A Machine Learning Approach to

Regime Modeling

Written by Alex Botte and Doris Bao

10/21 Executive summary

Market participants are often interested in understanding market regimes, how they
change over time, and how each regime might affect their portfolio. There are many
approaches to modeling market regimes. In this Street View, we offer a data-driven
approach by applying a Gaussian Mixture Model (a machine learning method) to the
factors in the Two Sigma Factor Lens.

© 2021 Two Sigma. All Rights Reserved. “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP. twosigma.com
Introduction each of those distributions would have its own properties,
like means and volatilities. In Exhibit 1 below, we show an
Financial markets have the tendency to change their
illustrative example of how a GMM might model a single time
behavior over time, which can create regimes, or periods
series. The green Cluster 2 captures the center part of the
of fairly persistent market conditions. Investors often look
asset’s return data, while the red and blue Clusters 1 and 3
to discern the current market regime, looking out for any
capture the tails.
changes to it and how those might affect the individual
components of their portfolio’s asset allocation. Modeling Thus, the GMM is able to use a combination of normal
various market regimes can be an effective tool, as it can distributions to model both the center and the tails of an
enable macroeconomically aware investment decision- asset’s distribution. We believe this is an especially helpful
making and better management of tail risks. method for modeling financial assets, as their return
distributions can often exhibit skew with a meaningful
In this Street View, we present a machine learning-based
number of observations in the tails.
approach to regime modeling, display the historical results
of that model, discuss its output for today’s environment,
Exhibit 1: Illustrative Example of Three Gaussians that
and conclude with practical use cases of this analysis for Describe a Single Time Series
allocators.

Determining the Regimes

In order to understand how a portfolio might react to various
regimes, one first needs to determine what the regimes are.
There are different approaches to establishing regimes.
One way is to specify regimes based on knowledge and
experience in the markets. For example, one might categorize
market regimes using “boom” and “bust” cycles, periods of
high or low equity market volatility, changes in monetary
policy, or “risk-on” versus “risk-off” sentiment, believing
those to be good indicators of meaningfully changing market Source: Towards Data Science as of June 24, 2021

conditions.
We want to provide the GMM with returns data from more
An alternative, more data-driven approach is letting than one asset to have a broader representation of the
historical data on assets and/or market risks delineate overall market and risks. So we will provide the GMM with
the regimes for you. A specific example of this approach the historical returns of the factors in the U.S. version of the
is a Gaussian Mixture Model (GMM), which is a type of Two Sigma Factor Lens, with most of the factor data dating
unsupervised learning method.¹ back to the early 1970s (see Appendix 1 for the start dates
by factor). Instead of modeling the distribution of a single
The GMM uses various Gaussian distributions (another
asset, like we did in Exhibit 1, we will ask the GMM to model
word for a normal, bell curve distribution) to model different
the joint distribution of all 17 factors in the lens.²
parts of the data. As a simple example, imagine we had a
single time series of an asset’s returns. As we know, returns
of financial assets do not always follow a normal distribution.
So a GMM would fit various Gaussian distributions to
capture different parts of the asset’s return distribution, and

1 “Unsupervised learning uses algorithms to analyze and cluster unlabeled datasets. These algorithms discover hidden patterns or data groupings without the need for
human intervention.” Source: https://www.ibm.com/cloud/learn/unsupervised-learning
2 We did not include the recently added Crowding equity style factor in the GMM analysis given data limitations.
2
© 2021 Two Sigma. All Rights Reserved. “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP.
Please see the last page of this document, which contains important disclaimer and disclosure information.
Four Market Conditions Exhibit 3: Factor Volatilities in the Four Market Conditions

The result of the GMM using the Two Sigma Factor

Lens factor data was four different clusters, or what we
think may correspond to four different types of market
conditions.³ As we mentioned earlier, regimes can be
defined as periods of fairly persistent market conditions,
so, in the next section, we will observe the behavior
of these four market conditions over time to identify
regimes.

