8.8.3. Momentum Crashes

Momentum Crashes

Momentum crashes (Daniel & Moskowitz, 2016) are infrequent but severe drawdowns of cross-sectional momentum strategies, in which the WML portfolio loses $-50\%$ to $-75\%$ over one or two months. Crashes are not random — they occur in identifiable panic states, defined empirically as periods following sustained bear markets when realized market volatility is elevated. The two largest U.S. momentum crashes on record are:

• August $1932$: WML returned approximately $-74.4\%$ in a single month, in the early-rebound phase out of the Great Depression bottom.
• April $2009$: WML returned approximately $-45.9\%$, in the early-rebound phase out of the Global Financial Crisis bottom.

The mechanism is conditional-beta time variation. Unconditionally, the WML portfolio has ${\beta }_{\text{WML}}\approx -0.08$ — close to zero. In panic states the conditional beta becomes sharply negative, often $-2$ or more, because the loser leg consists of high-beta survivors that have fallen the most and the winner leg consists of defensive low-beta names. When the market sharply rebounds from its panic-state low, the loser leg (which the strategy is short) skyrockets, while the winner leg (long) lags. Both legs lose simultaneously, and the long-short strategy implodes.

The Daniel-Moskowitz dynamic momentum strategy scales the WML position based on a panic-state indicator (typically built from the prior $24$-month market return and recent realized volatility) and approximately doubles the unconditional Sharpe ratio of static momentum. The dynamic strategy is the practical resolution to the crash problem: avoid being short losers in the late-bear-market / early-rebound regime.

Momentum crashes are the single most important risk-management consideration for momentum strategies. They imply that the unconditional CAPM beta of $-0.08$ from Jegadeesh & Titman is a deeply misleading summary statistic and that any risk model used to size momentum exposure must condition on the prevailing market state.

Intuition

Momentum works most of the time and then suddenly fails in dramatic fashion. The average Sharpe ratio of WML is around $0.5$, but that average hides an episodic pattern: long stretches of steady gains interrupted by single months in which the strategy loses half its value or more. Two months — August $1932$ and April $2009$ — together account for a large fraction of all the variance in WML over the $1927$–$2013$ U.S. sample.

The mechanism is the composition of the loser leg in panic states. After a long bear market, the stocks that have fallen the most tend to be high-beta financials, cyclicals, and distressed names whose equity is little more than a bet on survival. The winner leg, by contrast, has filled up with defensive stocks (utilities, staples) that held up. The “long winners, short losers” portfolio has quietly turned into “long defensives, short high-beta.” When the market sharply rebounds, the high-beta short leg explodes upward and the defensive long leg lags — both legs lose at once. Daniel & Moskowitz show this state can be flagged in real time using just the prior $24$-month market return and recent realized volatility; cutting momentum exposure in panic states roughly doubles the Sharpe.

Properties

(Panic-State Indicator) The two triggers are a negative trailing $24$-month market return and elevated recent realized volatility.

Both can be computed in real time from market data alone — no fundamentals, no factor model, no peeking at returns inside the formation window. Daniel & Moskowitz show that requiring both triggers (rather than either one) is what gives the rule its precision: bear markets without elevated volatility, and high-volatility periods that are not preceded by a sustained drawdown, both produce many false positives. The combined indicator fires for a small fraction of months but covers all $11$ of the worst WML months in the $1927$–$2013$ sample.

(Loser-Leg Composition) In a bear market the loser leg fills with high-beta survivors that behave like call options on market recovery.

Late in a long downturn the cross-section has reshuffled: the worst-performing names are typically distressed financials, cyclicals, and other high-beta firms whose equity has become a residual claim on survival. Their option-like payoffs have huge upside conditional on the market rebounding — exactly the wrong characteristic to be short. Symmetrically, the winner leg is dominated by defensive low-beta stocks (utilities, staples) that held up. The “long winners, short losers” portfolio has therefore quietly become “long defensive, short high-beta,” and unconditional WML beta statistics computed across all states miss this hidden compositional shift.

(Beta Flips Mid-Rebound) The conditional beta is negative late in the bear market and turns sharply positive once the rebound begins.

