Essay About Proprietary Models And Qsg Stock Selection Models

Measuring the Winds of August
Essay Preview: Measuring the Winds of August
Report this essay
Measuring The Winds Of August
Introduction
A lot has been writt en in the fi nancial media in the last few weeks about the struggles of quantitative strategies, particularly related
to their performance in the fi rst few days of August. We are not going to retell the stories, but we do want to put this months
performance into context. We also want to off er additional perspectives on what happened and any insights we might leverage to
avoid a similar disruption in the future.
The emerging consensus among quants is that the turbulence in the equity market was exported from the credit markets. The
liquidity demands of some multi-strategy hedge funds, funds who felt the pinch of margin calls on their credit portfolios, forced
them to urgently unwind many of their highly leveraged equity positions. The rapid unwinding, especially the short covering, was
on a leverage-driven scale that tested the equity markets infrastructure. This de-leveraging and the associated contagion eff ects
exacerbated the relative price behavior. This rush for the exits is certainly a concern, but a bigger concern is why did the biggest
quant hedge funds, despite their secrecy and oft en well-developed proprietary models, all experience the same fate? If the
fi rms were pursuing independent strategies with low historical correlations, is there a fundamental explanation for their shared
experience?
What We Saw
From Aug 1-9, virtually all QSG stock selection models (except the REIT model) had signifi cant negative long-short spreads (using
ranks as at Jul 31 within the Russell 1000) ranging from -4.2% to -12.7%. This was unprecedented for such a short period, especially
given that the market barely moved (Russell 1000 was down just 0.22%). The table below shows the stats for these models. Notice
that only 19% of the stocks in the top decile for the Value Analyst model outperformed the Russell 1000 during this period. In
contrast, 80% of the names in the bott om decile outperformed. A similar trend can be seen across all models, apart from the REIT
model, which had a healthy 2% long-short spread.
From Aug 10-31, we saw a complete reversal of the performance of the fi rst few days. All models (except the Bank & Thrift model
which was marginally negative) had signifi cantly positive long-short spreads. It is interesting to point out that the REIT model once
again had positive spreads during this period.
Measuring The Winds Of August Page 1
D1 D10 Spread Benchmark (BM) % D1 > BM % D10 > BM
Earnings Momentum (EMM) -5.51 2.82 -8.33 -0.22 25 69
Relative Value (RVM) -4.38 0.76 -5.14 -0.22 27 65
GARP (QGP) -5.11 0.57 -5.68 -0.22 26 66
Growth Analyst (QGA) -3.99 0.22 -4.21 -0.22 38 60
Momentum Analyst (QMA) -5.25 0.69 -5.94 -0.22 27 64
Value Analyst (QVA) -6.15 4.33 -10.48 -0.22 19 80
Retail -7.01 5.66 -12.67 -1.49 25 81
Banks & Thrift s -0.91 5.67 -6.58 2.90 28 60
REIT 3.60 1.54 2.06 1.90 67 38
Insurance -5.31 -0.39 -4.92 -3.05 30 68
Table 1: Selected Model Performance, 8.1.2007-8.9.2007
When Correlation Breaks Down
Focusing on the non-industry specifi c models for the Aug 1-9 period, perhaps the most surprising trend was that models that
stylistically should be good diversifi ers, given their low rank correlations, moved in tandem. For instance, the Earnings Momentum
model (EMM) which measures analysts expectations, revisions and earnings surprise among other earnings-related factors should
naturally be a good diversifi er when linearly combined with the Value Analyst model (QVA) which looks for cheap stocks poised to
make a run. Instead both were down -8.33% and -10.48% long-short respectively. The correlation matrix below shows that these 2
models had a low rank correlation of 0.26 on Jul 31. In fact, with the exception of the Relative Value (RVM) and GARP (QGP) models,
the rank correlation of all models was below 0.60. So, why did the models perform similarly despite the low rank correlations?
To answer that question, we need to remember that the concept of correlation, despite its appeal and elegance, is an oft en abused
statistical measure. Correlations tend to break down in times of high volatility and seemingly uncorrelated strategies tend to move
together (a concept called contagion). This obviously was the case during those challenging days in early August. In addition, in the
long-short world, where tail returns are the only thing that matt ers, a cross-sectional correlation measure across an entire universe
can be misleading.
Tail Overlap
A bett er way to gauge how quant strategies are similar, especially during high volatility periods, is a simple concept we have termed
“tail overlap”. This computes, for any 2 models, the percentage of stocks that overlap in each of the tails (i.e. Deciles 1 and 10). So
as shown in the matrix below, 32% of the stocks in EMMs top decile also belonged to RVMs top decile on Jul 31. Similarly, 16% of
QGAs decile 1 stocks were in QVAs decile 1.