But first, what are these four market conditions? Each

market condition from our GMM is characterized by a
17-dimensional Gaussian distribution (to account for the
17 factors in the Two Sigma Factor Lens). Each market So, again, while we don’t know exactly what these market
condition’s Gaussian distribution has different factor means conditions are capturing because they were generated
and volatilities, which are displayed in Exhibits 2 and 3 by data and a model, we can imply what each represents
respectively, as well as correlation structures.⁴ by examining the factor behavior in each. Let’s attempt to
classify and describe each.
One of the advantages of the GMM approach is that it is
entirely data-driven—that is, the model outputs various
Market Condition 1: Crisis
market conditions, but that doesn't tell us what those
conditions are intuitively, and they won’t necessarily map Market Condition 1 is likely one of the more interesting
exactly to well-known market environments that we or types of market conditions to investors when building
others could have specified ex-ante. This is both bad and portfolios, and we’ll spend the most time describing it.
good. Bad in the sense that it might be hard for us to put In this market condition, we see that several of the core
intuition behind the resulting market conditions, and good and secondary macro factors exhibited extremely poor
in the sense that it will hopefully tell us something that performance on average. For example, we observe that this
we wouldn’t have known without employing a technique was the only market condition where the global Equity and
like machine learning. After all, if we perfectly knew the Credit factors had negative mean returns, and meaningful
market conditions without bias, why would we use relatively ones to boot. The Emerging Markets factor, which captures
complicated models like GMMs? the risk-adjusted difference between emerging and
developed markets, also averaged a negative return. This
Exhibit 2: Annualized Factor Mean Returns⁵ in the Four means that emerging markets struggled more than their
Market Conditions
developed counterparts. At the same time, the Local
Equity factor averaged a positive return, indicating
U.S. equity markets outperformed global markets on a
risk-adjusted basis. Finally, the Equity Short Volatility
factor was negative; this factor underperforms when
equity market volatility is high.

3 The number of clusters is the only hyper-parameter in this model, and we used a cross-validation method to select the best number of clusters. The criterion used to
measure goodness-of-fit is the log-likelihood. We also tried using other criteria, e.g. AIC and p-value from an empirical goodness-of-fit test for multivariate distributions
(McAssey, 2013). The results are similar to using log-likelihood.
4 Factor means, volatilities, and correlations were fit using data for the period starting on the dates in Appendix 1 and ending in late 2020.
5 The factor means are estimators of the true means, so bear in mind that there are error bounds around these estimated means.
3
© 2021 Two Sigma. All Rights Reserved. “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP.
Please see the last page of this document, which contains important disclaimer and disclosure information.
The Interest Rates factor, representing global sovereign factors were nearly flat, indicating that the U.S. and the
bonds, exhibited a positive mean return, perhaps rest of the world experienced similar risk-adjusted returns.
demonstrating that investors flocked to lower-risk securities Finally, the Local Inflation factor exhibited a small positive
during this type of market condition. (This hypothesis is also mean return, indicating minimal benefit to a U.S. inflation
supported by the positive mean returns of the Low Risk and hedge.
Quality long-short equity style factors.) The Local Inflation
We’ll refer to Market Condition 2 as Steady State.
factor, which attempts to capture the returns of an inflation
hedge, was negative, indicating that U.S. inflation hedges
Market Condition 3: Inflation
didn’t pay off during these periods, as there was likely lower
demand and economic activity when equity markets were In Market Condition 3, the U.S.-specific Local Inflation factor
in crisis. exhibited a double-digit mean return, the highest mean
return for that factor across the four market conditions. This
In terms of the style factors, the equity styles exhibited suggests that U.S. inflation hedges were generally rewarded
mostly positive average performance, with the exception in Market Condition 3.
being Small Cap, indicating that larger cap companies do
better in this market condition. Periods like this, in which We find that the global Equity and Interest Rates factors
overall equity markets are in “crisis mode,” could affect the have small positive mean returns, underperforming most, if
viability of a small cap company more than a larger one, not all, of the other four market conditions (positive inflation
as worsening economic conditions are associated with a shocks tend to be negative for both stocks and bonds).
systematically larger decline in sales and investment for Additionally, central banks might combat higher inflation
smaller firms than larger firms.⁶ We also see that the Trend with higher interest rates, which would also serve as a
Following macro style factor exhibited a large positive headwind for bonds.
return; any directional trend in macro markets will benefit
We see that the Foreign Currency factor exhibited the
this factor, no matter whether the trend is up or down.
highest average return of any factor in this market condition,
Finally, Market Condition 1 exhibited the highest average indicating that the USD underperformed G10 currencies on
absolute correlation between the factors, although it was average. Inflation erodes purchasing power and therefore
still very close to zero. This is by design, as the factors would be expected to coincide with a weaker local currency.
are constructed to be lowly correlated with one another,
Based on the notable performance of the Local Inflation
especially over long periods. However, we do observe factor
factor, we’ll call this market condition Inflation.
correlations rising during shorter, crisis-like periods in both
this analysis and others that we’ve run in the past,⁷ though
Market Condition 4: Walking on Ice
the factor correlations don’t rise to the extreme values seen
in unresidualized asset classes. Market Condition 4 tends to occur around Crisis (and Steady
State) periods,⁸ potentially indicating market fragility. Global
Based on all these observations, we believe the most equity markets (as proxied by the Equity factor) do well here,
appropriate label for Market Condition 1 would be Crisis. but with a higher volatility than their long-term average. In
fact, the Equity factor exhibited its second highest volatility
Market Condition 2: Steady State in Market Condition 4 (the highest was in Crisis, Market
Market Condition 2 seems to cover the most normal and Condition 1). And more generally, factor volatilities were
healthy market periods, as there are no obviously large on average 1.6 percentage points higher in this market
drawdowns for any factor. Equity, Credit, and nearly every condition than their respective long-term averages.
style factor performed well on average. We see that the
mean returns for the Local Equity and Emerging Markets

