorting the commodity-equity portfolios weekly

Long portfolio Short portfolio Long-Short portfolio

Long portfolio Short portfolio Long-Short portfolio

MM Long-Short Ratio Growth

MM Long Proportion Growth

MM Short Proportion Growth

Sample period : August 2006 - September 2018

Money Managers’ Net Change

Panel A: Whole sample Panel B: Without Recession

Money Managers’ Long Short Ratio Growth

Panel A: Whole sample Panel B: Without Recession

Money Managers’ Short Proportion Growth

4.3. Fama-MacBeth Analysis

Money Managers’ Net Change

1 -0.198 (-0.88) -0.09 (-0.46)

Money Managers’ Long-Short Ratio Growth

1 -0.135 (-0.61) -0.065 (-0.33)

Money Managers’ Long Proportion Growth

1 -0.172 (-0.79) -0.066 (-0.34)

Money Managers’ Short Proportion Growth

1 0.457∗∗ (2.16) 0.383∗∗ (2.01)

4.4. Further Discussions

4.4.1. Jensen’s Alpha Analysis for the Monday-to-Friday Convention

4.4.2. Additional Results

4.5. Double-Sorting Results and Relationship with Market Fric-

Panel A: Equal-Weight Panel B: Value-Weight

Basak, Suleyman and Pavlova, Anna. (2016). ‘A model of financialization of commodities’,

White, Halbert. (1980). ‘A heteroskedasticity-consistent covariance matrix estimator and a

Then, we compound rtC to obtain weekly returns Rw C

Table B.1: Sample of Commodity Producers

Number of stocks where SHRCD:

Copper 1020; 1021; 3331 Copper Ores; 3 4 7 12

Total = 127 = 217 = 193 = 337

Sample period : August 2006 - September 2018

Net Long position of traders as percent of open interest

Money Managers Producers, Merchants, Processors and Users Swap Dealers

Long portfolio Short portfolio Long-Short portfolio

Money Managers’ Net Change

Money Managers’ Long Short Ratio Growth

Money Managers’ Short Proportion Growth

Open Interest Growth

Money Managers’ Net Change

Money Managers’ Long Short Ratio Growth

Money Managers’ Short Proportion Growth