Large Bets and Stock Market Crashes

Review of Finance, 2023, 1–41
https://doi.org/10.1093/rof/rfad008
Advance Access Publication Date: 28 March 2023
Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Large Bets and Stock Market Crashes
Albert S. Kyle1,* and Anna A. Obizhaeva2
1
Robert H. Smith School of Business, University of Maryland, College Park, MD, USA and 2New
Economic School, Moscow, Russia
Abstract
Some market crashes occur because of significant imbalances in demand and sup-
ply. Conventional models fail to explain the large magnitudes of price declines. We
propose a unified structural framework for explaining crashes, based on the insights
of market microstructure invariance. A proper adjustment for differences in business
time across markets leads to predictions which are different from conventional wis-
dom and consistent with observed price changes during the 1987 market crash and
the 2008 sales by Société Générale. Somewhat larger-than-predicted price drops
during 1987 and 2010 flash crashes may have been exacerbated by too rapid selling.
Somewhat smaller-than-predicted price decline during the 1929 crash may be due
to slower selling and perhaps better resiliency of less integrated markets.
Keywords: Crashes, Liquidity, Price impact, Market depth, Systemic risk, Market microstructure,
Invariance
JEL classification: G01, G28, N22
Received September 7, 2019; accepted March 13, 2023 by Editor Alex Edmans.
After stock market crashes, market participants, policymakers, and economists are usually
unable to explain what happened. Heavy selling pressure has been recorded during crashes.
It is known that large imbalances move prices in the direction of trades, as discussed by
Kraus and Stoll (1972), Grinold and Kahn (2000), and Gabaix (2009), but there is no com-
pelling quantitative explanation for why the seemingly small quantities sold might have led
to such large price dislocations in the highly liquid stock market.
*A.S.K. was a staff member of the Task Force on Market Mechanisms (“Brady Commission”) in 1987–
88, worked on a research project examining high-frequency trading in the CME S&P 500 E-mini futures
market for the Commodity Futures Trading Commission during 2009–10, and has worked with the
Securities and Exchange Commission, the Department of Justice, and the Federal Reserve Bank of
Atlanta. He has been a member of the CFTC Technology Committee and is a member of the FINRA
Economic Advisory Board. From 2010–22, he was an independent member of the Board of Directors of
an asset management company which trades equities on behalf on institutional investors.
C The Author(s) 2023. Published by Oxford University Press on behalf of the European Finance Association. All rights
V
reserved. For permissions, please email: [email protected]
2 A. S. Kyle and A. A. Obizhaeva
We investigate this issue through the lens of market microstructure invariance, a concep-
tual framework developed by Kyle and Obizhaeva (2016). By analyzing prices and quanti-
ties in market-specific “business time,” defined as time which is measured by counting
market events rather than seconds on a clock, we are able to explain why seemingly small
observed imbalances in demand and supply can indeed have created such large market

crashes.
We illustrate our approach by studying five crash events, chosen because data on the
magnitude of contemporaneous selling pressure became publicly available in their
aftermaths:
• After the stock market crash of October 1929, the Senate Committee on Banking and
Currency (1934) (the “Pecora Report”) documented the plunge in broker loans attrib-
uted to forced margin selling during the crash.
• After the October 1987 stock market crash, the U.S. Presidential Task Force on Market
Mechanisms (1988) (the “Brady Report”) reported quantities of stock index futures
contracts and baskets of stocks sold by portfolio insurers during the crash.
• After the futures market dropped by 20% at the open of trading 3 days after the 1987
crash, the Commodity Futures Trading Commission, Division of Economic Analysis
and Division of Trading and Markets (1988) identified large sell orders executed at the
open of trading; the press identified the seller as George Soros.
• After the Fed cut interest rates by 75 basis points in response to a worldwide stock mar-
ket plunge on January 21, 2008, Société Générale revealed that it had been quietly liqui-
dating billions of Euros in stock index futures positions accumulated earlier by rogue
trader Jérôme Kerviel.
• After the flash crash of May 6, 2010, the Staffs of the CFTC and SEC (2010a, 2010b)
cited as its trigger the large sales of futures contracts by one entity, later identified in the
press as Waddell & Reed.
We do not study the flash crash events in 1961 and 1989, the collapse of Long Term
Capital Management in 1998, the quant meltdown of August 2007, or the U.S. Treasury
note flash rally in October 2016 because data on the size of sales which precipitated these
events are not available.
Each of the five crashes is associated with a large sell bet, where we think of a “bet” (or
“meta-order”) as being a statistically independent decision either to speculate on informa-
tion or to hedge risks by buying or selling significant quantities of risky financial assets,
often implemented as sequences of orders executed over time. These bets resulted either
from trading by one entity or from correlated trading of multiple entities with the same
underlying motivation.
Many practitioners and academics believe that during these five crashes, selling pressure
was too small to induce significant price declines. We call this interpretation “conventional
wisdom.” Scholes (1972), Harris and Gurel (1986), and Wurgler and Zhuravskaya (2002)
illustrate conventional wisdom by claiming that the demand for financial assets is elastic in
the sense that selling 1% of the asset’s market capitalization has a price impact of less than
1%. When applied to market bets during the five crashes, this conventional thinking
implies tiny market impact. Indeed, given typical turnover rates of stocks, this implies that
sales of 5% of average daily volume are expected to have only modest impact on stock pri-
ces. Naively extrapolating these estimates to the highly liquid market for stock index
Large Bets and Stock Market Crashes 3
futures contracts further suggests that selling 5% of daily volume of the entire stock market
must have even smaller impact on the overall level of stock prices.
Microstructure invariance implies a different way to extrapolate price impact estimates
from stocks to index futures. This approach implies much larger price impacts of bets in li-
quid futures markets because it models trading in market-specific business time, not calen-

dar time.
The main intuition is as follows. As we show below, invariance principles imply that
business time passes about 225 times faster in the equities market as a whole than in mar-
kets for less liquid individual stocks. One calendar day of trading in stock index futures is
thus equivalent to 225 calendar days of trading in a stock. Hence, a sell bet of 5% of 1
day’s volume for index futures is analogous to selling 5% of daily volume each day for 225
consecutive calendar days (not 1 day) for a stock or a bet of 1125% (¼ 5% 225) of 1
day’s volume. A bet of 5% of 1 day’s volume for index futures therefore must have much
bigger (not smaller) price impact than a bet of 5% of 1 day’s volume for less liquid individ-
ual stocks. This argument goes against the standard intuition.
We calculate price impacts implied by quantities traded during crashes using a universal
formula for market impact, as suggested by invariance. The price impact is a function of
the dollar size of a bet, expected dollar volume, returns volatility, and a couple of invariant
parameters. Kyle and Obizhaeva (2016) calibrate these parameters using a database of
about 400,000 portfolio transition orders executed during the period 2001–05 in US stocks.
Portfolio transition orders are well suited for calibration of market impact functions be-
cause they can be thought of as exogenous shocks to demand and supply. In this article, we
extrapolate the estimates from the sample of relatively small individual US stocks to the
large US stock market as a whole and from 2001 to 2005 to other historical periods.
Except for the time frame of execution, the spirit of invariance suggests that institutional
details related to market structure, information asymmetries, or motivation of traders
should not affect market impact estimates much. We show that the implied estimates are
indeed large enough to explain crashes.
Table I summarizes our results for each of the five crash events. The table shows the ac-
tual percentage decline in market prices, the percentage decline predicted by invariance, the
percentage decline predicted by conventional wisdom, the dollar amount sold as a fraction
of average daily volume, and the dollar amount sold as a fraction of 1 year’s GDP.
Table I. Summary of five crash events
Actual and predicted price declines. The table columns show the actual price changes, pre-
dicted price changes, and bets as percent of average daily volume and GDP. The predictions
implied by invariance are much closer to actual price changes than those implied by conven-
tional wisdom.
Crash Event Actual Predicted Predicted %ADV %GDP

(%) invariance (%) conventional (%)
1929 market crash 25 46.43 1.36 265.41 1.136

1987 market crash 32 16.77 0.63 66.84 0.280
1987 Soros’s trades 22 6.27 0.01 2.29 0.007
2008 SocGén trades 9.44 10.79 0.43 27.70 0.401
2010 flash crash 5.12 0.61 0.03 1.49 0.030
We assume that the market impact of a bet is equal to the percentage of market capitaliza-
tion sold when calculating the estimates implied by conventional wisdom.
For all events, the price impact estimates based on conventional wisdom are minuscule
in comparison to actual price changes. Other estimates obtained from the analysis of
trades of institutional investors—such as studies by Grinold and Kahn (2000), Torre

(1997), Almgren et al. (2005), and Frazzini, Israel, and Moskowitz (2018)—produce
estimates higher than conventional intuition but far too small to explain crashes (reported
later). In contrast, predictions based on invariance are broadly similar to actual price
declines.
Across the five crash events, there are substantial differences between actual declines
and predictions based on invariance. The predictions of invariance do not take into account
the speed with which sales took place or the fragility of markets at the times of the crashes.
At the same time, the speed of sales may have influenced the magnitudes of price
declines. In 1929, efforts were made to spread the impact of margin selling out over several
weeks rather than several days. This may have made the actual price decline of 25% smaller
than the forecast of 46.43%. The 1987 Soros trades and the 2010 flash crash were both
“flash-crash” events in which prices declined rapidly and then recovered within minutes.
The unusually rapid rate at which bets were executed may have magnified temporary price
impact.
Variation in market integration across assets, availability of capital available to take the
other sides of large sales, and disruptions to the market mechanism may also help explain
differences between actual and predicted price declines. In 1929, smaller-than-predicted
price declines may have been reduced by markets being less integrated than today. Potential
buyers were keeping capital on the sidelines to profit from price declines widely expected to
occur if margin purchases were liquidated. In 1987, larger-than-predicted price declines
may have been exacerbated by breakdowns in the market mechanism documented in the
Brady Report.
These five bet-induced crashes differ from macroeconomic crises with sovereign
defaults, bank failures, exchange rate collapses, and bouts of high inflation catalogued by
Reinhart and Rogoff (2009). Recovery from economic crises in contrast takes many years,
even after significant changes in macroeconomic policies and market regulation. Bet-
induced crashes are likely to be short-lived, especially if followed by appropriate govern-
ment policy. For example, the looser monetary policy implemented by Federal Reserve
System immediately after the 1929 crash calmed down the market by the end of 1929. Even
though the wealth effect of declining equity prices may have helped trigger a recession by
reducing consumption, Friedman and Schwarz (1963) write that the Great Depression of
the 1930s resulted from a subsequent shift toward a deflationary monetary policy, not from
the 1929 crash itself. Similarly, the unwinding of Jérôme Kerviel’s large rogue bet in
January 2008 was followed by the collapse of Bear Stearns a few weeks later, but the deep
and long-lasting recession which unfolded in 2008–09 was triggered by the bursting of the
real-estate credit bubble, not from liquidation of his bet.
Unable to find rational quantitative explanations, some researchers believe that market
crashes result from irrational behavior. The “animal spirits” hypothesis of market crashes
says that price fluctuations occur as a result of random changes in psychology and emo-
tions, which may not be based on economically relevant information or rational calcula-
tions. Keynes (1936) said that financial decisions may be taken as the result of “animal
spirits—a spontaneous urge to action rather than inaction, and not as the outcome of a
weighted average of quantitative benefits multiplied by quantitative probabilities.” Akerlof

and Shiller (2009) echo Keynes (1936): “To understand how economies work and how we
can manage them and prosper, we must pay attention to the thought patterns that animate
people’s ideas and feelings, their animal spirits.” Promptly after the 1987 crash, Shiller
(1987) surveyed traders and found that “most investors interpreted the crash as due to the

psychology of other investors.”
In contrast, we believe that one does not need to invoke irrationality to explain crashes.
Before the 1929 crash, market participants widely discussed the possibility that forced liqui-
dations of margin accounts would lead to a collapse in prices. Before the 1987 crash, mar-
ket participants discussed that portfolio insurance sales might lead to a market meltdown.
In the absence of quantitative justification, these prescient views were largely dismissed due
to a deeply entrenched ideological belief that the demand for equities is elastic. Here, we
provide quantitative justification for a theory of crashes based on the market impact of
large bets, not based on psychology. The margin sales of 1929, portfolio insurance sales of
1987, and liquidation of Kerviel’s rogue positions were all large bets resulting from the
rapid execution of mechanical trading strategies, not from psychology. While the sales of
George Soros in 1987 and sales of Waddell and Reed in 2010 may reflect the animal spirits
of one person and one entity, the immediate recovery of prices suggests the opposite of
market-wide irrationality or psychological contagion.
The remainder of this article discusses the conventional wisdom in assessing market
impact, market microstructure invariance, particulars of each of the crash events, and
lessons learned.
1. Market Impact of Large Bets: Previous Literature

Previous studies have used different methodologies to obtain widely varying estimates of
the market impact of large bets. This literature can be divided into two strands. The first
strand, which we call “conventional wisdom,” examines the price effects of seasoned equity
offerings, changes in the composition of the S&P 500 index, and similar events. The second
strand examines the price impact of trades by institutional investors.
1.1 Conventional Wisdom