6 Crouzet, Nicolas and Neil R. Mehrotra (2017). “Small and Large Firms over the Business Cycle,” Research Division Federal Reserve Bank of Minneapolis.
7 See Exhibit 5 in the Two Sigma white paper Introducing the Two Sigma Factor Lens.
8 See Appendix 2 for a detailed breakdown of which market conditions preceded and followed Market Condition 4.
4
© 2021 Two Sigma. All Rights Reserved. “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP.
Please see the last page of this document, which contains important disclaimer and disclosure information.
We also find that most equity style factors’ mean returns The legend at the bottom of the exhibit includes the percent
were above average, with the main exception being of time the GMM found that market condition to have the
Momentum. Additionally, these factors in particular highest probability over this 1971 - 2020 period. Steady
experienced much higher volatilities compared to their State occurred most frequently since 1971, and each of the
long-term averages (e.g., Value exhibited 18.2% volatility in other three market conditions occurred in roughly 15-20%
Market Condition 4 vs. 8.9% long-term; Momentum 19.1% of the periods.
in Market Condition 4 vs. 10.5% long-term; Low Risk 20%
Starting from the top of the exhibit, Inflation was present
in Market Condition 4 vs. 10.4% long-term). This might
exclusively in the 1970s and 1980s, as expected, since that
mean that there is reversal behavior occurring within stocks
period was characterized by relatively high inflation and
exhibiting more choppy returns.
interest rates. Over that first decade and a half, Inflation was
Overall, it looks like this market condition potentially fairly persistent (i.e., limited interruption from other market
captures risk-on market periods where bubbles might exist conditions), as it took quite a bit of time to get soaring prices
or be forming. We’ll label it Walking on Ice (WOI). for goods and services under control. We don’t see Inflation
at any point in the last decade. Perhaps it will enter the
Historical Analysis of the Four Market picture in 2021 or 2022 if inflation does not prove transitory
Conditions (more on that in the next section).

