The Quants Ch. 2: August 2007 — The Week the Models All Broke at Once

"It was a hundred-year flood, and then it happened again the next day, and again the day after that." — the recurring description of the week from participants who lived it

Investment Background

Between Monday 6 August and Thursday 9 August 2007, quantitative equity market-neutral strategies — long undervalued stocks, short overvalued stocks, hedged against the market — suffered losses that their own risk models classified as effectively impossible.

Goldman Sachs's Global Equity Opportunities fund lost roughly 30% in a week. Its flagship Global Alpha fund was down sharply for the year. Renaissance's institutional equities fund reported an 8-9% loss in days. Highbridge, AQR, Tykhe, Black Mesa and dozens of smaller shops experienced the same pattern in near-perfect synchrony. Morgan Stanley's PDT desk, run by Peter Muller and one of the most consistently profitable trading operations on Wall Street, lost hundreds of millions.

Goldman's chief financial officer, David Viniar, gave the quote the episode is remembered by: the firm was seeing "25-standard-deviation moves, several days in a row."

That statement deserves careful attention, because it is not a description of the market. It is a confession about the model. A 25-standard-deviation event under a normal distribution has a probability so small that it would not be expected to occur once in a period vastly longer than the age of the universe. Observing several in a row does not mean the universe delivered a miracle. It means the distribution being used to measure the event was the wrong distribution.

And here is the detail that makes this week the single most instructive episode in quantitative finance: the strategies were not wrong. By late August, most of the positions that had been slaughtered had substantially recovered. The value stocks the models liked went on to outperform. The expensive stocks they were short went on to lag. The models had correctly identified the relationships. The funds still lost billions, and several never recovered their assets under management.

This is the defining lesson of the book: you can be right about the destination and be destroyed by the path.

The Wall Street Translation

What Actually Happened, Sequentially

The mechanism had nothing to do with equity valuation and everything to do with balance sheets elsewhere.

Stage one — the shock came from outside. By mid-2007, subprime mortgage credit was deteriorating. Two Bear Stearns credit hedge funds had collapsed in July. Losses were accumulating in structured credit books, in mortgage warehouses, and at multi-strategy funds with credit exposure. None of this had anything to do with the relative value of cheap versus expensive equities.

Stage two — a multi-strategy fund needed cash, and equities were the only liquid asset it owned. A fund carrying losses in illiquid mortgage positions cannot sell those positions; there is no bid. To raise cash or reduce gross exposure it sells what can be sold. Quantitative equity books are highly liquid by construction. So the healthy, working, profitable book was liquidated to fund the sick one.

Stage three — the liquidation moved prices against everyone holding the same book. Because the industry had converged on similar factor exposures, one large forced seller pushed down exactly the stocks that every other quant fund was long, and pushed up exactly the stocks every other quant fund was short. Long positions fell, short positions rose, and the market-neutral hedge provided no protection whatsoever, because the loss was not coming from market direction.

Stage four — the losses triggered risk limits, which produced more selling. Each fund's risk system observed an anomalous drawdown and instructed it to reduce leverage. Reducing leverage means selling longs and covering shorts — the identical action the first seller took. The response to the shock was mechanically the same as the shock itself.

Stage five — the loop tightened. More selling produced more losses, which triggered more limits, which produced more selling. For roughly three days, the dominant driver of quantitative equity prices was not information about companies. It was other quant funds deleveraging.

Stage six — it stopped. By Thursday afternoon and Friday, some participants had finished deleveraging, and a few funds with capital and nerve stepped in to buy. Prices rebounded sharply. Anyone who had been forced out on Wednesday missed the recovery entirely and locked in the loss permanently.

The Asymmetry That Determined Who Survived

Fund characteristic	Outcome that week
High leverage, tight risk limits	Forced to sell into the cascade; loss realised permanently
Redeemable investor capital, monthly liquidity	Faced redemptions on top of losses; forced selling compounded
Low leverage, permanent or locked capital	Able to hold, or to buy; largely recovered within weeks
Diversified across genuinely different mechanisms	Losses in one sleeve, offsets elsewhere; survived
Multi-strategy with correlated funding needs	Sick book forced liquidation of healthy book; worst outcome

Notice what is absent from this table: forecasting accuracy. Skill in identifying mispriced securities does not appear as a determinant of survival, because in that week it was not one. Survival was determined entirely by whether an institution retained the option to wait.

Why the Models Could Not See It

Risk models in 2007 estimated the distribution of returns from the historical behaviour of the positions. That approach embeds an assumption so basic it is rarely stated: that prices are generated by an external process the fund observes but does not influence.

