When Everyone Runs the Same Three Ideas

Quantitative investing spent most of its history as a lonely pursuit. Up until the early 2000s, a small group of managers sold systematic equity strategies to a small group of clients who already believed in them. Then the results kept coming, across many regions and many market environments, while other approaches disappointed.

Money followed. New managers entered, and the audience broadened well past the true believers. Nearly all of them used the same handful of ideas: value, momentum and earnings quality (the accruals signal that flags companies whose accounting props up their profits).

Large amounts of capital chasing identical signals produce a predictable result. Returns shrink, and some managers respond by adding leverage, which puts even more capital behind the same trades. We would call that a quant bubble, and the subprime mortgage crisis was the pin.

August 2007 brought a mass exodus. A slower, steady bleed of assets followed over the next several years, along with volatility and cross-manager correlations that looked nothing like what the industry had seen through 2006.

This matters again in September 2012 because the hangover has not cleared. Clients still ask whether systematic equity can earn its fee, and managers still debate whether the old signals have anything left. Two questions sit underneath that debate: where will alpha come from now, and how should we manage risk when the crowd is the biggest risk of all?

We expected risk management to be the part quants had mastered. It was the cornerstone of the whole approach, so the 2007 episode was uncomfortable. The models measured what we built them to measure, and the thing that hurt was something we never asked them about, namely how many other people held the same positions.

Why Correlation Alone Misleads Stock Pickers

Conventional wisdom says active management gets harder when stocks move together. If everything rises and falls in lockstep, how can anyone separate one company from another? It sounds right, and it is incomplete.

A better gauge of opportunity is volatility, split into its parts. Residual (specific) volatility measures how widely individual stock returns scatter after the common factors are removed, and a wide scatter means a big prize for good stock selection. Factor volatility measures how widely the factor returns themselves swing, which creates a separate prize for anyone who can pick factors.

The link between the two ideas comes from the simplest model in finance, the Sharpe market model. Assume every stock has a beta of one and the same residual risk, call it ω. Then the correlation between any two stocks is the market variance divided by the sum of market variance and residual variance.

That formula says high correlation can arise three ways: high market (or factor) volatility, low residual volatility, or some blend. Those three situations feel the same on a correlation chart and very different to a stock picker.

Take a quick example with round numbers. If market volatility is 15% and residual volatility is 15%, the implied correlation is 0.0225 divided by 0.045, or 50%. Now double the residual volatility to 30% while market volatility stays put: correlation falls to 0.0225 over 0.1125, which is 20%. Stock picking got easier, since the dispersion of outcomes is far wider, even though correlations fell as well. Move the other way, with residual volatility at 8% and market volatility at 15%, and correlation jumps to 78%. Opportunity is thin there.

So a rising correlation reading is a prompt to ask which ingredient moved. Reading it as a judgment on skill gets the analysis backward.

Reading the 1992 to 2010 Numbers

The talk builds a risk model, simple by design, to answer that question. Each month it takes daily returns for the 250 largest US stocks and fits two factors: beta, estimated from within-month regressions, and size, measured as log market cap. The data run from 1927 through 2010, which lets us see average stock volatility, factor volatility and specific volatility on one chart.

The long-run history shows stock-by-stock correlation drifting upward since the early 1990s, with sharp spikes in crisis periods. A separate chart of developed Europe tells a similar story from 1999 to 2010, with equal-weighted average daily correlations ranging from the low teens of percent up toward 40%.

Now the numbers that anchor the argument. For 1992 through 2010, factor volatility averaged 9.9% and specific volatility averaged 12.4%. Measured correlation was 26%, and the correlation implied by the market model was 39%.

For August 2007 through 2010 the figures shift. Factor volatility rose to 13.6%, specific volatility barely moved at 12.8%, measured correlation climbed to 39%, and implied correlation reached 53%. We checked the implied figures with the formula above: 0.0185 divided by 0.0185 plus 0.0164 gives about 53%, and the earlier period works out to 39%. The arithmetic holds.

What stands out is that residual volatility stayed almost flat. Correlations rose because factor volatility rose. Stocks did not become less distinguishable from each other in terms of their own idiosyncratic risk. The common factors got louder, and the gap between measured and implied correlation (26% against 39%, then 39% against 53%) is a reminder that a one-factor simplification overstates how tightly real stocks move together.

This suggests the stock-picking opportunity survived the crisis. What changed was the environment in which factor bets get paid, and a factor-heavy book takes the brunt of that change.