Scholes (1972) claimed that the price impact of large sales of equities is negligible based on
his analysis of secondary equity distributions. Harris and Gurel (1986), Wurgler and
Zhuravskaya (2002), and others study the price response to additions and deletions of
stocks to equity indices like the S&P 500 and infer that selling 1% of an individual stock’s
shares outstanding has a price impact of at most 1%. Wurgler and Zhuravskaya (2002,
Table IV, p. 603) provide a summary of demand elasticities from different papers, all of
which suggest an elastic demand for stocks; estimated demand elasticities vary from 1 to
3,000, an almost infinite elasticity from Scholes’s study.
When extrapolated from the market for individual stocks to the stock market as a whole
using the same demand elasticity, these empirical studies support the conventional wisdom
that observed selling pressure could not have created stock market crashes.
From a theoretical perspective, the conventional wisdom is based on the logic of perfect-
ly competitive capital markets, the capital asset pricing model, and the efficient markets hy-
pothesis. The market risk premium of 5–7% per year reflects compensation for bearing the
risk of the entire stock market for 1 year. When large bets are executed, market participants
taking the other side of these bets are exposed to risks of much smaller magnitude than the
entire stock market and hold positions over much shorter horizons than 1 year, usually a
few days or minutes. Hence, the compensation required for absorbing large bets should be
dramatically smaller than the equity market risk premium.
Indeed, when conventional wisdom was applied to the crash of 1987, prominent finan-

cial economists claimed that the price impact of reported sales was up to 100 times too
small to generate a crash. Leland and Rubinstein (1988), the academics most closely associ-
ated with portfolio insurance in 1987, say, “To place systematic portfolio insurance in per-
spective, on October 19, portfolio insurance sales represented only 0.2% of total U.S. stock
market capitalization. Could sales of 1 in every 500 shares lead to a decline of 20% in the
market? This would imply a demand elasticity of 0.01—virtually zero—for a market often
claimed to be one of the most liquid in the world.” Miller (1991) makes similar claims
about the 1987 crash: “Putting a major share of the blame on portfolio insurance for creat-
ing and overinflating a liquidity bubble in 1987 is fashionable, but not easy to square with
all relevant facts. . . . No study of price–quantity responses of stock prices to date supports
the notion that so large a price decrease (about 30%) would be required to absorb so mod-
est (1%–2%) a net addition to the demand for shares.” Using a calibrated theoretical model
of competitive capital markets, Brennan and Schwartz (1989) claimed that portfolio insur-
ance sales (of 0.63% of market capitalization) in the 1987 crash would have a price impact
of about 100 times smaller than the 32% price drop observed.
Since price pressure was thought to be too small to explain quantitatively market
crashes, some observers of the 1987 stock market crash, including Miller (1988, p. 477)
and Roll (1988), sought to explain the large price declines as market reactions to new fun-
damental information rather than response to trading, but it was difficult to find new infor-
mation to which market prices would have reacted so dramatically.
In our analysis below, we summarize conventional wisdom using the least conservative
conventional estimate and assume a unit demand elasticity for stocks: Selling 1% of capital-
ization moves prices down by 1%. Mathematically, suppose a stock’s price is P, outstand-
ing shares are N, and shares sold are Q. Then, the expected log-percentage market impact
DlnP is Q/N:
DP Q
DlnP ¼ : (1)
P N
Throughout this article, we adopt the convention that Q is the unsigned trade size and
DP=P is the expected unsigned price impact.1
The conventional market impact function can be also expressed in terms of average
daily volume. Let V denote daily volume in shares. Assume for simplicity that an asset’s
turnover is approximately 100% per year with 250 trading days. Since 1% of capitalization
is approximately equal to 250% of daily volume, the conventional wisdom [Equation (1)]
can be interpreted as
1 The size of market impact DlnP is either the expectation of the post-trade log-price minus pre-
trade log-price for buy bets or the expectation of the pre-trade log-price minus post-trade log-price
for sell bets. A similar formula can be written for simple percentage impact DP =P , where DP is ei-
ther the difference between post-trade price and pre-trade price for buy bets or the difference be-
tween pre-trade price and post-trade price for sell bets.
DP Q
DlnP ¼ : (2)
P 250 days V
The Brady Report used this intuition to compare daily volume elasticities in the 1987 crash
to the 1929 crash:
To account for the contemporaneous 28% decline in price, this implies a price elasticity of 0.9

with respect to trading volume which seems unreasonably high. As a percentage of total shares
outstanding, margin-related selling would have been much smaller. Viewed as a shift in the
overall demand for stocks, margin-related selling could have accounted realistically for no more
than 8% of the value of outstanding stock. On this basis, the implied elasticity of demand is 0.3
which is beyond the bound of reasonable estimates.
1.2 Estimates from Institutional Trades

The finance literature which studies price responses to large institutional bets typically finds
that the demand for stocks is less elastic than what conventional wisdom suggests.
Kraus and Stoll (1972) study block trades of large NYSE stocks. More recent estimates
of market impact from executions of large orders by institutional investors include Chan
and Lakonishok (1995, 1997) and Keim and Madhavan (1997).
Some studies find nonlinear price impact. The “square root model” or Barra model,
described by Grinold and Kahn (2000) and Torre (1997), says that the execution of an
order of size Q on average moves price by DP=P ¼ 1 r ðQ=VÞ1=2 . The square root model
implies that an order for 25% of 1 day’s volume in a stock with 2% daily volatility implies
a price impact of 100 basis points. With 100% annual turnover, this implies an elasticity of
demand of 0.10.
Frazzini, Israel, and Moskowitz (2018) estimate a more complicated version of the
square root model. Almgren et al. (2005) incorporate execution horizon into a model with
concave price impact similar to a square root model.
While less elastic demand from these models implies more price impact than convention-
al wisdom, a demand elasticity of only 0.10 is nevertheless not small enough to generate
crashes; this is shown in Appendix B, which extrapolates the estimates of these models to
crash events. For crashes to result from selling pressure, the elasticity needs to be approxi-
mately 0.01, yet another order of magnitude more inelastic than implied by these studies.
The square root model makes it especially difficult to explain crashes because its implied
concave price impact makes marginal price impact decrease as the size of large bets
increases. To explain crashes, we use a linear price impact model, which is popular with fi-
nance theorists because it excludes simple forms of arbitrage (Huberman and Stanzl, 2004).
2. Market Impact of Large Bets: Invariance

Invariance implies an alternative methodology for extrapolating price impact from the less
liquid markets for individual stocks to the more liquid market for indexes of all stocks.
While we agree with demand elasticities estimated for individual stocks in studies of institu-
tional trades, we suggest an alternative approach for how to extrapolate these estimates to
the stock market as a whole. By taking into account differences in business time, invariance
can explain stock market crashes.
2.1 Review of Invariance

Invariance is based on the simple intuition that trading in a speculative market is a game in
which financial risks are exchanged in business time. The speed of business time varies sig-
nificantly across assets. It is proportional to the rate at which new bets—or trading ideas—
arrive. Trading is fast in liquid markets and slow in illiquid markets.

Invariance consists of two conjectures: (1) The distribution of standard deviations of
dollar gains and losses on bets is the same across markets, when standard deviation is meas-
ured in units of business time.2 (2) The expected dollar costs of executing equivalent bets
are constant across markets, when equivalent bets are defined to transfer the same dollar
risks per unit of business time. These invariance conjectures imply specific scaling laws for
financial variables.
We derive these scaling laws using a simplified version of Kyle and Obizhaeva (2016).3
Kyle and Obizhaeva (2020) show how to obtain the same scaling laws in the context of an
equilibrium model of speculative trading with endogenous acquisition of private informa-
tion and endogenous entry into the market. In this model, scaling laws for bet sizes and
transaction costs ultimately follow from the assumption that the effort required to generate
private signals does not vary across markets. This is likely to hold, at least approximately,
in an equilibrium where traders allocate their skills optimally across markets.
2.2 Business Time

For a given stock, suppose that bets of average size Q arrive at rate c. For a typical stock,
we might have c ¼ 100 bets per day and Q ¼ 10; 000 shares. As c increases, market partici-
pants transfer risks more quickly and business time passes at a faster rate relative to calen-
dar time.
Since individual bets are difficult to observe, it is hard to measure c and Q in practice.
Nevertheless, the invariance hypotheses make it possible to infer c and Q from daily dollar
volume P V and daily returns volatility r up to a constant which is the same for all stocks.
The proof is based on two simple equations.
First, define trading activity W as the product of dollar volume and returns volatility:
W :¼ P V r: (3)
Trading activity better reflects the rate at which the market transfers risks than dollar vol-
ume P V because it takes into account that trading assets with higher volatility r transfers
more risk per dollar traded. Since all bets sum up to volume, V ¼ c Q, we can write W in
terms of c and Q:
W ¼ c Q P r: (4)
Since dollar volume P V has units of dollars=day and returns volatility r has units per
day1=2 , trading activity W has units dollars=day3=2 .
2 This conjecture does not say that dollar returns volatility or returns volatility are constant in busi-
ness time.
3 Kyle and Obizhaeva (2018b) obtain similar predictions using dimensional analysis and leverage neu-
trality. Kyle and Obizhaeva (2018a) derive them from a meta-model, a system of simple equations in-
herent to many microstructure models. Kyle, Obizhaeva, and Wang (2018) provide illustration using
a one-period equilibrium model.
Second, the first invariance conjecture says that dollar risk P r transferred by an aver-
age bet of Q shares per unit of business time 1=c is invariant across markets. Thus, for
some dollar constant C, such as C ¼ $2000, we have
Pr
Q pffiffiffi ¼ C: (5)
c

Equations (4) and (5) make up a system of two log-linear equations in two unknowns a :¼
P Q r and c:
ac¼W and a c1=2 ¼ C: (6)
The solution for a and c is

1=3 2=3
W W
a¼C and c¼ : (7)
C C
Now define H :¼ 1=c as the time interval between bets. For example, if c ¼ 100 bets per
day for some stock, then H is about 4 min during trading hours from 9:30 a.m. to 4:00
p.m. Equation (7) implies that average dollar bet size PQ and time between bets H are
given by
1=3 2=3
C W W
PQ ¼ and H :¼ 1=c ¼ ; where W :¼ P V r: (8)
r C C
Since W has units of dollars=day3=2 and C has units of dollars, Equation (8) has correct
units of dollars for P Q and days for H. Equation (8) shows how to extrapolate the size
and number of bets from one stock to another, under the invariance assumption that C is
constant across markets.
Define a benchmark stock as a security with stock price P ¼ $40 per share, expected
volume V ¼ 106 shares per calendar day, expected percentage returns volatility r ¼ 0:02
per day1=2 , and trading activity W ¼ P V r ; these parameters would approximately
correspond to a stock from the bottom of the S&P 500 index. Suppose the time interval be-
tween bets for this stock is H minutes, say, H is about 4 min. Equation (8) implies that H
is inversely proportional to the two-third power of trading activity,
2=3 2=3
1 1 W 1 PV r
¼ ¼ : (9)
H H W H P V r
Business time H represents different lengths of calendar time for different assets: 4 min
for the benchmark stock, an hour for thinly traded stocks, less than 1 min for
actively traded stocks, and about 1 s for the market as a whole. This is derived by
plugging values for volume and volatility into the equation above. Business time differs
across assets.
The conventional wisdom makes the mistake of extrapolating from one market to
another under the implicit assumption that business time H is constant across markets.
2.3 Distribution of Bet Size

~ not
The logic of invariance can be applied to the entire distribution of random bet sizes Q,
~
just the means Q. This logic implies that probability distributions of bet sizes Q must be
the same across markets if Q ~ is scaled by trading volume per business day V H, rather
than by trading volume per calendar day V. When bet size Q ~ is scaled by VH, Equation (9)
~
implies that the resulting scaled bet size Z has a mean of 1 and the same distribution for all
stocks:
2=3
~ d
Q ~
Q W

~
¼Z which implies
¼H ~
Z: (10)
VH V W
Equivalent bets transfer the same dollar risks in business time. In calendar time, equivalent
bets correspond to a smaller fraction of daily volume in markets with larger trading activity
and thus shorter time interval between bet arrivals.
2.4 Price Impact

The logic of invariance can also be extended to market impact. Think of bets in two differ-
ent markets as equivalent if they have the same scaled size Z ~
~ ¼ Q=ðVHÞ. Bet cost invari-
ance conjectures that equivalent bets have the same price impact when scaled by returns
volatility in business time.
pffiffiffiffiffi
Let DP=P denote price impact of a bet of size Q and let r H denote volatility in business
time. Bet cost invariance therefore implies an invariant price impact function f ðÞ such that
DP pffiffiffiffiffi Q
¼ r H f ðZÞ; where Z¼ : (11)
P VH
If the price impact function is modeled as a power function f ðZÞ ¼ a Zb with proportion-
ality constant a and exponent b, then Equation (11) takes the form
b
DP pffiffiffiffiffi Q
¼ar H : (12)
P VH
Plugging in H from Equation (9) and assuming linear market impact (b ¼ 1) yields
1=3
DP W Q PVr
¼a r ; where W :¼ : (13)
P C V C
In comparison with conventional intuition [Equation (2)] that bets of the same fraction of
daily volume Q/V must have the same percentage price impact, holding volatility r con-
stant, the linear specification [Equation (13)] has the additional factor ðW=CÞ1=3 , which
shows up due to the faster pace of business time in markets with higher trading activity.
This factor makes the demand more inelastic as trading activity increases. For example,
increasing W by a factor of 1,000 decreases the elasticity of demand by a factor of 10 and
reduces demand elasticity from say 0.10 to 0.01. As we shall see, if the market as a whole is
viewed as one big market with very fast business time, this makes the demand elasticity for
the market as a whole low enough for observed order imbalances to explain the average
size of price declines in market crashes.
2.5 Intuition behind Equivalent Bets

We next compare the magnitude of selling pressure during five market crashes with the
sizes of large institutional orders executed in US equities.
As a yardstick for measuring the size of institutional orders, we use the largest orders
from the 400,000þ portfolio transition orders studied by Kyle and Obizhaeva (2016). A
portfolio transition occurs when assets managed by one institutional asset manager are
transferred to another manager. Trades converting the legacy portfolio into the new port-
folio are typically handled by a professional third-party transition manager. Portfolio tran-
sitions represent some of the largest changes in portfolios held by institutional investors
during the year. We estimate the distributions of portfolio transition orders to be symmetric

around zero with unsigned order sizes close to log-normal random variables with different
log-means and the same log-variance of 2.53:

~
Q 2 W
ln N ð5:71 ln ; 2:53Þ: (14)
V 3 W
~
The empirical distribution of lnðQ=VÞ has a slope of 2=3 with respect to change in the log
of trading activity lnðWÞ, as expected given prediction (10).
Figure 1 summarizes the intuition of our article in one snapshot. It shows two types of
extrapolations across markets; one is based on conventional wisdom and another is based
on invariance. The vertical axis is the log of order size as a fraction of daily volume
lnðQ=VÞ. The horizontal axis is the log of scaled trading activity lnðW=W Þ. The point
lnðW=W Þ ¼ 0 corresponds to the benchmark stock with trading activity W from the bot-
tom of the S&P 500 index. Trading activity varies from lnðW=W Þ ¼ 12 for the least ac-
tively traded stocks to lnðW=W Þ ¼ 2:00 for the most actively traded stocks such as Apple,
by a factor of about 106 ð¼ expð12 þ 2ÞÞ. Trading activity in the overall stock market—
which includes both stocks and futures trading—is much higher, up to lnðW=W Þ ¼ 8:20;
it is about 500 ð¼ expð8:20 2:0ÞÞ times larger than the trading activity of the most liquid
stocks.
The (black) horizontal lines show the extrapolation direction implied by the conven-
tional wisdom. These isoquants mark orders of sizes equal to a given percentage of
calendar-day volume. For example, the horizontal line jQ=Vj ¼ 5% represents orders
equal to 5% of daily volume; conventional wisdom suggests that these bets will have the
same impact.
The diagonal (red and green) lines with slopes of 2=3 show equivalent bets as implied
by the theory of invariance. The lowest diagonal (red) line identifies log-medians of Q/V
for different markets, as implied by Equation (14); this line intersects the vertical axis at
5:71, a point corresponding to a median bet in the benchmark stock equal to
expð5:71Þ V or approximately 0:33% V. The six diagonal (green) parallel lines above
the median line mark orders whose log-sizes are one–six standard deviations above the log-
median sizes, respectively. Each log standard deviation represents an increase in bet size by
a factor of expð2:531=2 Þ 4:90.
For each of the 60 months from January 2001 to December 2005, the 400,000þ port-
folio transition orders are sorted into volume bins based on thresholds corresponding to the
30th, 50th, 60th, 70th, 75th, 80th, 85th, 90th, and 95th percentiles of the dollar volume
for NYSE-listed common stocks. The 600 blue diamonds in Figure 1 represent the largest
orders in each of 10 volume bins for each of 60 months. The diamond points form a cloud
tilted along invariance-implied iso-lines with slope of 2=3. These dots are certainly not on
a horizontal line, as would be predicted by the conventional wisdom. Since each bin con-
tains on average about 650 points, invariance and log-normality of order size suggest that
these largest portfolio transition orders should lie slightly below the 3-standard-deviation
diagonal with predicted slope of 2=3. As can be seen visually from the figure, this is