In August 2007 that assumption inverted. Prices were being generated by the funds themselves. The variable driving returns was the aggregate leverage and liquidation schedule of the quant industry — a variable that appeared in no model because no fund could observe the others' books.

The risk was not in the positions. It was in the ownership structure of the positions, and ownership structure was invisible from inside any single institution.

Execution Rules

1. Distinguish permanently from temporarily, and design your portfolio so losses stay temporary. A drawdown you hold through is a fluctuation; the identical drawdown you are forced to sell into is a permanent capital loss. The entire difference is whether you had to transact. Structure your portfolio so that no market event can compel you to sell — that is the mechanism that converts volatility from a threat into noise.

2. Hold enough non-correlated liquidity to fund every obligation for at least two years. In August 2007 the funds that survived were the ones that did not need cash that week. For a retiree the equivalent is two to three years of spending needs in Treasury bills, short-duration government bonds, or cash — assets that do not fall when equities fall, held specifically so that no market condition forces a sale of the equity core.

3. Refuse leverage in any form, including the implicit kinds. Margin debt is obvious. Less obvious: a mortgage carried into retirement against a portfolio funding the payments, leveraged or 2x ETFs, options sold naked, and a spending rate high enough that a bad sequence forces liquidation at depressed prices. Each of these removes your ability to wait, which is the only defence this chapter identifies.

4. Never add to a position because a model says it has become cheaper during a cascade. In August 2007 the signal to "buy more" was mathematically correct and financially fatal, because the price was being set by forced sellers rather than by information, and there was no way to know how much more selling remained. If you rebalance during stress, do it on a pre-committed calendar schedule with pre-committed amounts, never on a discretionary read of how attractive the opportunity looks.

5. Cap the total portfolio weight of every strategy that requires a functioning market to exit. Any position whose thesis depends on being able to trade out at a fair price — factor funds, liquid alternatives, individual small-cap holdings, tactical sleeves — belongs in a defined satellite allocation with a hard ceiling. Set that ceiling at a level where a total loss in the sleeve does not change your retirement plan, and hold the remainder in a broad, low-cost index core you never intend to trade tactically at all.

Retirement Application

The retiree's version of August 2007 is not a factor unwind. It is a market decline that coincides with a spending need — the sequence-of-returns problem — and the structural solution is identical.

Every fund that was destroyed that week was destroyed by being a forced seller. Every fund that survived, survived by being able to wait. A retiree drawing 4% from a portfolio in a 30% drawdown is a forced seller, mechanically, every single month, and is selling at the worst prices in the cycle. A retiree with two years of spending in Treasury bills is not.

That is the whole retirement lesson, and it costs nothing to implement. It requires no forecast, no model, no view on valuations, and no skill. It requires only the willingness to accept a slightly lower expected return on a small slice of the portfolio in exchange for never being on the wrong side of the mechanism that destroyed Goldman Sachs's quant funds.

The low-cost index core does the compounding. The short-duration cash and bond buffer preserves the option to wait. Nothing tactical is required, and any tactical position added must not compromise either of those two functions.

Risk Management

• Correlation risk is a liquidity phenomenon, not a statistical one. Assets do not become correlated because their fundamentals converge; they become correlated because the same investors are forced to sell them at the same time. This is why measured correlations are lowest exactly when they are least useful.

• Model failure risk compounds with model confidence. The funds with the most sophisticated risk systems reduced exposure fastest, which meant they sold hardest into the cascade. A superior risk model, in that specific week, produced an inferior outcome.

• Redemption risk applies to funds, not just to you. If you hold an actively managed fund or an alternative strategy, other investors' redemptions can force your manager to liquidate at bad prices. You inherit their behaviour. Broad index funds redeeming in kind are structurally far more robust here.

• Recovery-window risk. Prices rebounded within weeks in August 2007. If your plan required cash during those specific weeks, the eventual recovery was irrelevant to you. Always measure risk against your own cash-flow calendar, not against the eventual path of the index.

What This Chapter Cannot Do

This chapter cannot help you predict the next cascade. Its central point is that the cascade was invisible in advance to the best-informed participants in the market, using the best data available, and that no retail investor will do better.

It also cannot argue that quantitative strategies are worthless — several of the funds involved went on to produce excellent long-run records. What it can do is establish that returns are earned by surviving the path, and that the mechanism of survival is structural rather than analytical.

---

Key Takeaway: In August 2007 the models were right about the stocks and wrong about the world. Value recovered, the shorts underperformed, and the funds still lost billions — because correlation went to one, leverage removed the option to wait, and forced selling set prices for three days. The retiree's protection is not a better model. It is a low-cost index core, no leverage, and two to three years of spending in cash and short bonds, so that no week in the market can ever make you a forced seller. Any tactical position you hold is a hedge beside that core, never a replacement for it.