You are on page 1of 9

The Quant Scare

Cooler Heads

Headlines variously referred to it as a quant crisis, a panic and a meltdown. The implication was clear: The world of
quantitative management had been turned on its head and no one seemed to know quite why, or what to do about it.

While recent events obviously took quant managers by surprise, cooler heads prevailed. Quant managers did
what they do best: studied, analyzed, learned and adapted.

So...

> What happened?


> What can we infer from the data?
> Can we immunize for “hidden” risks?

And most importantly...

> What lessons have we learned?

These are some of the questions that Axioma and its clients have wrestled with in recent weeks. The dust has
yet to settle and the answers yet to be fully scoped, but here are our perspectives on what we prefer to call
“The Quant Scare.”

Sebastián Ceria
Chief Executive Officer
Axioma, Inc.
September 4, 2007

1
What happened?
Daily risk data show dramatic shifts in the magnitude, direction and correlations of factor returns

In the long run, fundamentals drive returns. So, when the sectional volatility. Chart 2 shows the daily factor returns for
market behaves in an unexpected fashion, it is essential the following four fundamental factors since the beginning
to understand whether fundamental factors are driving of July: Volatility, Short-Term Momentum, Leverage, and
that behavior. Recent events were characterized by a Medium-Term Momentum. The increase in factor cross-
trifecta of unprecedented changes in the magnitude, sectional volatility, in turn, affected the broad market’s
direction and correlations of factor returns, driving cross-sectional volatility.
(negative) returns in the equity markets and increased
risk for quant managers.

The magnitude of the shifts was startling. Chart 1 shows


the standardized daily factor returns for the style factors
in the Axioma US Robust RiskTM Model through mid-
August. Factor returns exceeding +2 or -2 should only
occur in about 5% of all observations. Yet almost one-third
of the returns during this period exceeded those bounds.
In contrast, only 2.27% of daily observations were found
outside these bounds for the period of January 3, 2005 to
Chart 2: Axioma US Factor Returns for the Period
July 31, 2007, and none of those observations outside the
July 2 – August 15, 2007
range exceeded the 5% absolute bound.
Recent factor behavior also was characterized by profound
changes in direction and correlations. During July and
August, we observed unprecedented changes in the
correlations between key factor returns. Table 1 compares
the changes—some quite dramatic—in factor correlations
between the periods of January 1, 2007 to June 5, 2007

Value Growth Market Short-Term Medium-Term Volatility


Sensitivity Momentum Momentum
Value - 0.2 -0.55 -0.24 0.08 -0.28
Growth 0.2 - -0.44 -0.2 0.35 0.003
Market -0.55 -0.44 - 0.39 -0.56 -0.02
Sensitivity
Short-Term -0.24 -0.2 0.39 - -0.47 0.18
Momentum
Chart 1: How Unusual Have August Daily Returns Been?
Medium-Term 0.08 0.35 -0.56 -0.47 - -0.52
Momentum
Four of the nine Axioma US Robust Risk Model style Volatility -0.28 0.003 -0.02 0.18 -0.52 -

factors1 experienced particularly large changes in cross- Table 1: Absolute Changes in Factor Correlations

2
and June 6, 2007 to August 15, 2007. The largest impact (Chart 3) demonstrates the recent rise in expected
was on Market Sensitivity and the two momentum volatility for the S&P 500. However, the chart fails to
factors. In effect, relationships normally taken for speak to the main drivers behind this expected increase
granted in quant models were turned on their heads. in future volatility. While previous rises in volatility during
Hence, corresponding approaches to risk management the LTCM and Tech Bubble (1998 and 2000, respectively)
based on the previous stability of these correlations were led by a sharp increase in asset-specific cross-sectional
became ineffective. volatility, the current volatility regime seems driven by an
increasing cross-sectional volatility of common factor
Market Sensitivity typically has a positive correlation with returns; four of them in particular. The implication for risk
Value, Growth, and Medium-Term Momentum and is management is that a risk-control strategy based on
uncorrelated with Short-Term Momentum2. In the recent asset diversification (i.e., increasing the number of assets
period this pattern was reversed, with a positive correlation held in the portfolio) would have been less effective in
between Short-Term Momentum and Market Sensitivity reducing risk in the current environment than a concerted
and no significant correlation of Market Sensitivity to effort to control the exposures to these common factors.
Value, Growth, and Medium-Term Momentum. In Conversely, alpha strategies based on factor tilts have
addition, Short- and Medium-Term Momentum reversed more opportunities to outperform (or underperform) their
their usually slightly positive correlation with each other benchmark than ones based on stock picking.
and Medium-Term lost its typically positive correlation
with Volatility, becoming essentially uncorrelated.
Table 2 details the correlations for Market Sensitivity in
the two periods.