Figure 1. Largest portfolio transition orders and market crashes. This figure shows the largest port-
folio transition orders for each month from January 2001 to December 2005 and for each of ten vol-
ume groups (blue points) as well as the bets during five market crashes (red points). Volume groups
are based on thresholds corresponding to 30th, 50th, 60th, 70th, 75th, 80th, 85th, 90th, and 95th per-
centiles of dollar volume for common NYSE-listed stocks. The vertical axis is jlnðQ=V Þj. The horizontal
axis is lnðW =W Þ, where W ¼ 40 106 0:02 and W ¼ P V r. The median order is 5:71 ð2=3Þ
pffiffiffiffiffiffiffiffiffiffi
lnðW =W Þ (red line). The x-standard deviation events are 5:71 ð2=3Þ lnðW =W Þ þ x 2:53 (green
lines). According to conventional wisdom, which extrapolates along horizontal lines, imbalances dur-
ing the five crash events (red dots) are similar in size to typical large orders (blue diamonds).
According to the invariance hypothesis, which extrapolates along lines with slope 2=3, imbalances
during five crash events are rare outlier events.
approximately the case. The scatter plot of largest portfolio transition orders thus confirms
the predictions of the invariance hypothesis.
Figure 1 also depicts the five crash events by big round red dots. Extrapolating along
horizontal lines, the conventional wisdom would say that these five events are not unusual
compared with the largest portfolio transition orders. These percentages of daily volume
are not very different from what is seen in the largest portfolio transition orders, which are
often about 25% of daily volume in liquid stocks and an even larger percentage of daily
volume in illiquid stocks (up to 700%). The two flash crashes are only 2.29% and 1.49%
of daily volume. Even the largest crashes—the 1929 crash, the 1987 crash, and liquidation
of Kerviel’s position—represent “only” 265%, 67%, and 28% of daily volume and only
1.36%, 0.63%, and 0.43% of market capitalization, respectively. Therefore, based on hori-
zontal extrapolation, nothing unusual would be expected to happen during crash episodes.
In contrast, in the context of invariance with its extrapolation along diagonal lines with
slope 2=3, the crash events are extremely large, even when compared with the largest in-
stitutional orders. The two flash crashes correspond to about 4.5-standard-deviation
events. The 1929 crash, the 1987 crash, and the liquidation of Jérôme Kerviel’s positions
correspond to about 6-standard-deviation events. This suggests it would not be surprising

that they caused significant price dislocations.
2.6 Invariance-Implied Market Impact Formulas

We use a log-linear version of the linear impact model [Equation (13)]. The expected per-

centage price impact from buying quantity Q of a security with share price P, expected
daily volume V, and daily expected volatility r is given by
1=3 4=3
k PV r Q
DlnP ¼ 4
V

: (15)
10 P r H V
This formula assumes benchmark stock values P ¼ $40 per share, V ¼ 106 shares per
day, r ¼ 0:02 per day1=2 , with H ¼ 0:01 days 4 min. Invariance says that this factor k
is the same for all markets and time periods.
Kyle and Obizhaeva (2016) estimate k using data on implementation shortfall of port-
folio transition orders.4 Introduced by Perold (1988), this metric is the difference between
the execution price and a “paper trading” benchmark price recorded before the order was
placed. The calibrated value of k is equal to 5.00, with standard error 0.38 (about 7% of
the price impact). The value k ¼ 5:00 is scaled so that the price impact of an order for 1%
of daily volume in the benchmark stock is 5 basis points. Linear price impact therefore
implies that an order for 10% of daily volume has a price impact of 50 basis points. A price
impact of 50 basis points is similar to estimates in the literature on institutional block trades
of similar size.
We use a log-linear version of the market impact model rather than a simple linear
model because our analysis deals with very large orders, sometimes equal in magnitude to
trading volume of several trading days. In contrast, Kyle and Obizhaeva (2016) consider
relatively smaller portfolio transition orders with an average size of 4.20% and a median
size of 0.57% of daily volume; for these smaller orders, the distinction between e6x and
16x is immaterial.
Equation (15) is a universal formula for market impact, which may be applied to differ-
ent markets and time periods. Having calibrated it on the sample of portfolio transition
orders in the individual US stocks for 2001–05, we extrapolate the same formula to large
market bets and to different time periods.
In Appendix A, we discuss several implementation issues for applying invariance to the
five crash events. These include defining boundaries of the market, choosing proxies for
expected volume and volatility, and understanding functioning of market institutions.
Appendix B reports estimates implied by a number of alternative models of market impact
based on conventional wisdom and the literature on institutional trades.
4 Equation (37) of Kyle and Obizhaeva (2016, p. 1400) uses a slightly different specification, which
estimates an average impact cost parameter of j I ¼ 2:50 basis points (standard error 0.19 basis
points) for transition orders, not a price impact coefficient k itself. Of course, there is a tight con-
nection between the two concepts. Assuming that orders are broken into pieces and executed at
prices which tend to increase along an upward sloping linear supply schedule, total price impact
k ¼ 2 2:50 must be about twice the average impact cost j I . Although invariance also has impli-
cations for bid–ask spread costs, these costs are negligible for large bets, and hence we ignore
them. The implied standard error of k is 2 0:19 basis points, about 7% of the estimate 2 2:50.
2.7 A Market Crash Scenario

Suppose both illiquid and liquid assets have annual turnover of 100% over 250 trading
days. The illiquid asset is a benchmark stock with P ¼ $40 per share, V ¼ 106 shares per
day, and r ¼ 0:02 per day1=2 . The liquid asset is the entire US stock market, which consists
of both the stock index futures market and the cash stock market.

The market has daily dollar volume of about P V ¼ $270 billion per day, about 6,750
(¼ 153 2) times the dollar volume of the benchmark stock, and daily returns volatility r ¼
0:01 per day1=2 , one-half of stock volatility. Since business time passes at a rate proportion-
al to ðP V r=CÞ2=3 , the stock market operates about 225 times faster (¼ ð6; 750 1=2Þ2=3 )
than the market for the benchmark stock, implying H ¼ H =225. If bets in the benchmark
stock arrive about once every 4 min, bets in the entire market arrive about once per second.
Now let us compare a bet of 25% of daily volume in the benchmark stock with a bet of
25% of daily volume in the market as a whole. For the benchmark stock, the sale would be
250,000 shares worth $10 million. For the market as a whole, the sale would be slightly
less than $70 billion. Such sales might represent the liquidation of a large institutional pos-
ition, similar in magnitude to the liquidation of Jérôme Kerviel’s rogue trades in 2008, or it
might represent many small investors withdrawing equity exposure from index mutual
funds or exchange-traded funds (ETFs) over a short period of a few days.
The conventional wisdom and invariance make sharply different predictions about the
magnitude of price impacts of these bets.
The conventional wisdom predicts that the price impact of both bets would be minis-
cule. Since 100% turnover per year implies daily turnover of 0.40% of market capitaliza-
tion, a bet of 25% of 1 day’s volume represents 0.10% of market capitalization. Unit
demand elasticity therefore implies a price decline of 10 basis points for both assets, which
the market would barely notice.
In contrast, the invariance-implied extrapolation [Equation (13)] leads to very different
predictions. For the individual stock, since 1% of daily volume implies a price impact of 5
basis points, linear impact implies that a sale of 25% of daily volume in the benchmark
stock has a price impact of 125 basis points. The implied demand elasticity of 0.08 is far
smaller than the elasticity of one which represents conventional wisdom and is consistent
with the academic literature on the price impact of institutional bets.
For a bet on the entire market, the invariance-implied elasticity is much lower and the
price impact is correspondingly greater. Equation (13) implies that since trading activity is
higher by a factor of 153, price impact would be fifteen times higher if volatility was the
same. Since market volatility of 1% per day is half of the daily volatility of 2% for the
benchmark stock, price impact is reduced by a factor of 2 from 15 to 7.5 times the price im-
pact of 125 basis points for the individual stock. The price impact of a bet about $70 billion
in the market as a whole is therefore about 937 basis points, similar to the price declines
observed when Kerviel’s trades were liquidated. The implied demand elasticity for the mar-
ket as a whole is only 0.01, about 7.5 times smaller than for an individual stock and 100
times smaller than conventional wisdom.
While we disagree with the conventional wisdom that the demand elasticity for financial
assets is greater than one, we do not disagree with the greater overall level of price impact
for liquid stocks documented by the literature on institutional trades. Instead, by using the
invariance hypotheses, we suggest an alternative way of extrapolating price impact for indi-
vidual stocks to the overall stock market.
Our calculations suggest that the overall stock market is much more fragile than most
economists believe. Sudden equity index ETF or mutual fund liquidations of $200 billion
over a few days would potentially result in a 30% crash in stock prices, matching the crash
of 1987.

3. Examples of Five Market Crashes
We next discuss five crash episodes. The actual price changes during crash events reflect
not only sales by particular groups of traders placing large bets but also many other events
occurring at the same time, including arrival of news, trading by other traders, and func-
tioning of the trading infrastructure.5
3.1 The Stock Market Crash of October 1929

The stock market crash of October 1929 is the most infamous crash in the history of the
USA. It became associated with even larger stock price declines from 1930 to 1932, bank
runs, and the Great Depression.6
The Dow Jones average declined by about 25% during the last week of October 1929
(from 305.85 on October 23 to 230.07 on October 29) and 34% during the last 3 months
of 1929 (from 352.57 on September 25 to 234.07 on December 25). These price changes
included a 11% drop in the morning on Black Thursday, October 24; a 13% drop on Black
Monday, October 28; and another 12% drop on Black Tuesday, October 29.
In the late 1920s, many Americans became heavily invested in a stock market boom. A
significant portion of stock investments was made in leveraged margin accounts. Between
1926 and 1929, both the level of margin debt and the level of the Dow Jones average
doubled in value. Both the stock market boom and the boom in margin lending came to an
abrupt end during the last week of October 1929.
During the week before Black Thursday, October 24, the Dow Jones average fell 9%,
including a drop of 6% on Wednesday, October 23, and this led to a self-reinforcing cycle
of liquidations of stocks in margin accounts.
To quantify the margin selling which occurred during the last week of October 1929,
we follow the previous literature and contemporary market participants by estimating mar-
gin selling indirectly from data on broker loans and bank loans collateralized by securities.
For the last week of October 1929, we estimate margin selling as $1.181 billion. For the 3
months from September 30, 1929, to December 31, 1929, we estimate total margin selling
as $4.348 billion. Details of the estimations are presented in Appendix C.
These liquidations exerted downward price pressure on the stock market. To estimate
its magnitude, we treat the 1929 stock market as one market, rather than numerous mar-
kets for different stocks, and plug estimates of expected dollar volume and volatility for the
entire stock market into Equation (15).
Historical volatility during the month prior to October 1929 was about 2.00% per day.
Historical volume was $342.29 million per day in 1929 dollars. Prior to 1935, the volume
reported on the ticker did not include “odd-lot” transactions and “stopped-stock”
5 The data for the stock market are from CRSP and the data for index futures are from Bloomberg.
6 Our analysis is based on several documents: Board of Governors of the Federal Reserve System
(1927–1931, 1929); Galbraith (1954); Senate Committee on Banking and Currency (1934); Friedman
and Schwartz (1963); Smiley and Keehn (1988); and Haney (1932).
transactions, which have been estimated to be equal about 30% of “reported” volume
(Board of Governors of the Federal Reserve System, 1943, p. 431). We thus multiply
reported volume by 13/10, obtaining an estimate of $444.97 million per day. The margin
sales of $1.181 billion during the last week of October were approximately 265% of aver-
age daily volume.

Equation (15) implies that margin-related sales of $1.181 billion were expected to trig-
ger a price decline of 46.43%7:
!1=3 4=3
5:00 444:97 106 9:42 0:0200
46:43% ¼ 1 exp
104 40 106 0:02
!
1:181 109
:
ð0:01Þð444:97 106 Þ
As a robustness check, Table II reports other estimates using volume and volatility calcu-
lated over the preceding m months, with m ¼ 1, 2, 3, 4, 6, 12. Invariance predicts price
declines ranging from 26.79% to 46.43%, somewhat larger than the actual price change of
25%.8
In contrast, since the reduction of broker loans of $1.181 billion was only a very small
fraction of the $87.1 billion market capitalization of NYSE issues at the end of September
1929 (Brady Report, p. VIII-13), conventional intuition [Equation (1)] predicts a price
change of only 1.36%, much smaller than actual price decline of 25% and about forty
times smaller than the magnitude predicted by invariance.
We also make price impact calculations for margin sales of $4.348 billion during the
last 3 months of 1929. Conventional wisdom implies a price drop of 4.99%. Invariance
implies a much larger price decline ranging from 68.28% to 89.95%, more than the actual
price decline of 34% during the last 3 months of 1929 and the price decline of 44% from
high point in late September 1929 to low point in mid-November 1929.
3.2 The Market Crash in October 1987

From Wednesday, October 14, 1987, to Tuesday, October 20, 1987, the US equity market
suffered the most severe 1-week decline in its history. The Dow Jones index dropped 32%
from 2,500 to 1,700; as of noon Tuesday, October 20, the S&P 500 futures prices had
dropped about 40% from 312 to 185. On Black Monday alone, October 19, 1987, the
Dow Jones index fell 23% and the S&P 500 futures market dropped 29%.
It has long been debated whether the market crash resulted from the sales by institutions
implementing portfolio insurance. Portfolio insurance is a trading strategy that replicates
put option protection for portfolios by dynamically adjusting stock market exposure in re-
sponse to market fluctuations. Since portfolio insurers sell stocks when prices fall, the strat-
egy amplifies downward pressure on prices in falling markets. We calculate the price
impact of portfolio insurance sales implied by invariance.
7 To convert 1929 dollars to 2005 dollars, we use the GDP deflator, which equates $1 in 1929 to $9.42
in 2005. We use the year 2005 as a benchmark because the estimates of Kyle and Obizhaeva (2016)
are based on the sample period 2001–05, with more observations occurring in the latter part of the
sample.
8 The 7% standard error for price impact implies a 2-standard-deviation interval to be
46:43%ð162 %7Þ, which is larger than the actual price decline of 25%.
Table II. 1929 stock market crash: implied price impact of margin sales
The table shows the invariance-implied impact of $1.181 billion of margin sales during the
week October 24–30, 1929, and $4.343 billion of margin sales from September 25 to December
25, along with average daily 1929 dollar volume and daily volatility for m ¼ 1, 2, 3, 4, 6, and 12
months preceding October 24, 1929. The conventional wisdom predicts a price decline of 1.36%

from October 24 to 29 and 4.99% from September 25 to December 25. The actual price decline
was 25% from October 24 to 29 and 34% from September 25 to December 25.
Months preceding October 24, 1929
m: 1 2 3 4 6 12
ADV (1929 $M) 444.97 461.45 436.49 427.20 387.18 390.45