What Crowded Factors Mean for C, GS and QUAL

Consider how this plays out in a real portfolio. Citigroup (C) and Goldman Sachs (GS) were both deeply entangled with the credit shock, and during the worst stretch their returns were driven far more by the financials and market factors than by anything company-specific. A quant model that held both for value reasons was not diversified in the way its optimizer believed.

We do not have position data for any particular manager, so treat this as an illustration. A value screen in 2007 would have flagged cheap financials, and so would the screens of dozens of competitors. When losses began, everyone holding the same cheap names had the same reason to cut, and the cutting itself pushed prices against all of them.

That is how crowding turns a factor exposure into a liquidity exposure. The risk model sees volatility and correlation estimated from history. It does not see the overlap between your book and everyone else's.

Factor products raise the same question for ordinary investors. A fund built around quality, such as QUAL, packages one of the three famous themes into a single ticker. The idea behind it is sound, but the more money that flows into any packaged factor, the more its return depends on flows rather than on the characteristics of the companies.

A rough scenario shows the exposure. Suppose a portfolio holds 20% in a quality-tilted fund, and the quality factor has a bad month in which it falls 4% relative to the market. That costs the portfolio 0.8% from that sleeve alone. If the same investor also owns a value fund and a momentum fund from the same family of ideas, and all three sell off together as in August 2007, the damage compounds. A 4% relative drop across 60% of the portfolio is a 2.4% hit that no single fund's risk report would have warned about.

The opportunity side of the same fact is the stable residual volatility. Single-stock dispersion held up through the crisis, so portfolios built on different information still had something to work with.

Building a Risk Process That Expects a Crowd

The talk's main prescription for alpha is original thinking. Consistent alpha needs a steady supply of new and different ideas, found by looking in places the crowd has not looked and thinking deeply about why an idea should work. Generic signals have lost their consistency, and they also tie a manager's returns to everyone else's during crowded periods.

For risk, the implication is that we must ask our models different questions. Conventional factor risk models describe the covariance of returns. They say little about how many other holders sit in the same names, or how quickly those holders might all try to leave.

A practical process adds at least three kinds of check (we count them because they are separate jobs, not because the number is special). The first is a crowding measure: how much does your portfolio overlap with the holdings peers publish or we can infer, and how has that overlap trended? The second is a stress test that takes August 2007 itself and applies the factor moves of that month to today's book. The third is a review of leverage against the speed at which we could sell the positions.

None of this replaces the standard risk model. It supplements the model with information that history alone cannot supply, because history only shows what happened to portfolios as they were then, not as they would behave once more capital stacks on top of them.

We also think the volatility decomposition deserves a permanent place on the dashboard. If factor volatility rises while residual volatility holds, as in 2007 to 2010, the right response differs from the one you would choose if residual volatility had collapsed. In the first case, factor timing and factor diversification matter more. In the second, the stock-picking opportunity has dried up, and a manager should be asking whether we can still justify the fee.

Steps to Take Before the Next August

For individual investors, the first step is an inventory. List every fund you own and look up what factor tilts it carries: value, momentum, quality, low volatility. Many people discover that three different funds share a single dominant exposure.

Second, size your factor exposure on the assumption that it can lose several percent in a month when the crowd heads for the exit. If a 4% relative loss on a sleeve is tolerable, hold it. If not, cut the sleeve until it is.

Third, prefer exposures that you can explain without referring to a backtest. A quality tilt in a fund such as QUAL is easier to defend when you know why profitable, stable companies might be undervalued, and you know the weakness: everyone else read the same academic papers.

For anyone hiring or monitoring a quant manager, the questions are concrete. Ask what the manager's signals are beyond value, momentum and earnings quality. Ask how crowding is measured, how much leverage is used, and what the portfolio would have done in August 2007. A manager who answers with specific numbers has thought about it, and one who changes the subject to a Sharpe ratio has not.

For managers themselves, the work is to find new sources of alpha and to treat the volatility split as a standing diagnostic. The data show that stock-level opportunity stayed intact, with specific volatility near 12% to 13% through the whole period, while factor volatility rose from 9.9% to 13.6%. That points to where the new effort belongs: ideas that earn their return from stock-specific information, and risk controls that account for the other people sharing the trade.

Crowds will form again around whatever works next. The investors who handle it best will be the ones who measured the crowd before it showed up in their returns.