Market-Sensitivity Value Growth Short-Term Medium-Term


Correlations Momentum Momentum
2007-01-01 to 0.36 0.33 -0.02 0.39
2007-06-05
2007-06-06 to -0.18 -0.10 0.37 -0.16
2007-08-15

Table 2: Comparison of Market-Sensitivity Correlations

The magnitude of the change in correlations from a


Chart 3: Recent Increase in Expected Market Volatility
longer-term average in the recent 50-day period starting
June 6 was approximately three times greater than the The following study illustrates this point. We conducted a
average change over other 50-day periods. daily performance attribution on three variants of the
following strategy from July 23rd to August 7th. In Test 1,
In addition to the substantial shifts in magnitude and we constructed a Value-tilt portfolio by constraining every
direction, the composition of risk also changed. The VIX common factor (i.e., Non-value Style Bounds = “0”) other

3
than the Value factor (Table 3). In this variant we under- Daily Attribution from the Close on July 23 to the Close on August 7
(11 trading days)
performed the benchmark S&P 500 due to the negative
Non-Value Active Factor Specific
performance of the Value factor during this period, but Style Bounds Return Return Return
Test 1 0.00 -0.92% -0.22% -0.71%
the underperformance was limited to just -0.92%. In Test Test 2 0.40 -2.33% -1.14% -1.19%
2, we relaxed our control on exposures to the other fun- Test 3 0.80 -2.74% -1.65% -1.08%

damental factors, allowing for an active exposure of up Table 3: Performance Attribution Test Results
to + or - 0.40. In this variant, we underperformed our
benchmark by -2.33%, taking on additional bets on This attribution shows that if factor exposures remain
Size (-0.34) and Short-Term Momentum (-0.26) in par- constant and are not changed to respond to changing
ticular. In Test 3, we relaxed our control of other factors market conditions, the risk and subsequent returns of
even further to + or - 0.80. This gave us even larger the portfolio will rise (and fall) along with factor cross-
(negative) exposures in Size (-0.63) and Short-Term sectional volatility.
Momentum (-0.29) and an even greater underperfor-
mance of -2.74%. In this last variant, the contribution
to (negative) active returns from common factors sur-
passed that of asset-specific returns.

4
What can we infer from the data?
Was it a liquidity crisis caused by the sub-prime debacle? Or did other, perhaps
“hidden,” factors contribute to the turmoil?

There is general agreement in the quant community the effect of mortgage defaults on interest rates).
that the recent scare was caused by a liquidity crunch While some analysts attributed the fall off to Value or
stemming from the sub-prime lending debacle. Liquidity, those particular factors—while clearly under-
Matthew S. Rothman of Lehman Brothers was among performing—were not the principal drivers, according
those who first articulated the most plausible explanation to Axioma’s factor returns.
linking these two events, as follows: Many multi-strategy
hedge funds running quantitative equity strategies
were also investing in highly leveraged credit
instruments. When the value of these credit instruments
began to fall due to the sub-prime crisis, the funds
were forced to raise cash to cover margin calls. Since
the market values of these illiquid instruments were
hard to gauge, the multi-strategy funds moved to
generate cash by selling equities instead. This
liquidation triggered a steep decline in the equity
assets held by quant managers as a whole. The Chart 4: Cumulative Returns of Axioma Factors:
decline increased the overall volatility in the market, July 2, 2007 - August 15, 2007

thus forcing even those managers not invested in the Cumulative returns are shown in Chart 4. To verify the
credit instruments to begin reducing leverage in an assumption that portfolios with large exposures to
attempt to manage risk. Due to similarities in the Volatility and Leverage performed poorly over the period
positions that quant managers held, the effects of the in question, we constructed a portfolio with a strong
reduction in leverage spilled well beyond the world of volatility tilt relative to the S&P 500. The daily performance
market-neutral hedge fund strategies. attribution results on this portfolio from July 2nd to August
15th were as expected: the principal factors contributing
If we assume that the quant scare did, in fact, originate to underperformance were Volatility and Leverage, as
in the sub-prime market, then portfolios with a leverage shown in Table 4.
tilt should have been hit the hardest. This is precisely
what Axioma’s risk model shows. The principal style While Leverage certainly appears to have been an
factors responsible for the market’s underperformance important component of underperformance, we
in recent weeks were Volatility and Leverage. This strongly believe that no single a priori-defined risk
concurs with anecdotal evidence (i.e., the market has factor could have anticipated the absolute and relative
been extremely volatile, with much being written about lack of performance.