Daily volatility 0.0200 0.0159 0.0145 0.0128 0.0119 0.0111
10/24–10/30 sales (% ADV) 265% 256% 271% 276% 305% 302%
Price impact 46.43% 36.26% 33.75% 29.93% 29.00% 26.79%
9/25–12/25 sales (% ADV) 977% 942% 996% 1,018% 1,123% 1,114%
Price impact 89.95% 80.95% 78.04% 73.01% 71.66% 68.28%
We consider the entire stock market to be one market; this is consistent with the Brady
Report. Accordingly, we estimate daily volume as the sum of average daily volume in the
futures market and the NYSE for the previous month. Some portfolio insurers abandoned their
reliance on the futures markets and switched to selling stocks directly because futures contracts
became unusually cheap relative to the cash market. We construct a proxy for sales as the sum
of portfolio insurance sales in the futures market and the NYSE from tables in the Brady
Report, Figures 13–16, pp. 197–198, obtaining results similar to Gammill and Marsh (1988).
Over the 4 days, October 15, 16, 19, and 20, 1987, portfolio insurers sold S&P 500 futures
contracts representing $10.48 billion in index futures and $3.27 billion in NYSE stocks. The
gross sales amount of $13.75 billion in futures and stocks is combined for the purpose of ana-
lyzing price impact of portfolio insurance sales. Reported values are all 1987 dollars.
In the month prior to the crash, the historical volatility of S&P 500 futures returns was
about 1.35% per day, similar to estimates in the Brady Report. The average daily volume
in the S&P 500 futures market was equal to $10.37 billion. The NYSE average daily vol-
ume was $10.20 billion. Portfolio insurance gross sales were equal to about 67% of 1 day’s
combined volume.
Plugging portfolio insurance gross sales and market parameters into Equation (15)
yields a price decline of 16.77%9:
!1=3
5:78 ð10:37 þ 10:20Þ 109 1:54
16:77% ¼ 1 exp
104 40 106
!
4=3 ð10:48 þ 3:27Þ
0:0135
0:02 :
ð0:01Þð10:37 þ 10:20Þ
Table III reports other estimates based on historical trading volume and volatility calcu-
lated over the preceding m months, with m ¼ 1, 2, 3, 4, 6, 12. These estimates range from
11.87% to 16.77%. For robustness, estimates under several alternative assumptions are
9 The GDP deflator of 1.54 converts 1987 dollars to 2005 dollars.

Table III. 1987 stock market crash: effect of portfolio insurance sales
The table shows the invariance-implied impact triggered by portfolio insurers’ net sales of S&P
500 futures contracts ($9.51 billion) and NYSE stocks ($1.60 billion), portfolio insurers’ gross
sales of S&P 500 futures contracts ($10.48 billion) and NYSE stocks ($3.27 billion), portfolio
insurers’ sales of S&P 500 futures adjusted for purchases of index arbitrageurs ($10.48 billion

minus $3.27 billion), and portfolio insurers’ sales of NYSE stocks adjusted for sales of index
arbitrageurs ($3.27 billion plus $3.27 billion) in 1987 dollars. Average daily dollar volume and
daily volatility are based on m months preceding October 14, 1987, with m ¼ 1, 2, 3, 4, 6, and
12, both for the S&P 500 futures and CRSP stocks. Conventional wisdom predicts price declines
of 0.51% for portfolio insurers’ net sells and 0.63% for their gross sells. The actual price decline
was 32% for the Dow Jones average and 40% for S&P 500 futures.
m: 1 2 3 4 6 12
S&P 500 Fut ADV (1987 $B) 10.37 11.29 11.13 10.12 10.62 9.85
NYSE ADV (1987 $B) 10.20 10.44 10.48 10.16 10.04 9.70
Gross sells (% ADV) 66.84% 63.28% 63.65% 67.82% 66.53% 70.33%
Price impact 16.77% 14.18% 12.23% 11.87% 13.20% 13.64%
Price impact of net sales combined 13.78% 11.62% 10.00% 9.71% 10.81% 11.17%
Price impact of S&P 500 sales 14.11% 11.67% 10.10% 10.00% 10.93% 11.45%
Price impact of NYSE sales 13.00% 11.18% 9.56% 9.09% 10.32% 10.53%
presented in Table III. Details are presented in Appendix C. The similarity between pre-
dicted and observed price declines is consistent with our hypothesis that heavy selling by
portfolio insurers played a dominant role in the crash of October 1987.
The estimates based on conventional wisdom are much smaller. According to the Brady
Report, there were 2,257 issues of stocks listed on the NYSE, with a value of $2.2 trillion
on December 31, 1986. Conventional wisdom implies that gross sales of $10.48 billion in
futures and $3.27 billion in individual stocks, representing 0.63% of shares outstanding in
total, would have an impact of only 0.63%. Other alternative models yield estimates not
higher than 2%. Citing similar arguments, many experts have rejected the idea that sales of
portfolio insurers caused the 1987 market crash.
The invariance-implied estimate of 16.77% is smaller than the price drops of 32% in
the cash equity market and 40% in the S&P 500 futures market. The 1987 crash may have
been triggered by negative news about anti-takeover legislation as well as trade deficit stat-
istic on October 14. It may have been further aggravated by break-downs in the market
mechanism which disrupted index arbitrage relationships, as documented in the Brady
Report.
3.3 Trades of George Soros on October 22, 1987

On Thursday, October 22, 1987, just 3 days after the 1987 market crash, George Soros lost
$60 million in minutes by selling a large number of S&P 500 futures contracts as prices
spiked down 22% at the opening of trading. These sales have been attributed to pessimistic
predictions that Robert Prechter made based on the Elliott Wave Theory—a form of
technical analysis that looks for recurrent long-term price patterns related to persistent
changes in investor sentiment. He pointed out similarities between the 1929 crash and the
1987 crash.
The Commodity Futures Trading Commission, Division of Economic Analysis and
Division of Trading and Markets (1988) issued a report describing the events of October

22, 1987, without mentioning Soros by name. At 8:28 a.m. Central Time (CT), approxi-
mately 2 min before the opening bell at the NYSE, a customer of a clearing member submit-
ted a 1,200-contract sell order at a limit price of 200, more than 20% below the previous
day’s close of 258. Over the first minutes of trading, the price dropped to 200, at which
point the sell order was executed. At 8:34 a.m., a second identical limit order for 1,200 con-
tracts from the same customer was executed by the same floor broker. These transactions
liquidated a long position acquired on the previous day at a loss of about 22% or about
$60 million in 1987 dollars. Within minutes, S&P 500 futures prices rebounded and, over
the next 2 h, the market recovered to the levels of the previous day’s close. Within days,
Soros’s Quantum Fund sued the brokerage firm which handled the order, alleging a con-
spiracy among traders to keep prices artificially low while his sell orders were executed.
Two other events may have exacerbated the decline in prices in the morning of October
22 by increasing the selling pressure. First, when the broker executed the second order, he
mistakenly sold 651 more contracts than the order called for. The oversold contracts were
taken into the clearing firm’s error account and liquidated at a significant loss to the broker.
Second, the Commodity Futures Trading Commission, Division of Economic Analysis and
Division of Trading and Markets (1988) reports that between 9:34 a.m. and 10:45 a.m. the
same clearing firm also entered and filled four large sell orders for another customer—a
pension fund—with a total of 2,478 contracts sold at prices ranging from 230 to 241.
These additional orders are for almost exactly the same size as Soros’s orders. This fact sug-
gests information leakage or coordination regarding the size of these unusually large
orders.
We compare the actual price decline of 22% with predictions based on invariance.
During the prior month, average daily volatility was 8.63% and average daily volume in
the S&P 500 futures market was $13.52 billion in 1987 dollars. The very high volatility es-
timate based on crash data is reasonable because market participants expected high volatil-
ity to persist. Since Soros’s sales started just before the opening of NYSE trading, the
arbitrage mechanism which connects stock and futures markets did not have time to work;
indeed, futures contracts traded at levels about 20% cheaper than stocks. We thus consider
only S&P 500 futures market in this example, not combining it with the market for NYSE
stocks.
Each S&P 500 contract had a notional value of 500 times the S&P 500 index. With an
S&P 500 level of 258, one contract represented ownership of about $129,000. Soros’ sale
of 2,400 contracts, or about $309.60 million, was equal to 2.29% of average daily volume.
Given the prior month estimates, Equation (15) predicts a price decline of 6.27%10:
!1=3 4=3 !
5:00 13:52 109 1:54 0:0863 309:60 106
6:27% ¼ 1 exp :
104 40 106 0:02 ð0:01Þð13:52 109 Þ
10 The GDP deflator of 1.54 converts 1987 dollars to 2005 dollars.

Table IV. October 22, 1987: effect of Soros’s trades
The table shows the invariance-implied price impact of (A) Soros’s sell order of 2,400 contracts;
(B) Soros’s sell order of 2,400 contracts plus 651 contracts of error trades (3,051 contracts in
total); and (C) Soros’s sell order of 2,400 contracts, plus 651 contracts of error trades, plus the
sell order of 2,478 contracts by the pension fund (5,529 contracts in total). Calculations use aver-

age daily 1987 dollar volume and daily volatility for m ¼ 1, 2, 3, 4, 6, and 12 months preceding
October 22, 1987, for the S&P 500 futures contracts. Conventional wisdom predicts price
declines of 0.01%, 0.02%, and 0.03%, respectively. The actual price decline in the S&P 500
futures market was 22%.
m: 1 2 3 4 6 12
S&P 500 Fut ADV (1987 $B) 13.52 11.72 11.70 10.99 10.75 10.04
2,400 contracts as % ADV 2.29% 2.64% 2.65% 2.82% 2.88% 3.08%
Price impact A 6.27% 4.50% 3.40% 2.96% 2.36% 1.67%
Price impact B 7.90% 5.68% 4.30% 3.75% 2.99% 2.12%
Price impact C 13.85% 10.06% 7.66% 6.69% 5.36% 3.81%
Table IV presents three sets of estimates based on the historical volume and volatility of
S&P 500 futures contracts calculated over the preceding m months, with m ¼ 1, 2, 3, 4, 6,
12. Invariance implies (i) price impact of 1.67–6.27% based on 2,400 contracts alone; (ii)
price impact of 2.12–7.90% adding 651 error contracts (3,051 contracts in total); and (iii)
price impact of 3.81–13.85% adding 2,478 contracts sold by the pension fund (5,529 con-
tracts in total).
The actual price decline of 22% is significantly larger than our estimate. Factors which
could have led to large impact include potentially underestimated expected volatility, front-
running based on leakage of information about the size of the order, and the peculiarly ag-
gressive execution strategy of placing two limit orders with a limit price of 200, more than
20% below the previous day’s close.
Conventional wisdom would imply minuscule price changes. Given the total value of
$2.2 trillion of issues listed on the NYSE at the end of 1986, the Soros’s order, the errone-
ous sales, and the sales by the pension fund would be expected to have a combined impact
of only 0.03%.
3.4 Liquidation of Kerviel’s Rogue Trades in January 2008

On January 24, 2008, Société Générale issued a press release stating that the bank had
“uncovered an exceptional fraud.” Subsequent reports by Société Générale (2008a, 2008b,
2008c) revealed that rogue trader Jérôme Kerviel had used “unauthorized” trading to place
large bets on European stock indices.
Kerviel had established long positions in equity index futures contracts with underlying
values of e50 billion: e30 billion on the Euro STOXX 50, e18 billion on DAX, and e2 bil-
lion on the FTSE 100. He acquired these naked long positions mostly between January 2
and January 18, then concealed them using fictitious short positions, forged documents,
and emails suggesting his positions were hedged. The fall in index values in the first half of
January led to losses on these secret directional bets. Internal investigators became strongly
suspicious about the nature of the positions on Friday, January 18.
Société Générale informed the heads of the central bank and the Financial Markets
Authority (AMF), the French stock market regulator. The AMF allowed the bank to delay
public announcement of the fraud for 3 days so that Kerviel’s positions could be liquidated

quietly. The head of the central bank also delayed informing the government. After liqui-
dating the positions between Monday, January 21, and Wednesday, January 23, the bank
had sustained losses of e6.4 billion which—after subtracting out e1.5 billion profit as of
December 31, 2007—were reported as a net loss of e4.9 billion.
As Société Générale liquidated the positions, prices fell all across Europe. The Stoxx
Europe Total Market Index (TMI)—which represents all of Western Europe—fell by
9.44% from the close on January 18 to its lowest level on January 21. On Monday,
January 21—a bank holiday with muted US financial markets activity—the Fed held an un-
scheduled meeting of the Federal Open Market Committee (FOMC) via conference call at
6:00 p.m. New York time, several days before its scheduled meeting. At 8:30 a.m. the next
day, the Fed announced an unprecedented 75-basis point cut in interest rates, which pushed
all prices up and helped Société Générale to liquidate the rest of the position on better
terms. We do not know whether Fed officials were aware of Société Générale’s situation
when this decision was made. According to the Fed’s Minutes (Board of Governors of the
Federal Reserve System, 2008), published five years later, the purpose of the meeting was
to “to update the Committee on financial developments over the weekend and to consider
whether we want to take a policy action,” but there is no mention of Société Générale. In
his memoir, Bernanke (2015, pp. 195–196) said the Fed “had no idea the rogue-trading
bombshell was coming,” but he mentions (p. 195) “a conference call the morning of
January 19 Paris time,” during which “senior SocGen managers in Paris and New York
had told New York Fed supervisors that the bank would report positive earnings for the
fourth quarter, even after taking write-downs on its subprime mortgage exposure.”
On the one hand, the surprise early announcement of an interest rate cut could have
helped the bank obtain more favorable execution prices on some portion of its trades. On
the other hand, January 21 was a bank holiday in the USA; in the previous year, the futures
markets had only one-third of the typical volume on days when US markets were closed.
Low volume on the bank holiday could have reduced liquidity, making the unwinding of
positions more expensive.
Due to significant correlations among European markets, we perform our analysis under
the assumption that all European stock and futures markets are one market. Based on data
from the World Federation of Exchanges, the seven largest European exchanges by market
capitalization in 2008 (NYSE Euronext, London Stock Exchange, Deutsche Börse, BME
Spanish Exchanges, SIX Swiss Exchange, NASDAQ OMX Nordic Exchange, and Borsa
Italiana) had average daily volume for the month ending January 18, 2008, equal to e69.51
billion.
We also sum average daily volume across the ten most actively traded European equity
index futures markets (Euro Stoxx 50, DAX, CAC, IBEX, AEX, Swiss Market Index SMI,
FTSE MIB, OMX Stockholm 30, and Stoxx 50 Euro) and find average daily futures volume
of e110.98 billion. The total daily volume in both European stock and equity futures mar-
kets was equal to e180.49.
Our estimate of expected volatility is 1.10%, the previous month’s daily standard devi-
ation of returns for the Stoxx Europe TMI.
Table V. January 2008: effect of liquidating Kerviel’s positions
The table shows the invariance-predicted losses of liquidating Kerviel’s positions of e50 billion
under the assumption that the major European cash and futures markets are integrated, and
one Euro is worth $1.4690. Results are provided based on average daily volume of the major
European stock exchanges and index futures as well as daily volatilities of Stoxx Europe TMI,