Now that we have an understanding of the various market WOI occurred mostly during the tech bubble in the late
conditions, let’s look back through history to see when each 1990s and early 2000s. Markets were fragile for a while,
occurred. This analysis will be able to tell us the extent to as the bursting of the tech bubble occurred over multiple
which there have been fairly persistent market conditions, or years. There were a handful of Crisis periods during this
regimes, throughout history. time as well, which correspond to days where the market
had relatively large drawdowns, while the WOI periods were
For any given historical period, the GMM will estimate
times where the market either temporarily reversed and/
probabilities that the market was in the four market
or experienced large volatility. WOI was also the highest
conditions. So each market condition will have a probability
probability market condition in the immediate post-crisis
for a particular period, and the four probabilities will sum
performance reversals after the Global Financial Crisis
to 100%. Exhibit 4 shows the highest probability market
(GFC) and COVID market crises, indicating that the market
condition for periods throughout history. We should note
was recovering but still in a fragile state.
that each period displayed in Exhibit 4 is independent and
identically distributed. This means the GMM evaluates each As expected, Crisis occurred during notable periods like
period completely independently, without awareness of what the stock market crash in 1987, the GFC in 2008, and the
market conditions occurred in the past or future. COVID market crisis in 2020.
Exhibit 4: Highest Probability Market Conditions We find that Steady State dominated the last decade with a
Throughout History
sprinkling of short-lived Crisis and WOI periods. This period
coincided with the “decade of the central bank” where the
Federal Reserve and its counterparts around the world
exhibited major influence over the economy and markets.
Over this time, Steady State was interrupted here and there
by some Crisis flare ups (e.g., European Sovereign Debt
Crisis in the early 2010s and the Taper Tantrum in mid-
2013), but the central banks would often step in to steady
the markets through quantitative easing and rate cuts, rarely
allowing WOI periods to form and returning markets to
Steady State.⁹

Steady State, 47% WOI, 16% Crisis, 17% Inflation, 20%

9 https://www.cnbc.com/2019/12/17/decade-of-the-central-bank-ends-as-the-fed-shifts-to-new-paradigm.html
5
© 2021 Two Sigma. All Rights Reserved. “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP.
Please see the last page of this document, which contains important disclaimer and disclosure information.
To wrap-up and bring this back to regimes, we can certainly Exhibit 6: Market Condition Probabilities Since the
discern patterns of fairly persistent market conditions in Beginning of 2020
Exhibit 4 (e.g., Inflation in the 1970s and 1980s, WOI in early Period: January 1, 2020 - July 9, 2021.

2000s, and Steady State in the recent decade). However,

it was rare for one market condition to go for long periods
without interruption from another. Transient themes

Market Condition Probability

can enter the picture on a short-term basis, and market
conditions can change rather rapidly (think COVID market
crisis in February and March 2020, followed by a sharp
rebound). The GMM was reactive to those changes, resulting
in abrupt market condition switches at times.

Where We Are Today

Now let’s use this model to understand market behavior
over the past ~1.5 years. Exhibit 5 is an updated, “zoomed- To explore this result further, in Exhibit 7 we display the
in” version of Exhibit 4, displaying the highest probability factor returns since the beginning of 2020 and find that the
market condition in each period since the beginning of 2020. factors that outperform the most in Inflation (Local Inflation
We provide additional detail on each market condition’s and Foreign Currency) were negative.
probability over the same period in Exhibit 6.
We should also comment that the current period might be
We find that Crisis was expectedly present in the COVID inflationary, but different from what we saw in the 1970s
market crisis in February and March 2020, followed and 1980s (where Inflation showed up the most in our
immediately by WOI. In the second half of 2020, Steady training period). Additionally, this model is not predictive, so
State was the predominant market condition. Then, in 2021, we are not able to say whether Inflation will be increasingly
we observed a shift toward WOI, which as we mentioned important later in 2021 or the years ahead. However,
earlier, might be a proxy for market fragility. Crisis entered stay tuned for future Street Views where we may provide
the picture sporadically since then, most notably in early to updates on the model’s output.
mid-March 2021.
Exhibit 7: Factor Returns Since the Beginning of 2020
Finally, Inflation’s probability has been zero over this period Period: January 1, 2020 - July 9, 2021.