5
Account Summary volume would have driven asset prices significantly
Return Risk IR Confidence
against them. Suppose, for example, that a hedge fund
Managed -7.84% 5.04% n/a n/a
Benchmark -5.04% 4.29% n/a n/a has $10 billion in AUM in a dollar-neutral equity fund.
Active -2.80% 1.06% -6.65 0.80%
The liquidity measures put in place during the rebalancing
Factor Contribution -2.20%
+ Specific Return Contribution -0.61% (i.e., market-impact constraints) would have been
Active Return -2.80% sufficient only for that fund. Assume that other hedge
Contributors to Active Return by Style
Contribution Avg Wtd Exp HR IR
funds invest in similar strategies so that a combined total
Short-Term Momentum 0.27% -5.65% 42.42% 6.80 of $400 billion in AUM is now invested in similar quant
Size 0.03% 7.35% 48.48% 1.76
Liquidity 0.03% 7.36% 48.48% 1.37
strategies. Under de-leveraging or cash generation, the
Value 0.01% 0.62% 60.61% 2.36 cost of trading becomes far greater than expected. In a
Growth -0.01% 0.48% 45.45% -4.59
Medium-Term Momentum -0.02% 0.33% 45.45% -3.86 typical rebalancing exercise of the $10 billion quant fund
Market Sensitivity -0.18% 10.06% 42.42% -4.04 where we assume that the total dollars traded both long
Volatility -1.14% 40.80% 48.48% -5.88
Leverage -1.24% 55.28% 33.33% -10.13 and short is 20% of AUM, the expected execution shortfall
(or market impact3) as predicted by the Goldman Sachs
Table 4: Factors Contributing to Underperformance
Execution Shortfall Model (GS-ESM) would have been
Other sources of risk were clearly afoot—let’s call them 0.33%4 of AUM. However, if we assume that $100 billion
“hidden” factors—and demand for liquidity was foremost in AUM were trading the same 20% of AUM that day,
among them. then the predicted execution shortfall would have been
1.04% of AUM. If $400 billion in AUM had the same 20%
One of the features of the event was a steep increase in turnover, the predicted shortfall climbs to 1.69% of AUM.
trade volumes, resulting in huge demand for liquidity. By This means that the value of the fund would have been
Axioma’s calculations, the average increase in average expected to drop by 1.69% from trading similar assets to
daily volume (ADV) among assets in the Russell 1000 those of other quants; 1.36% more than if it had traded
was roughly 30% during the week starting August 6 . th
by itself.
The average increase in ADV among assets in the
Russell 2000 was more than twice that value. Trading Based on the situation over the last few weeks, we
desks reported record numbers of tickets. Assuming that believe that there is also a nonlinearly increasing risk
these trades were initiated by quants, this additional associated with holding more illiquid assets.

6
Can we immunize for “hidden” risks?
New techniques can help

Recognizing the inherent weakness of risk models to strategies but with the inclusion of the Alpha Factor into the
deal with risk arising from factors not included in the risk constraint with a volatility parameter of 30%5.
models, Axioma recently introduced an approach to help
deal with hidden factors. The technique identifies one (or In viewing the results in Table 5, a casual observer might
more) "missing" factors orthogonal to the original factors conclude that the Alpha Factor merely imposes more
in the risk model. The presence of the additional portfolio aggressive risk controls, resulting in conservative portfolios
factor recovers a component of risk that is otherwise that might even hinder performance during regimes of sta-
unaccounted for. This adjustment efficiently mitigates the ble volatility. This is not the case, as illustrated by the figures
underestimation of risk. Recent events provided a unique in Table 6 for the above portfolios as of June 29, 2007, prior
opportunity to further evaluate the effectiveness of the to the market shakeup. Enabling the Alpha Factor risk con-
Axioma Alpha FactorTM methodology. trols yielded superior results—greater return with less risk.