based on m months preceding January 18, 2008, with m ¼ 1, 2, 3, 4, 6, and 12. Conventional
wisdom predicts price decline of 0.43%. The actual price decline in the Stoxx Europe TMI was
9.44%.
Months preceding January 18, 2008
m: 1 2 3 4 6 12
Stk Mkt ADV (2008 eB) 69.51 66.51 67.37 67.01 66.73 66.32
Fut Mkt ADV (2008 eB) 110.98 114.39 118.05 117.46 127.17 121.26
Order as % ADV 27.70% 27.64% 26.97% 27.11% 25.79% 26.66%
Price impact 10.79% 12.66% 11.94% 11.53% 12.93% 10.59%
Total losses (2008 eB) 2.77 3.27 3.08 2.97 3.34 2.72
Losses: Adj. A (2008 eB) 5.08 5.58 5.39 5.28 5.65 5.03
Losses: Adj. B (2008 eB) 7.39 7.89 7.70 7.59 7.96 7.34
According to Equation (15), the liquidation of Kerviel’s e50 billion position—equal to

about 27.70% of the average daily volume in aggregated stock and futures markets—is
expected to trigger a price decline of 10.79% across European markets:
5:00
10:79% ¼ 1 exp
104
!1=3 4=3 !
180:49 1:4690 0:92 109 0:0011 50
:
40 106 0:02 ð0:01Þ 180:49
In this equation, we use an exchange rate of $1.4690 per Euro to convert Euro volume into
US dollar volume and convert 2008 dollars into 2005 dollars to be able to use them in our
calibrated formulas.11
Table V shows the estimates of price impact based on historical trading volume and
volatility calculated over the preceding m months, with m ¼ 1, 2, 3, 4, 6, 12. Invariance pre-
dicts price changes ranging from 10.59% to 12.93%. The Stoxx TMI index actually fell by
9.44% from the market close of 316.73 on January 18 to its lowest level of 286.82 on
January 21.
In contrast, conventional wisdom predicts that sales of e50 billion would have a much
smaller price impact of 0.43%, given that it represents less than 1% of the total capitaliza-
tion of European markets, which was about e11.752 trillion in December 2007, as reported
by Federation of European Securities Exchanges.
We also estimate the dollar costs of liquidating the rogue position to range from e2.72
billion to e3.34 billion under different assumptions about expected volume and volatility.
Adding mark-to-market losses sustained prior to liquidation leads to estimated losses
11 The GDP deflator of 0.92 converts 2008 dollars into 2005 dollars.
ranging from e5.03 billion to e7.96 billion. These estimates are consistent with officially
reported losses of e6.30 billion. Appendix C presents the details.
In explaining the costs of liquidating the positions to shareholders already concerned
about the bank’s losses on subprime mortgages, bank officials blamed “the very un-
favorable market conditions” (see the explanatory note about the exceptional fraud

released by Société Générale on January 27). Expressing conventional wisdom, the bank
announced that its trades accounted for not more than 8% of turnover on any one of
the futures exchanges on which they were conducted and thus did not have a serious
market impact. In contrast, when examined through the lens of invariance, the reported
losses are almost exactly of the magnitude expected from the price impact of the trades
on European stock markets.
3.5 The Flash Crash of May 6, 2010

During the morning of May 6, the S&P 500 declined by 3%. Rumors of a default by
Greece had made markets nervous in a context where there was already uncertainty about
elections in the UK and an upcoming jobs report in the USA.
During the 5-min interval from 2:40 p.m. to 2:45 p.m. ET, the E-mini S&P 500 futures
contract suddenly dropped 5.12% from 1,113 to 1,056. After a pre-programmed circuit
breaker built into the CME’s Globex electronic trading platform halted trading for 5 s, pri-
ces went up 5% over the next 10 min, recovering losses.
After the flash crash, the Staffs of the CFTC and SEC (2010a, 2010b) issued a joint re-
port. They said that an automated execution algorithm sold 75,000 S&P 500 E-mini
futures contracts between 2:32 p.m. and 2:51 p.m. on the CME’s Globex platform, exactly
during the V-shaped flash crash. The E-mini contract represents exposure of fifty times the
S&P 500 index, one-tenth the multiple of 500 for the older but otherwise similar contract
sold by portfolio insurers in 1987. The program sold S&P 500 exposure of $4.37 billion.
The joint report did not mention the name of the seller, but journalists identified the seller
as Waddell & Reed.
Many people did not believe that selling 75,000 contracts could have triggered a
price decline of 5%, because they implicitly relied on conventional intuition: The $4.37-
billion sale represented only 3.75% of the daily volume of about 2,000,000 contracts
per day in the S&P 500 E-mini futures market, and its impact was expected to be tiny.
Thus, many accused high-frequency traders of failing to provide liquidity as prices
collapsed.
We examine whether such an order of 75,000 contracts could have resulted in a
flash crash. During the preceding month, the volume in E-mini contracts was about
$132 billion per day, and the volume in the stock market was about $161 billion
per day; the combined daily volume was $292 billion. The historical daily volatility
was 1.07%.
Equation (15) implies that the sales of $4.37 billion—equal to about 3.31% of daily
volume in S&P 500 E-mini futures market in the previous month or 1.49% for futures
and stock market combined—are expected to trigger a price decline of 0.61%12:
12 The GDP deflator of 0.90 converts 2010 dollars into 2005 dollars.
Table VI. Flash crash of May 6, 2010: effect of 75,000 contract futures sale
The table shows the invariance-implied price impact of sales of 75,000 S&P 500 E-mini futures
contracts. Calculations are based on average daily volume and volatility of the S&P 500 E-mini
futures for the m months preceding May 6, 2010, with m ¼ 1, 2, 3, 4, 6, and 12. Conventional
wisdom predicts a price decline of 0.03%. The actual price decline in the S&P 500 E-mini futures

market was 5.12%.
Months preceding May 6, 2010
m: 1 2 3 4 6 12
S&P 500 Fut ADV (2010 $B) 132.00 107.49 109.54 112.67 100.65 95.49
Stk Mkt ADV (2010 $B) 161.41 146.50 142.09 143.03 132.58 129.30
Order as % ADV 1.49% 1.72% 1.73% 1.71% 1.87% 1.94%
Price impact (hist r) 0.61% 0.49% 0.44% 0.53% 0.55% 0.73%
Price impact (r ¼ 2%) 1.39% 1.52% 1.53% 1.52% 1.61% 1.65%
!1=3
5:00 ð132 þ 161Þ 0:90 109
0:61% ¼ 1 exp
104 40 106
!
4=3 75; 000 50 1; 164
0:0107
0:02 :
0:01 ð132 þ 161Þ 109
Table VI shows estimates based on historical volume and volatility of S&P 500 E-mini
futures contracts calculated over the preceding m months, with m ¼ 1; 2; 3; 4; 6; 12. These
estimates range from 0.44% to 0.73%. Appendix C provides estimates under the alterna-
tive assumptions of 2% volatility and less integrated markets. These estimates tend to be
higher, ranging from 0.76% to 2.91%.
The predicted price impact of 0.61% is smaller than the actual decline of 5.12%. As dis-
cussed in the next section, unusually fast execution may have significantly increased the
temporary impact of these trades and led to rapid rebound in prices afterward. Given that
the capitalization of US market was about $15.077 trillion at the end of 2009, conventional
wisdom would predict a price decline of only 0.03%.
4. Policy Implications and Lessons Learned

Application of microstructure invariance to intrinsically infrequent historical episodes
requires an exercise in judgement to extract appropriate lessons learned. In some cases, our
theory predicts price declines that differ from actual declines. We discuss next factors that
can potentially explain these differences. While speculative in nature, our discussion sug-
gests important lessons for policymakers concerned with measuring and predicting crash
events of a systemic nature, for asset managers worried about managing market impact
costs associated with execution of large trades that might potentially disrupt markets, and
for researchers interested in understanding of how financial markets work.
4.1 Price Impact Is Large in Liquid Markets

Market participants often execute large orders by restricting quantities traded to be not
more than 5% or 10% of average daily volume over a period of several days. This heuristic
strategy is usually believed to be reasonable for individual stocks and thus certainly reason-
able for more liquid markets such as markets for stock index futures.

While this strategy is reasonable for trading individual stocks, our analysis shows that it
may incur much larger-than-expected costs when implemented in more liquid markets.
Invariance implies that the price impact (in volatility units) of trading a given fraction of
daily volume is proportional to the cube root of dollar volume and returns volatility. For
example, if dollar volume P V increases by a factor of 1,000—approximately consistent
with the difference between the benchmark stock and stock index futures—the market im-
pact of a given fraction of volume in Equation (15) increases by a factor of ð1000Þ1=3 ¼ 10.
The larger impact in more liquid markets follows from the invariance hypotheses com-
bined with linear price impact.
4.2 Rapid Execution Magnifies Transitory Price Impact

Our analysis sheds some light on how price dynamics in response to large bets depends on
the speed of execution. While the long-term impact of bets is likely to depend on their infor-
mation content, the short-term price dynamics is probably affected by the speed of trad-
ing.13 Speeding up execution exacerbates temporary impact.
If uninformative sell bets are executed faster, prices temporarily fall further and execu-
tion prices are worse than with slower execution of the same bet. After order execution is
finished, prices revert to the same levels which would have prevailed under slower execu-
tion. Extremely rapid bet execution therefore leads to V-shaped price paths, in which prices
first plunge sharply and then recover rapidly.
The model of smooth trading of Kyle, Obizhaeva, and Wang (2018) provides a theoret-
ical framework for modeling short-term price reactions to unusually rapid execution of
large bets.14 The model implies that markets interpret extremely rapid, heavy selling as an
indication that extremely negative information is about to flow into the market. Prices col-
lapse immediately when a heavy rate of selling is detected. When the expected negative in-
formation does not materialize, prices rebound, even though much of the heavy selling
continues.
The impact formula (15) contains parameters calibrated using the size of portfolio tran-
sition orders but not the speed of execution. Most transitions were executed over a period
of a few days, and only the most complex of them were carried out over a period of a few
weeks. Executions at a prudent pace were designed to keep impact costs low. Extrapolating
estimates from portfolio transitions to sales during crashes implicitly makes the identifying
13 Financial crises eventually followed the crash events of 1929 and liquidation of Kerviel’s rogue
trades in 2008. Whether margin sales in 1929 or Kerviel’s trades in 2008 had information content is
a difficult question to frame in a meaningful manner. For example, perhaps Kerviel traded against
informed traders who correctly foresaw the impending financial crisis, delaying the incorporation
of this information into prices until his own positions were liquidated.
14 The model of smooth trading gives rise in the equilibrium to both endogenous permanent and tem-
porary impacts, k and j. In most of the traditional models such as Kyle (1985), price impact is per-
manent and transaction costs of an informed trader do not depend on the speed of trading as
long as he trades continuously.
assumption that crash orders were also executed at the same “natural” or prudent pace as
portfolio transition orders.
During 1987 and 2010 flash crashes, larger-than-predicted price declines followed by
rapid price recoveries suggest that transitory price impact may have been exacerbated by
the extremely rapid rate at which selling took place. According to the Staffs of the CFTC

and SEC (2010b), for example, the May 2010 flash crash order was executed extremely
rapidly in just 20 min, while in previous months two orders of similar magnitudes had been
executed over periods of 5 and 6 h, which is fifteen times slower.15
4.3 Policies to Maintain Credit Availability Mitigate Crashes

Some policy responses may help to mitigate the negative effects of crashes. These policies
have to aim at easing flow of credit as well as providing funds that will make up the gap in
demand and supply. Our analysis suggests that the amount of funds necessary to counteract
the shock must be comparable to the size of the shock itself.
A good example is the 1929 market crash, during which price declines were somewhat
smaller than predicted by invariance and the crash was well contained until the end of
1929. First, immediately after the initial stock market break on Black Thursday, a group
of prominent New York bankers put together an informal fund of about $750 million to
buy securities in order to support prices. The fund was similar in size to the margin sales
shock of about $1.181 billion. When their decisions were publicized, the sense of panic
subsided. Similar actions were undertaken by J.P. Morgan and other bankers after a crash
in 1907.
Second, the New York Fed acted prudently in 1929. In the 1920s, bankers and their reg-
ulators were aware that if non-bank lenders suddenly withdrew funds from the broker loan
market, there would be pressure on the banking system to increase lending to make up the
difference. By discouraging banks from lending into the broker loan market prior to the
1929 crash, the New York Fed increased the ability of banks to support it after the stock
market crashed. During the last week of October 1929, the New York Fed wisely reversed
course and encouraged banks to provide clients with bank loans collateralized by securities.
As lenders withdrew funds from the broker loan market, the resulting unprecedented in-
crease in demand deposits at New York banks gave banks plenty of cash for financing
increased loans on securities. The New York Fed also encouraged easy credit by purchasing
government securities and cutting its discount rate twice. Some brokers cut margins from
40% to 20%.
All of these stabilizing policies smoothed the margin selling out and allowed brokers to
liquidate the large positions of under-margined stock investors gradually over 5 weeks, ra-
ther than selling collateral off at fire-sale prices over several days. This appears to have
helped the market to digest imbalances and reduce temporary price impact, thus avoiding a
sudden, brutal bursting of the stock market bubble.
15 The smooth trading model implies that the temporary price impact is linear in the speed of trading.
Since selling during the flash crash occurred about fifteen times faster than normal order execu-
tion, the model implies transitory price impact to be fifteen times greater than in the case when
selling occurs at a “normal” rate, followed by a reversal. Multiplied by 15, our estimates of 0.61%
price decline become even larger than the actual decline of 5.12%.
4.4 Effect of Large Bets May Propagate across Integrated Markets

During the 1987 crash, both US markets and many major world markets experienced severe
declines, despite the fact that the portfolio insurance selling was confined to the USA.
According to Roll (1988), this justifies concluding that portfolio insurance did not trigger
the 1987 crash.