(the maximum daily probability was only 0.01%), which is Annualized Return
a particularly interesting result given the market’s fears of Equity 19.11%
Trend Following 11.92%
higher inflation and the high CPI prints in May and June Momentum 7.58%
2021. Local Equity 4.43%
Quality 2.86%
Exhibit 5: Highest Probability Market Conditions Since the Interest Rates 2.44%
Small Cap 0.20%
Beginning of 2020
Local Inflation -0.93%
Period: January 1, 2020 - July 9, 2021.
Emerging Markets -0.99%
Commodities -1.04%
Foreign Exchange Carry -4.34%
Foreign Currency -5.37%
Equity Short Volatility -5.41%
Fixed Income Carry -6.73%
Credit -11.33%
Value -17.30%
Low Risk -19.28%

Source: Two Sigma Factor Lens as of August 16, 2021, using daily data.

6
© 2021 Two Sigma. All Rights Reserved. “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP.
Please see the last page of this document, which contains important disclaimer and disclosure information.
Conclusion: Applying This Analysis to
Investment Decisions
One way to approach modeling regimes is to determine
them based on experience and knowledge of the markets.
An alternative approach (and one that Two Sigma generally
takes when solving problems) is more data-driven in nature.
The unsupervised learning method presented in this Street
View can add value by letting a large amount of historical
data determine the regimes for you. The output of this model
applied on the factors in the Two Sigma Factor Lens was four
clusters, or market conditions. We then labeled those market
conditions, based on the properties of each, as follows:
Crisis, Steady State, Inflation, and Walking on Ice (WOI).
We analyzed their behavior throughout history to identify
regimes, or periods where market conditions showed some
persistence.

We believe there are multiple use cases for allocators

looking to apply this type of analysis. One use case is
risk management. Allocators can enhance their scenario
analysis by sampling from the distributions of these market
conditions to stress test their portfolios.

Another use case is assistance with asset allocation

decisions. Predicting market returns over long periods is
difficult even under very stable market conditions. The
evidence of rather quickly-changing market conditions
presented in this Street View makes adherence to those
long-term forecasts even more challenging. The implications
for asset allocation might be that allocators should design
portfolios that can withstand market condition volatility over
the long-term, while potentially seeking opportunities for
tactical shifts on shorter horizons.

7
© 2021 Two Sigma. All Rights Reserved. “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP.
Please see the last page of this document, which contains important disclaimer and disclosure information.
Appendix 1: Two Sigma Factor Lens Start Dates for the GMM
Factor Start Date
Equity 1/1/71
Interest Rates 1/1/71
Credit 1/1/71
Commodities 1/5/72
Emerging Markets 12/22/72
Foreign Currency 1/5/71
Local Inflation 1/1/71
Local Equity 1/4/71
Equity Short Volatility 7/1/86
Fixed Income Carry 5/9/91
Foreign Exchange Carry 1/6/72
Trend Following 3/26/92
Low Risk 1/4/71
Momentum 1/4/71
Quality 1/4/71
Value 1/4/71
Small Cap 1/4/71

Appendix 2: Market Conditions Preceding and Following Market Condition 4 (WOI)

Regime # Before WOI # After WOI % Before WOI % After WOI
Steady State 139 137 6.9% 6.8%
WOI 1782 1782 89.1% 89.1%
Crisis 76 80 3.8% 4.0%
Inflation 4 2 0.2% 0.1%

8
© 2021 Two Sigma. All Rights Reserved. “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP.
Please see the last page of this document, which contains important disclaimer and disclosure information.
Important Disclosure Information
This report is prepared and circulated for informational and educational purposes only and is not an offer to sell or the solicitation of an offer to buy any
securities or other instruments. The information contained herein is not intended to provide, and should not be relied upon for investment, accounting, legal,
tax or other advice. This report does not purport to advise the recipient personally concerning the nature, potential, value or suitability of any particular
sector, geographic region, security, portfolio of securities, transaction, investment strategy or other matter. No consideration has been given to the specific
investment needs or risk-tolerances of any recipient. The recipient is reminded that an investment in any security is subject to a number of risks including the
risk of a total loss of capital, and that the discussion herein does not contain a list or description of relevant risk factors. The recipient hereof should make an
independent investigation of the information described herein, including by consulting its own tax, legal, accounting and other advisors. This report does not
constitute any form of invitation or inducement by Two Sigma Investments, LP or its affiliates (“Two Sigma”) to engage in investment activity. Past performance
is no guarantee of future results.