Return Predicted Risk Realized Risk Avg Turnover Sharpe Ratio


Long Medium-Term Momentum, long Value
No AF -19.32% 6.00% 9.58% 42% -2.02
AF -11.82% 6.00% 6.63% 11% -1.78

Long Medium-Term Momentum, short Short-Term Momentum


No AF -3.63% 6.00% 11.15% 75% -0.33
AF -0.39% 6.00% 9.21% 46% -0.04

Table 5: For the Period January 3, 2007 – August 15, 2007

It may be premature to assert whether events in the past


month and a half constitute a change in volatility regime.
To suggest that all traditional assumptions have broken
down, rendering the current breed of risk (and alpha) models
Chart 5: Portfolio Performance, with and without Alpha Factor
useless, is equally unhelpful and likely to be a vast over-
To illustrate the effectiveness of the Alpha Factor, we consid- statement. Irrespective of market conditions, deficiencies will
ered various long-short, dollar-neutral strategies with strong always exist in any factor risk model. This is precisely what
factor biases toward those risk factors that have been the Alpha Factor methodology recognizes, making it an
behaving perversely during this crisis. Using the Russell invaluable addition to the portfolio construction process.
1000 as the universe, optimized portfolios of 100-150 assets
Return Predicted Risk Realized Risk
were generated at the start of 2007 (Chart 5). Leveraging Long Medium-Term Momentum, long Value
No AF -6.93% 6.00% 8.29%
Axioma’s daily risk models, the portfolios were allowed to
AF -4.49% 6.00% 5.19%
rebalance daily. Each rebalancing imposed a total risk limit of Long Medium-Term Momentum, short Short-Term Momentum

6%, minimum and maximum asset holdings thresholds of No AF -3.32% 6.00% 5.99%
AF 1.99% 6.00% 5.16%
0.5% and 10%, respectively, and a maximum period
Table 6: For the Period January 3, 2007 – June 29, 2007
turnover of 75%. We then compared results for the same

7
What lessons have we learned?
“In the view of several big-time quants I spoke to, their big mistake was in not realizing that their little corner of Wall
Street had become so crowded with imitators—and that when others were forced to sell, they were going to get hurt.
Now they are all trying to figure out how to factor that into their thinking for the future—Mr. (Clifford) Asness (co-founder
of AQR Capital Management) included. ‘We have a new risk factor in our world,’ he said.”
“Markets Quake, and a ‘Neutral’ Strategy Slips”
By Joe Nocera - The New York Times
August 18, 2007

The events of the last several weeks have been sobering. quants. But this explanation is, at least in part, a self-
We have been reminded—and powerfully so—that fulfilling prophecy. Liquidity exposures also explained
successful investment ideas are invariably imitated. Yet less than expected. This largely has to do with the way
truly successful portfolio managers are defined not by their in which the exposures are computed in a fundamental
capacity for mimicry, but by their ability to differentiate factor risk model. The style factors only explain exposure
themselves successfully from their peers. At Axioma, our to the style relative to the market. But if the market
responsibility is to provide our quant clients with innovative behavior is being driven largely by liquidity, the factor
tools that will help them achieve this differentiation. attribution would not reflect this phenomenon.
Axioma is investigating alternatives to computing
But it is not only about differentiation. Recent events style exposures, along with other modifications to its
have highlighted the shortcomings in current risk robust risk model, to help its clients adapt to constantly
management techniques to control hidden—and quite evolving market conditions.
asymmetric—risks. This issue demands a solution.
Fortunately, we already see opportunities for improving Market impact must also be given greater consideration
the tools that provide portfolio managers with early and in the portfolio construction process. In the future, portfolio
timely warnings, and help them better understand the managers must be better prepared to manage liquidity
sources of performance. Managers also need to use risk. In particular, quant managers must assume that others
portfolio optimizers to construct hedges that will allow are investing in similar strategies. During liquidity crunches,
them to react quickly to minimize losses, when they must assume that their trades will have much larger
confronting the next scare. impact on prices, as we have just witnessed.

Our ability to obtain better insights will lean heavily on Regarding early warnings, we strongly believe that
the development of enhanced performance attribution current risk management practices can be further
methodologies. Perhaps risks could be perceived earlier improved. At Axioma, we will continue to deliver
if returns were attributed differently when using factor- more timely, transparent and robust risk models, as
based attribution. During the scare, part of the underper- well as new tools for incorporating unforeseen risks.
formance suffered by quant strategies was attributed to The Axioma Alpha Factor methodology is the first
the Value factor, because Value is a popular signal with among these.