Common patterns across markets were also documented during liquidation of Kerviel’s
positions in January 2008. Even markets where Société Générale did not liquidate positions
had very similar performance. The bank thus argued that its own impact on prices was lim-
ited and large price declines in multiple markets had to be attributed to other factors.
In contrast, we believe that financial markets are integrated and heavy selling in one
market is likely to affect correlated markets. Due to this connectedness, market impact esti-
mates should take into account how market liquidity is shared across markets for related
securities. For the 1987 crash, it supports treating the stock and futures markets as one big
market. For Société Générale’s trades, it supports aggregating across all European markets
rather than focusing only on countries where the bank liquidated positions. At the same
time, the question of how to define the boundaries of the market for the purpose of apply-
ing invariance hypotheses and predicting price impact in correlated markets remains to be
understood.
4.5 “Inefficient” Liquidity Silos May Enhance Stability

There may be a trade-off between efficiency and stability. In markets with less efficient
trading arrangements, more capital is required to sustain orderly trading, but this capital
also makes the systems more stable during volatile times. Invariance may help to assess the
effect of market integration on liquidity.
The inefficiency of financial markets in 1920s may explain their remarkable resilience.
During the 1929 crash, the disproportionably large amount of selling related to liquidation
of margin loans was more than fifteen times greater than selling during the 1987 crash, as a
percentage of GDP, but the price decline was only half as large.16
In the 1920s, speculative capital may have been compartmentalized into numerous sep-
arate silos. Speculative trading and intermediation associated with underwriting of new
stock issues often took place in “pools,” which played a role similar to today’s hedge funds.
The pools were typically dedicated to trading only one stock, and investors in the pools
often had close connections to the company whose stock the pool traded; there were no
prohibitions against insider trading and no SEC requiring firms to disclose material infor-
mation to the market. These pools traded actively, used leverage, took short positions, and
arbitraged stocks against options, particularly when facilitating distribution of newly issued
equity. There were no futures markets or ETFs allowing investors to trade large baskets of
stocks easily. When faced with massive liquidations of margin loans, the market may have
more speculative capital available to stabilize the situation than in a more “efficiently”
16 The 1987 portfolio insurance trades of $13 billion were equal to only about 0.28% of GDP in that
year (1987 GDP was $4.7 trillion); stock prices fell 32%. During the last week of October 1929, the
margin-related sales of $1.181 billion were equal to about 1% of GDP (1929 GDP was $104 billion),
approximately four times the levels of the 1987 crash; yet stock prices fell by only 25%. Inclusion
of additional sales equal to about 3% of GDP in subsequent weeks makes margin selling in 1929 to
be more than fifteen times greater than selling during the 1987 crash, as a percentage of GDP.
leveraged system in which institutional investors can spread their capital across markets by
trading hundreds of stocks simultaneously.
Invariance-implied estimates would change significantly if instead of being considered
as one large market, the stock market in 1929 was thought of as a set of many small, iso-
lated, and thus less liquid markets for individual stocks. One would expect market impact

to be much smaller in these less liquid markets for the same market bet. As a hypothetical il-
lustration, suppose the 1929 stock market consisted of 125 separate markets for 125 differ-
ent stocks, and assume all of them were of the same size and turnover. Compared with one
large integrated market, 125 small markets would absorb the same shock 1252=3 ¼ 25
times more slowly and its impact would be 1251=3 ¼ 5 times smaller, as implied by
Equations (9) and (15).
4.6 Early Warning Systems May Be Useful and Practical

Some strategies are inherently destabilizing. They have built-in features of negative feed-
backs: as prices go down, more selling is required and this pushes prices further down. The
more capital is invested into these strategies, the bigger is their potential destabilizing
effects on prices. Equipped with quantitative invariance formulas for market impact, one
may detect instances when destabilizing strategies become so large that they may put finan-
cial markets at risk.
Tuzun (2012) uses invariance to assess the effect of leveraged ETFs on markets. He finds
that short ETFs and leveraged long ETFs in financial stocks were close to the tipping point
in 2008 and 2009. A price decline of 1% would induce leveraged ETFs to sell about $1 bil-
lion; invariance implies that this imbalance would lead to a further price decline of another
1% and thus potentially trigger a downward spiral.
For some of the five crash events in our article, policymakers or stock market partici-
pants also had in hand the information required to quantify the price impact and foresee
the systemic risks looming from sudden liquidations of large stock market exposures.
Instead, they mistakenly trusted in conventional intuition when assessing the potential mag-
nitudes of price declines, dismissing the danger of meltdowns.
Contrary to the beliefs of some, the market crashes in 1929 and 1987 were not com-
pletely unexpected. In both cases, data were publicly available before the events. Data on
broker loans were published by the Federal Reserve System and the NYSE before the 1929
crash. Estimates of assets under management by portfolio insurers were available before the
1987 crash.
In both cases, the potential price impact of liquidations was a topic of public discussion
among policymakers and market participants. In the months prior to the 1929 stock market
crash, brokers were raising margin requirements to protect themselves from a widely dis-
cussed collapse in prices which might be induced by rapid unwinding of stock investments
financed with margin loans. Market participants watched statistics on broker loans careful-
ly, noting the tendency for total lending in the broker loan market to increase as the stock
market rose. Markets were aware that margin account investors were buyers with “weak
hands,” likely to be flushed out of their positions by margin calls if prices fell significantly.
Discussions about who would buy stocks if a collapse in stock prices forced margin account
investors out of their positions resembled similar discussions in 1987 concerning who
would take the opposite side of portfolio insurance trades.
The debate about the extent to which portfolio insurance contributed to the 1987 crash
started long before the crash itself. On the day the 1987 crash occurred, academics were
holding a conference on a topic of potential “market meltdown” induced by portfolio in-
surance sales. The term “market meltdown,” popularized by then NYSE chairman John
Phelan, was used in the year or so before the stock market crash to describe a scenario of

cascading portfolio insurers’ sell orders resulting in severe price declines and posing system-
ic risks to the economy. Months before the 1987 crash itself, the SEC’s Division of Market
Regulation, Securities and Exchange Commission (1987)—responding to worries that port-
folio insurance had made the market fragile—published a study describing in some detail a
potential meltdown scenario induced by portfolio insurance sales, which closely resembled
the subsequent crash in October 1987, but the study dismissed the risk of such an event as a
remote possibility, accepting the conventional wisdom at the time.
Many market participants were firmly convinced that, given the substantial trading vol-
ume in the US equity markets—and especially the index futures market—there was enough
liquidity available to accommodate sales of portfolio insurers without any major down-
ward adjustment in stock prices. During hearings before the House Committee on Energy
and Commerce (1987) on July 23 prior to the 1987 crash, Hayne E. Leland defended port-
folio insurance:
We indicated that average trading will amount to less than 2% of total stocks and derivatives
trading. On some days, however, portfolio insurance trades may be a greater fraction. . . .In the
event of a major one-day fall (e.g., 100 points on the Dow Jones Industrial Average), required
portfolio insurance trades could amount to $4 billion. Almost surely this would be spread over
2–3 day period. In such a circumstance, portfolio insurance trades might approximate 9–12%
of futures trading, and 3–4% of stock plus derivatives trading.
If regulators had applied simple principles of invariance prior to the 1987 crash, they
would have been alarmed by Hayne Leland’s projection of potential sales of 4% of stock-
plus-futures volume over 3 days in response to a decline in stock prices of about 4% (i.e.,
100 points on the Dow Jones average). They would see that the stock market was already
close to a tipping point. Historical volume and volatility in July 1987 implied that sales of
$4 billion in response to a 4% price decline would lead to another drop in prices, just slight-
ly smaller than 4%. Absent stabilizing trades by investors trading in an opposite direction,
potential portfolio insurance sales were already on the verge of triggering precisely the cas-
cade meltdown scenario practitioners dismissed as a near-zero probability event.
5. Conclusion
Crash-like events continue to occur. The Staffs of the Fed, the CFTC, and SEC (2015) de-
scribe the “flash rally” in the US Treasury market on October 15, 2014, during which pri-
ces rose rapidly for several minutes and then fell back down. Since the report was not based
on audit trail data identifying individual traders, it does not rule out the possibility that the
flash rally resulted from rapid buying by one trader. Obizhaeva and Piftankin (2023) de-
scribe how the sharp V-shaped devaluation of Russian currency on December 16, 2014,
was likely caused by a large multi-billion-dollar bet. The collapse of the Chinese stock mar-
ket in the summer of 2015 was likely caused by liquidations of margin accounts, as dis-
cussed in Bian et al. (2018); this crash was in many ways similar to the crash of 1929 in the
US market; in both cases, extraordinary steps were taken to stabilize the markets.
Our study offers several practical insights about why stock market crashes happen, how
to prevent them if possible, and how to respond to them if they occur. Large price disloca-
tions may occur because of imbalances in demand and supply. Even in highly liquid mar-
kets and even if quantities traded are restricted to 5% or 10% of daily volume, execution of
large bets may lead to significant price changes. Rapid execution is likely to magnify transi-

tory impact. Heavy selling in one market is likely to spread to economically related mar-
kets. Policies aiming at easing flow of credit and providing funds to make up the gap in
demand and supply may help to mitigate the adverse effects of crashes, but the amount of
funds necessary must be comparable to the size of the shock itself. Early warning systems
may potentially help to prepare for crashes and act accordingly to mitigate or avoid them.
Data Availability
The data underlying this article were obtained from public sources. Stock market data were
obtained from CRSP. Index futures data were obtained from Bloomberg. Other data were
obtained from cited documents.
Appendix A: Implementation Issues

In order to apply microstructure invariance to data on the five crash events, several imple-
mentation issues need to be addressed.
First, the volume and volatility inputs into our formulas should not be thought of as
parameters of narrowly defined markets of a particular security in which a bet is placed but
rather as parameters based on the market as a whole. Securities and futures contracts may
share the same fundamentals and have a common factor structure. When a large order
moves prices in the S&P 500 futures market, index arbitragers usually insure that prices for
the underlying basket of stocks move by about the same amount as well. It is difficult to
identify the boundaries of the market. Consistent with the spirit of the Brady Report, we
take the admittedly simplified approach of adding together cash and futures volume for
three of the four crash events in which stock index futures markets existed. In our analysis
of the Soros trades, we ignore cash market volume because his trades were executed so
quickly that price pressure in the futures market was not transferred to cash markets, which
had not yet opened at the time of his trades.
Second, the spirit of the invariance hypothesis is that volume and volatility inputs into
the market impact [Equation (15)] are market expectations prevailing before the bet is
placed. Expected volume and expected volatility determine the sizes of bets investors are
willing to make and the market depth intermediaries are willing to provide. Different price
impact estimates are possible, depending on whether volatility estimates are based on
implied volatilities before the crash, implied volatilities during the crash, historical volatil-
ities based on the crash period itself, or historical volatilities based on months of data be-
fore the crash. For robustness, we present results based on historical data for different
windows prior to the crash event.
Third, it is likely that the price impact of an order is related to the speed with which it is
executed. The market impact model [Equation (15)] assumes that orders are executed at a
“normal” speed in the relevant units of business time. For example, a very large order in a
small stock may be executed over several weeks or even months, while a large order in the
stock index futures market may be executed over several hours. The impact model leaves
open the possibility that unusually rapid execution of very large orders may increase their
temporary price impact, but these effects are hard to quantify properly. We discuss this
issue further in Section 4.
Fourth, there have been numerous changes in market mechanisms between 1929 and
2010, including better communications technologies, introduction of electronic handling of

orders, a reduction in tick size, and the migration of trading volume from face-to-face trad-
ing floors to anonymous electronic platforms. Such changes may have lowered bid–ask
spreads, but we believe—in the spirit of Black (1971)—that they have had little effect on
market depth, which largely determines the price impact of large bets. We thus apply esti-
mates of market depth based on portfolio transitions during 2001–05 to the entire period
1929–2010.
Fifth, Kyle and Obizhaeva (2016) calibrate both linear and square root impact models
consistent with invariance. From an empirical perspective, the square root specification
explains price impact for individual stocks somewhat better than the linear model, as con-
sistent with the empirical econophysics literature (Bouchaud, Farmer, and Lillo, 2009), but
the linear model explains the price impact of the largest 1% of bets in the most active stocks
slightly better than the square root model. Crash events are explained by applying invari-
ance to a linear model. To make this point, “invariance” implicitly assumes a linear impact
function in the main part of the article. Due to its concavity, the square root model predicts
much smaller price declines during crash events. Appendix C presents these estimates along
with estimates based on alternative models.
Appendix B: Estimates for Different Market Impact Models

We compute estimates of predicted price changes based on several alternative models of
market impact. Market impact is expected to depend on market characteristics such as mar-
ket capitalization N, daily share volume V, returns volatility r, and the corresponding GDP
deflator dgdp ; unsigned bet size Q; and perhaps the time horizon T over which the bet is
executed.
We consider several specifications when calculating the implied magnitudes of simple
(non-logged) market impacts DP=P:
• The invariance-implied linear model (“Inv-LIN”), discussed in Kyle and Obizhaeva

(2016):
1=3
DP 2 2:50 P V dgdp r 4=3 Q
¼ 4
6
: (16)
P 10 40 10 0:02 ð0:01Þ V
• The invariance-implied square-root model (“Inv-SQRT”), discussed in Kyle and

Obizhaeva (2016):
1=2
DP 2 12:08 r Q
¼ : (17)
P 104 0:02 ð0:01Þ V
• The conventional model (“Conv-N”), based on market capitalization:
DP Q
¼ : (18)
P N
• The Conv-V, based on daily volume:
DP Q
¼ : (19)
P 250 V
• The Barra model, discussed in Torre and Ferrari (1999) and Grinold and Kahn (2000):
1=2
DP Q
¼r : (20)
P V

• Almgren–Chriss (“AC”) model, discussed in Almgren et al. (2005):
1=4 3=5
DP Q N Q
¼ 0:314 r þ 2 0:142 r : (21)
P V V VT
• Frazzini–Israel–Moskowitz (“FIM”) model, discussed in Frazzini, Israel, and

Moskowitz (2018) in Table AI, Column (9):
1=2 !
DP Q Q pffiffiffiffiffiffiffiffi 2
¼ 0:2 lnð1 þ N 109 dgdp Þ þ 0:35 þ 9:32 þ 0:13 r 252 100 4 :
P 0:01 V 0:01 V 10
(22)
In the last two models, the estimates are multiplied by a factor of 2 to convert transaction
costs estimates to price impact estimates.
The AC model [Equation (21)] explicitly depends on the execution horizon T. For the
1929 crash, we assume selling occurred over 5 days (T ¼ 5). For the 1987 crash, we assume
selling occurred over 4 days (T ¼ 4). For Soros’ trades, we assume selling occurred over
6 min from 8:28 a.m. to 8:34 a.m. (T ¼ 6/420). For the liquidation of Kerviel’s trades, we
assume selling occurred over 3 days (T ¼ 3). For the flash crash of 2008, we assume selling
occurred over about 20 min, or 1/20 of a day (T ¼ 1/20).
Panel A of Table AI presents impact estimates based on six impact models for percentage
market impact along with actual price declines for the five crashes. First, all estimates are
much lower than actual price declines, except for the Inv-LIN estimates. Second, the Conv-
N and Conv-V estimates based on the conventional intuition usually generate the smallest
estimates among models. Third, calibrated on the sample of institutional transactions, the
Barra, AC, and FIM estimates are all similar in magnitude; they are slightly larger than con-
ventional estimates but still much lower than the actual price declines. Fourth, the AC esti-
mate is significantly larger than other alternative estimates for the Soros bet because this
estimate explicitly accounts for the very short execution horizon of this bet.
For all five crashes, the Inv-SQRT estimates are quantitatively similar to the Barra esti-
mates. Due to its concavity, the square-root models predict much smaller price declines
than the linear model. Thus, invariance alone does not explain magnitudes of price declines
during crash events; instead, crash events are explained by applying invariance to a linear
model.
Panel B of Table AI presents impact estimates based on six impact models for log-
percentage market impact DlnP along with actual price declines for the five crashes. These
estimates are obtained from models (16)–(22), where DP=P on the left-hand side of these
equations is replaced with DlnP. The Inv-LIN and the Conv-N based on market capitaliza-
tion for log-impact are the two models discussed on detail in the main part of our article.
The estimates based on log-returns are smaller than the estimates based on simple
returns, but this difference is negligible for most models. The only exception is the Inv-LIN
Table AI. Alternative models
The table presents actual price declines for five market crashes along with price declines as
implied by seven price impact models: the Inv-LIN, the Inv-SQRT, the Conv-N based on market
capitalization, the Conv-N based on daily volume, the Barra model (“Barra”), AC model, and
FIM model. Panel A shows the estimates for models with simple returns DP =P and panel B

shows the estimates for models with log returns DlnP .
Panel A: Simple percentage impact DP=P
Crash Event Actual Inv-LIN Inv-SQRT Conv-N Conv-V Barra AC FIM