The views expressed herein are not necessarily the views of Two Sigma. Such views (i) may be historic or forward-looking in nature, (ii) reflect significant
assumptions and subjective judgments of the author(s) of this report, and (iii) are subject to change without notice. The use of pronouns such as “we” and
“our” are references to the authors of this report only (and/or methodologies or assumptions utilized by the authors in this report only). Two Sigma may have
market views or opinions that materially differ from those discussed, and may have a significant financial interest in (or against) one or more of such positions
or theses.

By their receipt, the recipient of this report acknowledges the disclaimers contained within this report and, in particular, acknowledges that this report and
the data contained within this report are provided on an “as is” basis. Two Sigma, its employees, information providers (including, without limitation Barra,
LLC), and any other third party involved in or related to the making or compiling of this report or the data contained within this report make no representation
or warrant of any kind, either express or implied, with respect to this report or the data (or the results to be obtained by the use thereof). Two Sigma, its
employees, information providers (including, without limitation Barra, LLC) and any other third party involved in or related to the making or compiling of this
report or the data, expressly disclaim any and all implied warranties of originality, accuracy, completeness, non-infringement, merchantability and fitness for
a particular purpose.

As a recipient of this report, you assume the entire risk of any use you may make of this report or of the data. In no event shall Two Sigma, its employees,
information providers (including, without limitation Barra, LLC) or any third party involved in or related to the making or compiling this report or the data, be
liable to the recipient or any other third party, for any direct or indirect damages, including, without limitation, any lost profits, lost savings or other incidental or
consequential damages, arising out of your use or your inability to use the report or the data, regardless of the form of action, even if Two Sigma, its employees,
any of their information providers (including, without limitation Barra, LLC), or any other third party involved in or related to the making or compiling of this
report or the data has been advised of, or otherwise might have anticipated the possibility of such damages.

This report may include certain statements and projections regarding the anticipated future performance of various securities, sectors, or geographic regions.
These forward-looking statements are inherently subject to significant business, economic and competitive uncertainties and contingencies, many of which
are beyond Two Sigma’s control. Factors which could cause actual results to differ materially from those anticipated include, but are not limited to: competitive
and general business, economic, market and political conditions from those expected; changes in the legal, regulatory and legislative environments in global
markets; and the ability of management to effectively implement certain strategies. Words like “believe,” “expect,” “anticipate,” “promise,” “plan,” and other
expressions or words of similar meanings, as well as future or conditional verbs such as “will,” “would,” “should,” “could,” or “may” are generally intended to
identify forward-looking statements.

References in this report to third-party content or websites do not imply any endorsement by the third party of this report; nor do they imply any endorsement
by this firm of the third-party content or website. As noted, Barra, LLC’s analytics and data (www.barra.com) were used in preparation of this information.
Copyright 2021 Barra, LLC. All Rights Reserved. In addition, some of the images and other material used herein may otherwise be protected by copyright
and/or trademark. If so, such copyrights and/or trademarks may be owned by the entity that created the material and, where relevant, are used purely for
identification and comment as fair use under international copyright and/or trademark laws. Use of such an image, copyright or trademark does not imply any
association with such organization (or endorsement of such organization) by Two Sigma, nor vice versa.

© 2021 Two Sigma | All Rights Reserved | “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP.

Important Notice to Venn Subscribers: This document is not an offer to, or solicitation of, any potential clients for Venn or otherwise for the provision of
investment management, advisory or any other services. Nothing in this document should be considered a representation of how the Two Sigma Factor
Lens may be used on Venn or about Venn or the Two Sigma Factor Lens in any respect. Importantly, any use by Venn of the Two Sigma Factor Lens may differ
materially from any content, research or methodologies discussed herein. This document is subject to change without notice.

9
© 2021 Two Sigma. All Rights Reserved. “Two Sigma” and the “2σ” logo are registered trademarks of Two Sigma Investments, LP.
Please see the last page of this document, which contains important disclaimer and disclosure information.