8
Of course, while immunizations and early warnings are immunization against asymmetric risks, fat tails and
extremely valuable, our clients also need tools that enable liquidity risks. Discussions highlighted the need for tools
them to react constructively in the presence of unforeseen that do a better job of providing early warnings. Talk
circumstances, such as the scare we just experienced. By focused on the need to enhance risk management and
incorporating realistic market-impact measures in portfolio portfolio construction by incorporating nonlinear assets,
construction, our clients will be able to make better deci- such as options, and realistic market-impact measures in
sions about which stocks to buy or sell in a liquidity crisis. the portfolio construction process. We’ve heard the wake
Additionally, clients must be able to build, sometimes in real up call: efforts to improve these tools must be redoubled.
time, better hedges that protect their portfolios from certain
exposures. Hedging is important because it allows portfolio Fortunately, we are not starting from scratch. Axioma is
managers to control performance—and risks—without already collaborating with its clients on the development
necessarily being forced to liquidate positions. of new tools that address these issues.

In numerous conversations with our clients after the Success in meeting these challenges will inevitably be
Amaranth meltdown, a number of these very same measured as it always has: by our ability to work
issues were flagged as potential research topics. collaboratively with our clients to generate new ideas
Concerns were expressed regarding the need for and to create innovative tools that drive their success.

1. The Axioma US Robust Risk Model has nine style factors. Volatility gives a measure of an asset’s relative volatility over time according to its
historical behavior. It is calculated as the square-root of the asset’s absolute return averaged over the last 60 days, divided by the cross-sectional volatility of
the market. Short-Term Momentum gives a measure of a stock’s recent performance. It is defined as an asset’s cumulative return over the last 20 trading
days. Leverage provides a measure of a company’s exposure to debt levels. The Leverage factor is calculated as total debt divided by market capitalization.
Total debt is the sum of long-term debt and debt in current liabilities (short-term debt) taken on the most recent date over the last 520 trading days for which
both values have been filed. For market capitalization, the 20-day average is used. Medium-Term Momentum gives a measure of a stock’s past performance
over the medium-term. It is defined as an asset’s cumulative return over the last 300 trading days, excluding the last 20 trading days. Size allows us to differ-
entiate between large and small stocks and is defined as the natural logarithm of the market capitalization, averaged over the last 20 trading days. Market
Sensitivity is a measure of a stock’s under or over performance relative to the broad market from historical data. It is calculated by regressing the time-series
of an asset’s return against the market return. We define the market as the portfolio of the top 3,000 assets by market capitalization. Liquidity is a turnover
factor that provides a measure of a stock’s liquidity, or lack thereof. It is defined as the natural logarithm of the last 20-day average volume divided by the
natural logarithm of the last 20-day average market capitalization. Value, or Book to Price, gives a measure of a stock’s value within the market; how fairly is
it priced. The book-to-price factor is calculated as the ratio of common equity to the current market capitalization. Growth gives an indication of a company’s
rate of growth historically. The Growth factor is calculated as the product of one minus the dividend pay-out rate and the one-year return on equity.
2. Note that all correlations among factors are with reference to the factors in Axioma’s US Fundamental Risk Model. These correlations are
computed using daily factor returns that are the results of robust cross-sectional regressions.
3. The most advanced quant tools model market impact as a nonlinear increasing function of the amount traded. Contemporary approaches model
market impact as a function of the dollar trade amount, where the average cost per share is assumed to be a multiple of the square root of the
trade size in currency. When multiplied by the shares traded, the resulting market-impact function is the sum over the cost of the assets, where the
cost per asset is a multiple of the currency value traded raised to the 3/2 power.
4. The values given are percentages of the value of the portfolio, not the dollar amount traded. The trades were generated by turning over about 20% of a portfo-
lio. The portfolio was generated from a backtest over the Russell 2000 universe with a value tilt and realistic constraints on asset and factor bounds and tracking
error. These numbers were generated for illustrative purposes to show the nonlinear increases in cost due to increased trading. The trades made in the $100B
and $400B portfolios are multiples of ADV and are therefore outside of the normal range of trading for which execution shortfall models are calibrated.
5. 30% corresponds roughly to the average level of specific risk in the Axioma US estimation universe during the first half of 2007.

Axioma Robust, Axioma Alpha Factor and Axioma Portfolio are trademarks of Axioma, Inc.

"This analysis is provided for informational purposes only as a service by Axioma to clients and other interested parties. It does not constitute and
should not be construed as investment advice. In order to obtain answers to specific questions dealing with the subject matter and relating to a
client's or a potential client's particular circumstances, please contact Axioma directly."

You might also like