(%) (%) (%) (%) (%) (%) (%) (%)
1929 market crash 25.00 62.56 3.94 1.36 1.06 3.26 6.62 4.95
1987 market crash 32.00 18.31 1.33 0.63 0.27 1.10 1.04 2.02
1987 Soros’s trades 22.00 6.47 1.58 0.01 0.01 1.31 3.47 0.62
2008 SocGén trades 9.44 11.40 0.70 0.43 0.11 0.58 0.35 1.18
2010 flash crash 5.12 0.61 0.16 0.03 0.01 0.13 0.16 0.24
Panel B: Log-percentage impact DlnP
Crash Event Actual Inv–LIN Inv–SQRT Conv–N Conv–V Barra AC FIM

(%) (%) (%) (%) (%) (%) (%) (%)
1929 market crash 25.00 46.43 3.86 1.36 1.06 3.21 6.41 4.83
1987 market crash 32.00 16.77 1.32 0.63 0.27 1.10 1.04 1.99
1987 Soros’s trades 22.00 6.27 1.57 0.01 0.01 1.30 3.41 0.62
2008 SocGén trades 9.44 10.79 0.70 0.43 0.11 0.58 0.35 1.17
2010 flash crash 5.12 0.61 0.16 0.03 0.01 0.13 0.16 0.24
model, for which large estimates based on the simple return are reduced when log-returns
are used instead; the biggest difference is observed for the 1929 crash, for which the simple-
return model implies price decline of 63% and the log-return model implies price decline of
only 46%.
Appendix C: Estimation Details

Estimation Details for the Crash of 1929
A significant portion of stock investments in the late 1920s was made in leveraged margin
accounts. To finance their leveraged purchases of stocks, individuals and non-financial corpo-
rations relied either on bank loans collateralized by securities or on margin account loans at
brokerage firms. When investors borrowed through margin accounts at brokerage firms, the
brokerage firms financed only a modest portion of the loans with credit balances from other
customers. To finance the rest, brokerage firms pooled securities pledged as collateral by cus-
tomers under the name of the brokerage firm (in “street name”) and then re-hypothecated
these pools by using them as collateral for broker loans. The broker loan market of the late
1920s resembled the shadow banking system of the early 2000s in its lack of regulation, per-
ceived safety, and the large fraction of overnight or very short maturity loans.
The broker loan market was controversial during the 1920s, just as the shadow banking
system was controversial during the period surrounding the financial crisis of 2008–09.
Some thought the broker loan market should be tightly controlled to limit speculative trad-
ing in the stock market on the grounds that lending to finance stock market speculation
diverted capital away from more productive uses in the real economy. Others thought it
was impractical to control lending in the market because the shadow bank lenders would
find ways around restrictions and lend money anyway. The New York Fed chose to dis-

courage New York banks from lending money against stock market collateral. As a result,
loans to brokers by New York banks declined after reaching a peak in 1927.
Attracted by the high interest rates on broker loans—typically 300 basis points or more
higher than loans on otherwise similar money market instruments—non-New York banks
and non-bank lenders continued to supply capital to the broker loan market. Many of these
loans were arranged by the New York banks; sometimes, non-bank lenders bypassed the
banking system entirely, making loans directly to brokerage firms.
Investment trusts (similar to closed end mutual funds) placed a large fraction of the
newly raised equity into the broker loan market rather than buying expensive common
stocks. Corporations, flush with cash from growing earnings and proceeds of securities issu-
ance, invested a large portion of these funds in the broker loan market rather than in new
plant and equipment.
To quantify the margin selling which occurred during the last week of October 1929, we
follow the previous literature and contemporary market participants by estimating margin
selling indirectly from data on broker loans and bank loans collateralized by securities.
In the 1920s, data on broker loans came from two sources. First, the Fed collected week-
ly broker loan data from reporting member banks in New York City supplying the funds or
arranging loans for others. Second, the New York Stock Exchange collected monthly bro-
ker loan data based on demand for loans by NYSE member firms. The broker loan data
reported by the New York Stock Exchange include broker loans which non-banks made
directly to brokerage firms without using banks as intermediaries; such loans bypassed the
Fed’s reporting system. Since loans unreported to the Fed fluctuated significantly around
the 1929 stock market crash, we rely relatively heavily on the NYSE numbers in our ana-
lysis below but also pay careful attention to the weekly dynamics of the Fed series for meas-
uring selling pressure during the last week of October 1929.
We calculate weekly proxies for margin sales as follows. (1) We difference the weekly
Fed series to construct weekly changes. (2) We interpolate the monthly NYSE series to con-
struct a weekly series by assuming that these loans changed at a constant rate within each
month, except for October 1929. For October 1929, the Fed series shows little change, ex-
cept for the last week, and we therefore assume that the entire monthly change in the NYSE
series represents unreported changes in broker loans which occurred during the last week of
October 1929. (3) Finally, we add changes in bank loans collateralized by securities to take
into account the fact that some changes in broker loans do not represent margin sales be-
cause they were converted into bank loans collateralized by securities. The last adjustment
also has a significant effect because there was an unprecedented increase in banks loans col-
lateralized by securities during the last week of October 1929, followed by offsetting reduc-
tions during November.
Figure AI shows the weekly levels of the Fed’s broker loan series and the monthly levels
of the NYSE broker loan series. Two versions of each series are plotted, one with bank
loans collateralized by securities added and one without (“Fed Broker Loans,” “Fed Broker
Loans þ Bank Loans,” “NYSE Broker Loans,” “NYSE Broker Loans þ Bank Loans”). The
figure also shows the level of the Dow Jones Industrial Average from 1926 to 1930. The

Figure AI. Broker loans and the 1929 market crash. The figure shows weekly dynamics of seven varia-
bles from January 1926 to December 1930: NYSE broker loans (red solid line), Fed broker loans (red
dashed line), the sum of NYSE broker loans and bank loans (black solid line), the sum of Fed broker
loans and bank loans (black dashed line), changes in NYSE broker loans (red bars), changes in the
sum of NYSE broker loans and bank loans (black bars), and the Dow Jones average (in blue). Monthly
levels of NYSE broker loans are marked with solid dots. Weekly levels of NYSE broker loans are
obtained using a linear interpolation from monthly data, except for October 1929, when all changes in
NYSE broker loans are assumed to occur during the last week.
time series on both broker loans and stock prices follow similar patterns, rising steadily
from 1926 to October 1929 and then suddenly collapsing. According to Fed data, broker
loans rose from $3.141 billion at the beginning of 1926 to $6.804 billion at the beginning
of October 1929. According to NYSE data, the broker loan market rose from $3.513 bil-
lion to $8.549 billion during the same period. As more and more non-banks were getting
involved in the broker loan market, the difference between NYSE broker loans and Fed
broker loans steadily increased until the last week of October 1929, when non-bank firms
pulled their money out of the broker loan market and the difference suddenly shrank.
During the period 1926–30, weekly changes in broker loans were typically small and
often changed sign, as shown in the tiny bars at the bottom of Figure AI. Starting with the
last week of October 1929, there were five consecutive weeks of large negative changes, al-
most twenty times larger than changes during preceding weeks. This de-leveraging erased
the increase in broker loans which had occurred during the first 9 months of the year.
For the last week of October 1929, we estimate margin selling as $1.181 billion (the dif-
ference between the estimated reduction in broker loans of $2.440 billion from $8.549 bil-
lion to $6.109 billion and an increase in bank loans on securities of $1.259 billion from
$7.920 billion to $9.179 billion). For the 3 months from September 30, 1929, to December
31, 1929, we estimate margin selling as $4.348 billion (the difference between the reduction
in NYSE broker loans of $4.559 billion from $8.549 billion to $3.990 billion and an in-
crease in bank loans on securities of $0.211 billion from $7.720 billion to $7.931 billion).
Estimation Details for the Crash of 1987

Along with our main estimates in Table III, we present several other estimates for robust-

ness. First, some of the market participants classified as portfolio insurers in the Brady
Report abandoned their portfolio insurance strategies as prices crashed and switch to buy-
ing securities. Even though we believe that for the purpose of analyzing the price impact of
portfolio insurance sales it is better to use the gross sales amount, we also report estimates
for net sales of $11.11 billion of futures contracts and stocks combined ($9.51 billion in
futures and $1.60 billion in stocks). Their predicted impact ranges from 9.71% to 13.78%.
Second, we show implied estimates if we treat markets for futures contracts and NYSE
stocks separately. To avoid radically different price impacts in two markets, we adjust
quantities sold in both markets by the NYSE’s estimate of net NYSE index-arbitrage sales
of $3.27 billion (Brady Report, Figures 13 and 14). We add this number to portfolio insur-
ance sales in NYSE stocks and subtract the same amount from portfolio insurance sales in
the futures market because arbitrageurs transferred some price pressure from futures to
stocks. This results in net sales of $7.21 billion in the futures market with impact ranging
from 10.00% to 14.11% and $6.54 billion in NYSE stocks with impact ranging from
9.09% to 13.00%. The fact that index-arbitrage sales make price impact estimates similar
in both markets is consistent with the interpretation that portfolio insurance sales were in-
deed driving price dynamics in both markets.
Estimation Details for Liquidation of Kerviel’s Trades in 2008

We also examine whether implied cost estimates are consistent with officially reported
losses of e6.30 billion. We assume that average impact cost is equal to half of predicted
price impact since—assuming no leakage of information about the trades—a trader can the-
oretically walk the demand curve, trading only the last contracts at the worst expected pri-
ces. Accounting for compounding, invariance predicts that the total cost of unwinding
Kerviel’s position is equal to 5.55% of the initial e50 billion position, that is, e2.77 billion.
Officially reported losses also include mark-to-market losses sustained by hidden naked
long positions as markets fell from the end of the previous reporting period on December
31, 2007, to the decision to liquidate the positions when the market re-opened after
January 18, 2008. From December 28, 2007, to January 18, 2008, the Euro STOXX 50 fell
by 9.18%, DAX futures fell by 9.40%, and FTSE futures fell by 8.68%. If we assume that
Kerviel held a constant long position from December 31, 2007, to January 18, 2008, then
these positions would have sustained e4.62 billion in mark-to-market losses during that
period. Société Générale reported, however, that Kerviel acquired his hidden long position
gradually over the month of January. If we assume that Kerviel acquired his position grad-
ually by purchasing equal quantities of futures contracts at each lower tick level from the
end-of-year 2007 close to January 18 close, we estimate that such positions would be under
water by only half as much, that is, e2.31 billion, at the close of January 18.
Table V reports that the estimated market impact costs of liquidating the rogue position
range from e2.72 billion to e3.34 billion under different assumptions about expected vol-
ume and volatility. Adding mark-to-market losses sustained prior to liquidation leads to
estimated losses ranging (i) from e5.03 billion to e5.65 billion if positions were acquired
gradually and (ii) from e7.34 billion to e7.96 billion if positions were held from the end of
2007. These estimates are similar in magnitude to losses of e6.30 billion reported by the
bank.
As a robustness check, we also estimate market impact under the assumption that the
Euro STOXX 50, the DAX, and the FTSE 100 futures markets are distinct markets, not
components of one bigger market.

In the month preceding January 18, 2008, historical volatility per day was 98 basis
points for futures on the Euro STOXX 50, 100 basis points for futures on the DAX, and
109 basis points for futures on the FTSE 100. Average daily volume was e55.19 billion for
Euro STOXX 50 futures, e32.40 billion for DAX futures, and £7.34 billion for FTSE 100
futures. Kerviel’s positions of e30 billion in Euro STOXX 50 futures, e18 billion in DAX
futures, and e2 billion in FTSE 100 futures represented about 54%, 56%, and 20% of daily
trading volume in these contracts, respectively. We use an exchange rate of e1.3440 for £1
on January 17.
Our calculations estimate a price impact of 12.08% for liquidation of Kerviel’s position,
10.77% for liquidation of his DAX futures position, and 4.12% for liquidation of his FTSE
futures position. Indeed, from the close on January 18 to the close on January 23, Euro
STOXX 50 futures fell by 10.50%, DAX futures fell by 11.91%, and FTSE 100 futures fell
by 4.65%. Note that from the close on January 18 to the lowest point during January 21–
23, Euro STOXX 50 futures fell by 11.67%, DAX futures fell by 12.71%, and FTSE 100
futures fell by 9.54%. The similarity of actual price declines for the STOXX 50, DAX, and
FTSE suggests substantial integration of European markets, consistent with our strategy of
thinking about them as one market.
From the close on January 18 to low points on January 22, the Spanish IBEX 35, the Italian
FTSE MIB, the Swedish OMX, the French CAC 40, the Dutch AEX, and the Swiss Market
Index fell by 12.99%, 10.11%, 8.63%, 11.53%, 10.80%, and 9.63%, respectively. By
January 24, all of these markets had largely reversed these losses. Euro Stoxx 50 and FTSE
reversed losses as well, but DAX recovered only partially. Large price declines in markets
where Kerviel did not hold positions suggest that the markets are well integrated as well.
Estimation Details for the Flash Crash of May 6, 2010

Since the price drop in the morning may have reset market expectations about volatility, as
a robustness check, we also report results for expected volatility of 2.00% per day; they
range from 1.39% to 1.65%.
If we do not treat the cash market and the futures market as one market but focus only
on the futures market, then the estimates range from 0.76% to 1.29% for historical volatil-
ity and from 2.35% to 2.91% for volatility of 2%.
Appendix D: The Frequency of Market Crashes

Market microstructure invariance can be used to quantify the frequency of crash events,
including both the size of selling pressure and the resulting price impact.
Using portfolio transitions orders as proxies for bets, Kyle and Obizhaeva (2016) find
that the invariant distributions of buy and sell bet sizes can be closely approximated by a
log-normal. The distribution of unsigned bet size X ~ of a stock with expected daily volume
of P V dollars and expected daily returns volatility r can be approximated as a log-normal
!
~
X 2 rPV pffiffiffiffiffiffiffiffiffiffi
ln ¼ 5:71 ln þ 2:53 Z; ~ (23)
V 3 6
ð0:02Þð40Þð10 Þ
~ N ð0; 1Þ. Under the assumption that there is one unit of intermediation trade
where Z
volume for every bet, the bet arrival rate c per day is given by
!

2 rPV
lnðcÞ ¼ lnð85Þ þ ln : (24)
3 ð0:02Þð40Þð106 Þ
These equations have the following implications for a benchmark stock with dollar volume
of $40 million per day and volatility 2% per day1=2 . The estimated mean of –5.71 implies a
median bet size of approximately $132; 500, or 0.33% of daily volume. The estimated log-
variance of 2.53 implies that a one-standard-deviation increase in bet size is a factor of
about 4.91. The implied average bet size is $469,500 and a four-standard-deviation bet is
about $77 million, or 1.17% and 192% of daily volume, respectively [0:33% expð2:53=2Þ
and 0:33% expð2:53 4Þ]. There are 85 bets per day. The standard deviation of daily order
imbalances is equal to 38% of daily volume [851=2 expð5:71 þ 2:53Þ]. Half the variance
in returns results from fewer than 0.10% of bets and suggests significant kurtosis in returns.
Now let us extrapolate these estimates to the entire market, where volume is the sum of
the volume of CME S&P 500 futures contracts and all individual stocks. Using convenient
round numbers based on the 2010 flash crash, the volume for the entire market is about
$270 billion per day or 6,750 times the volume of a benchmark stock. The volatility of the
index is about 1% per day or half of 2% volatility of a benchmark stock. With 6,750 con-
veniently equal to 153 2, invariance implies that market volume consists of 19,125 bets
(85 152 ) with the median bet of about $4 million ($132; 500 15 2) or 0.0014% of daily
volume. The implied average bet size is $14 million or 0.0052% of daily volume, and a
four-standard-deviation bet is $2.310 billion ($469; 500 15 2 and $77 106 15 2) or
0.86% of daily volume. The implied standard deviation of cumulative order imbalances is
2.55% of daily volume (38%=15).
Equations (23) and (24) can be used to predict how frequently crash events occur. The
three large crash events—the 1929 crash, the 1987 crash, and the 2008 Société Générale
trades—are much rarer events than the two smaller crashes—the 1987 Soros trades and the
2010 flash crash.
We estimate the 1929 crash, the 1987 crash, and the 2008 liquidation of Kerviel’s posi-
tions to be 6.15, 5.97, and 6.19 standard deviation bet events, respectively. Given corre-
sponding estimated bet arrival rates of 1,887, 5,606, and 19,059 bets per day, such events
would be expected to occur only once every 5,516, 597, and 674 years, respectively.
Obviously, either the far right tail of the distribution estimated from portfolio transitions is
fatter than a log-normal or the log-variance estimated from portfolio transition data is too
small. In the far right tail of the distribution of the log-size of portfolio transition orders in
the most actively traded stocks, Kyle and Obizhaeva (2016) do observe a larger number
observations than implied by a normal distribution. It is also possible that portfolio transi-
tion orders are not representative of bets in general. If the true standard deviation of log bet
size is 10% larger than implied by portfolio transition orders, then 6.0 standard deviation
events become 5.4 standard deviation events, which are expected to occur about thirty-four
times more frequently.
We estimate the 1987 Soros trades and the 2010 flash crash trades to be 4.45- and 4.63-
standard-deviation bet events, respectively. Given estimated bet arrival rates of 14,579 bets
and 29,012 bets per day, respectively, bets of this size are expected to occur multiple times
per year. We believe it likely that large bets of this magnitude do indeed occur multiple
times per year, but execution of such large bets typically does not lead to flash crashes be-

cause such large bets would normally be executed more slowly and therefore have less tran-
sitory price impact.
References
Akerlof, G. A. and Shiller, R. J. (2009): Animal Spirits: How Human Psychology Drives the
Economy, and Why It Matters for Global Capitalism, Princeton University Press, Princeton, NJ.
Almgren, R., Thum, C., Hauptmann, E., and Li, H. (2005): Direct estimation of equity market im-
pact, Risk 18, 58–62.
Bernanke, B. S. (2015): The Courage to Act: A Memoir of a Crisis and Its Aftermath, Norton,
New York, NY.
Bian, J., He, Z., Shue, K., and Zhou, H. (2018): Leverage-induced fire sales and stock market
crashes. Working paper.
Black, F. (1971): Towards a fully automated exchange, Part I, Financial Analysts Journal 27,
28–35.
Board of Governors of the Federal Reserve System. (1927–1931): Annual Report of the Board of
Governors of the Federal Reserve System. 13th through 17th Annual Reports Covering
Operations for the Years 1926–1930.
Board of Governors of the Federal Reserve System. (1929): Federal Reserve Bulletin, multiple
issues, Board of Governors of the Federal Reserve System, Washington, DC.
Board of Governors of the Federal Reserve System. (1943): Banking and Monetary Statistics,
1914–1941, Board of Governors of the Federal Reserve System, Washington, DC.
Board of Governors of the Federal Reserve System. (2008): Meeting of the Federal Open Market
Committee on January 29–30, 2008, Board of Governors of the Federal Reserve
System, Washington, DC.
Bouchaud, J.-P., Farmer, D. J., and Lillo, F. (2009): How markets slowly digest changes in supply
and demand, in: Hens, T. and Schenk-Hoppé, K. R. (eds.), Handbook of Financial Markets:
Dynamics and Evolution, North Holland, Amsterdam, pp. 47–160.
Brennan, M. J. and Schwartz, E. S. (1989): Portfolio insurance and financial market equilibrium,
Journal of Business 62, 455–472.
Chan, L. K. C. and Lakonishok, J. (1995): The behavior of stock prices around institutional trades,
Journal of Finance 50, 1147–1174.
Chan, L. K. C. and Lakonishok, J. (1997): Institutional equity trading costs: NYSE versus Nasdaq,
Commodity Futures Trading Commission, Division of Economic Analysis and Division of
Trading and Markets. (1988): Final Report on Stock Index Futures and Cash Market Activity
during October 1987 to the U.S. Commodity Futures Trading Commission, U.S. Government
Printing Office, Washington, DC.
Division of Market Regulation, Securities and Exchange Commission. (1987): The Role of
Index-Related Trading in the Market Decline of September 11 and 12, 1986, U.S. Government
Frazzini, A., Israel, R., and Moskowitz, T. J. (2018): Trading costs. Working paper.
Friedman, M. and Schwartz, A. J. (1963): A Monetary History of the United States, 1867–1960,
NBER Book Series Studies in Business Cycles, Princeton University Press, Princeton, NJ.
Friedman, M. and Schwarz, A. (1963): A Monetary History of the United States, Princeton
University Press.
Gabaix, X. (2009): Power laws in economics and finance, Annual Review of Economics 1,
255–294.
Galbraith, J. K. (1954): The Great Crash, 1929, Houghton Mifflin Co., Boston.
Gammill, J. F. and Marsh, T. A. (1988): Trading activity and price behavior in the stock and stock

index futures markets in October 1987, Journal of Economic Perspectives 2, 25–44.
Grinold, R. C. and Kahn, R. N. (2000): Active Portfolio Management, McGraw Hill, New York,
NY.
Haney, H. L. (1932): Brokers’ loans, Journal of the American Statistical Association 27, 65–72.
Harris, L. and Gurel, E. (1986): Price and volume effects associated with changes in the S&P 500
list: New evidence for the existence of price pressures, Journal of Finance 41, 815–829.
House Committee on Energy and Commerce, U.S. Congress (1987): Program Trading: Hearing
before the Subcommittee on Telecommunications and Finance of the Committee on Energy and
Commerce, Vol. 100th Cong., 1st Sess., testimony of Hayne E. Leland, U.S. Government
Huberman, G. and Stanzl, W. (2004): Price manipulation and quasi arbitrage, Econometrica 72,
1247–1275.
Keim, D. B. and Madhavan, A. (1997): Transaction costs and investment style: An inter-exchange
analysis of institutional equity trades, Journal of Financial Economics 46, 265–292.
Keynes, J. M. (1936): The General Theory of Employment, Interest and Money, Palgrave
Macmillan, London.
Kraus, A. and Stoll, H. R. (1972): Price impact of block trading on the New York Stock Exchange,
Kyle, A. S. (1985): Continuous auctions and insider trading, Econometrica 53, 1315–1335.
Kyle, A. S. and Obizhaeva, A. A. (2016): Market microstructure invariance: Empirical hypotheses,
Econometrica 84, 1345–1404.
Kyle, A. S. and Obizhaeva, A. A. (2018a): Adverse selection and liquidity: From theory to practice.
Working paper, available at SSRN: https://ssrn.com/abstract=3236030 or http://dx.doi.org/10.
2139/ssrn.3236030
Kyle, A. S. and Obizhaeva, A. A. (2018b): Dimensional analysis, leverage neutrality, and market
microstructure invariance. Working paper, available at SSRN: https://papers.ssrn.com/sol3/
papers.cfm?abstract_id=2785559 or https://dx.doi.org/10.2139/ssrn.2785559
Kyle, A. S. and Obizhaeva, A. A. (2020): Market microstructure invariance: A dynamic equilib-
rium model. Working paper, available at SSRN: https://ssrn.com/abstract=3326889 or http://
dx.doi.org/10.2139/ssrn.3326889
Kyle, A. S., Obizhaeva, A. A., and Wang, Y. (2018): Smooth trading with overconfidence and mar-
ket power, Review of Economic Studies 85, 611–662.
Leland, H. E. and Rubinstein, M. (1988): Comments on the market crash: Six months after,
Journal of Economic Perspectives 2, 45–50.
Miller, M. H. (1988): Explaining the events of October 1987, in R. J. Mackay (ed.), After the
Crash: Linkages between Stocks and Futures, American Enterprise Institute for Public Policy
Research, University Press of America, AEI Studies.
Miller, M. H. (1991): Financial Innovations and Market Volatility, Blackwell, Cambridge,
Massachusetts.
Obizhaeva, A. A. and Piftankin, G. (2023): The Russian Rouble Crisis of December 2014:
Structure and liquidity of a foreign exchange market, Russian Journal of Money and Finance 82,
104–136.
Perold, A. (1988): The implementation shortfall: Paper vs. reality, Journal of Portfolio
Management 14, 4–9.
Reinhart, C. M. and Rogoff, K. (2009): This Time Is Different: Eight Centuries of Financial Folly,
Princeton University Press, Princeton, NJ.
Roll, R. (1988): The international crash of October 1987, Financial Analysts Journal 44, 19–35.
Scholes, M. S. (1972): The market for securities: Substitution versus price pressure and the effects
of information on share prices, Journal of Business 45, 179–211.
Senate Committee on Banking and Currency, U.S. Congress (1934): Stock Exchange Practices:

Report of the Committee on Banking and Currency Pursuant to S. Res. 84 (72nd Congress) and
S. Res. 56 and S. Res. 97 (73rd Congress), Vol. 73rd Congress, Second Session, U.S.
Government Printing Office, Washington, DC, “The Pecora Commission Report.”
Shiller, R. J. (1987): Investor behavior in the October 1987 stock market crash: Survey evidence.
Working Paper 2446, NBER.
Smiley, G. and Keehn, R. H. (1988): Margin purchases, brokers’ loans and the bull market of the
twenties, Business and Economic History 17, 129–142.
Société Générale. (2008a): Explanatory note about the exceptional fraud, Paris, January 27, 2008.
Société Générale. (2008b): Mission green, summary report: Interim conclusions as of February 20,
2008, General Inspection Department.
Société Générale. (2008c): Mission green, summary report, General Inspection Department, May
20, 2008.
Staffs of the CFTC and SEC. (2010a): Preliminary Findings Regarding the Market Events of May
6, 2010, Report of the Staffs of the CFTC and SEC to the Joint Advisory Committee on
Emerging Regulatory Issues. May 18, 2010, “Preliminary Report,” U.S. Government Printing
Office, Washington, DC.
Staffs of the CFTC and SEC. (2010b): Findings Regarding the Market Events of May 6, 2010,
Report of the Staffs of the CFTC and SEC to the Joint Advisory Committee on Emerging
Regulatory Issues. September 30, 2010, “Final Report,” U.S. Government Printing Office,
Washington, DC.
Staffs of the Fed, the CFTC, and SEC. (2015): The U.S. Treasury Market on October 15, 2014,
joint staff report by U.S. Department of the Treasury, Board of Governors of the Federal
Reserve System, Federal Reserve Bank of New York, U.S. Securities and Exchange Commission,
U.S. Commodity Futures Trading Commission, July 13, 2015, U.S. Government Printing Office,
Washington, DC.
Torre, N. G. (1997): BARRA Market Impact Model Handbook, BARRA Inc., Berkeley, CA.
Torre, N. G. and Ferrari, M. J. (1999): The Market Impact Model, Barra, Berkeley, CA.
Tuzun, T. (2012): Are leveraged and inverse ETFs the new portfolio insurers? Working paper.
U.S. Presidential Task Force on Market Mechanisms. (1988): Report of the Presidential Task
Force on Market Mechanisms, U.S Government Printing Office, Washington, “Brady Report,”
U.S. Government Printing Office, Washington, DC.
Wurgler, J. A. and Zhuravskaya, E. (2002): Does arbitrage flatten demand curves for stocks?,
Journal of Business 75, 583–608.

Large Bets and Stock Market Crashes

Uploaded by

Copyright:

Available Formats

Large Bets and Stock Market Crashes

Uploaded by

Document Information

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

Large Bets and Stock Market Crashes

Uploaded by

Copyright:

Available Formats

Review of Finance, 2023, 1–41

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Table I. Summary of five crash events

Crash Event Actual Predicted Predicted %ADV %GDP

1929 market crash 25 46.43 1.36 265.41 1.136

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

weighted average of quantitative benefits multiplied by quantitative probabilities.” Akerlof

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

1. Market Impact of Large Bets: Previous Literature

1.1 Conventional Wisdom

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

1.2 Estimates from Institutional Trades

2. Market Impact of Large Bets: Invariance

2.1 Review of Invariance

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

2.2 Business Time

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

ac¼W and a c1=2 ¼ C: (6)

The solution for a and c is

2.3 Distribution of Bet Size

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

2.4 Price Impact

2.5 Intuition behind Equivalent Bets

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

correspond to about 6-standard-deviation events. This suggests it would not be surprising

2.6 Invariance-Implied Market Impact Formulas

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

2.7 A Market Crash Scenario

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

3.1 The Stock Market Crash of October 1929

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

3.2 The Market Crash in October 1987

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Months preceding October 24, 1929

ADV (1929 $M) 444.97 461.45 436.49 427.20 387.18 390.45

9 The GDP deflator of 1.54 converts 1987 dollars to 2005 dollars.

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Months preceding October 14, 1987

3.3 Trades of George Soros on October 22, 1987

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

10 The GDP deflator of 1.54 converts 1987 dollars to 2005 dollars.

Table IV. October 22, 1987: effect of Soros’s trades

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Months preceding October 22, 1987

3.4 Liquidation of Kerviel’s Rogue Trades in January 2008

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Table V. January 2008: effect of liquidating Kerviel’s positions

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Months preceding January 18, 2008

According to Equation (15), the liquidation of Kerviel’s e50 billion position—equal to

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

3.5 The Flash Crash of May 6, 2010

Downloaded from https://academic.oup.com/rof/advance-article/doi/10.1093/rof/rfad008/7091473 by guest on 01 June 2023

Months preceding May 6, 2010

4. Policy Implications and Lessons Learned

ac¼W and a c1=2 ¼ C: (6)