This flip is what makes the crash so violent. Early in a panic state the strategy is short-loser-leg-of-high-beta, so it has a deep negative beta and loses if the market rises. Once the rebound is underway, the loser-leg shorts have already gone up so much that they now constitute a long position in market beta from the strategy’s risk-management perspective — momentum then becomes long-beta into a market that has already moved. Both regimes lose: first short-beta into the rebound, then long-beta into the inevitable consolidation. A static hedge sized off the unconditional beta cannot capture either phase.

(Cross-Asset Generalization) The crash mechanism is not unique to U.S. equities — time-series momentum suffers parallel tail events for the same structural reason.

Time-series momentum crashes occur when correlations across asset classes spike (March $2020$, late $2008$): the diversification that holds the cross-asset Sharpe above $1$ in normal months disappears, every contract reverses at once, and the realized portfolio volatility overshoots the target the position sizes were calibrated against. Different asset, same lesson — momentum strategies carry tail risk that is invisible in long-run averages and clustered in identifiable regimes.

Examples

August $1932$: the prototype momentum crash

By summer $1932$, the U.S. equity market had fallen approximately $80\%$ from its $1929$ peak. The WML portfolio formed at the end of July $1932$ was long stocks that had held up best (defensive utilities and consumer staples) and short stocks that had fallen the most (financials, railroads, and other high-beta cyclicals). The cumulative $24$-month market return was deeply negative, putting the strategy in a panic state by Daniel & Moskowitz’s definition.

In August $1932$, the market rallied approximately $+38\%$ in a single month — the largest one-month gain in U.S. equity history. The high-beta loser leg (which WML was short) gained roughly $+80\%$. The defensive winner leg (which WML was long) gained only about $+15\%$. The long-short combination therefore returned approximately $-74\%$ in that single month — wiping out years of accumulated WML returns.

The conditional beta of the strategy in July $1932$ was approximately $-1.7$, vastly larger in magnitude than the unconditional $-0.08$. A naive Value-at-Risk calculation based on the unconditional moments would have predicted essentially no tail risk in the strategy. The actual loss was an order of magnitude larger.

April $2009$: the modern replication

By March $2009$, the S&P 500 had fallen approximately $57\%$ from its October $2007$ peak. Past-loser stocks were dominated by financials (Citigroup, Bank of America, AIG, Lehman survivors) trading at low single-digit prices, and past-winner stocks were dominated by defensive names that had held up. The cumulative $24$-month market return was deeply negative; realized volatility was at a multi-decade high. Both panic-state triggers fired.

In March and April $2009$, the market rallied roughly $+25\%$ from its trough. The financial stocks in the loser leg gained $+150\%$ to $+300\%$ from their March lows. The WML portfolio lost approximately $-46\%$ in April $2009$ alone, the second-worst month in the entire $1927$–$2013$ sample. As in $1932$, the strategy was caught with a large negative conditional beta into a sharp rebound, and the short-loser leg blew up.

The April $2009$ crash is more recent and better-documented than $1932$, and it is the empirical anchor for the Daniel-Moskowitz paper. It also provides a test for any candidate panic-state indicator: any indicator worth its name must have flagged March $2009$ as a panic state in real time.

The dynamic strategy's Sharpe doubling

Daniel & Moskowitz construct a dynamic WML strategy whose position size scales with a function of the panic-state indicator: full long-short exposure in normal states, partial or zero exposure in panic states. They estimate the optimal scaling in-sample and report out-of-sample performance separately to guard against overfitting.

The static unconditional WML strategy has a Sharpe ratio of approximately $0.5$ in their U.S. sample. The dynamic strategy achieves a Sharpe ratio of approximately $1.0$ — roughly double — almost entirely by avoiding the worst $5\%$ of months. The average return in normal states is essentially unchanged. The improvement is not the result of a clever new signal; it is the result of not trading the static signal in the small handful of months when the conditional risk is unmanageably large.

The deeper takeaway is methodological: any factor strategy whose unconditional Sharpe ratio is computed without regard to state-dependent tail risk is reporting a number that is both too high (because it ignores the crash months) and too dangerous (because it implies the strategy is safer than it is).

See Also

• 8.8. Momentum — parent hub.
• 8.8.1. Cross-Sectional Momentum — sibling: the strategy whose conditional-beta failure mode is documented here.
• 8.8.2. Time-Series Momentum — sibling: trend-following analogue with parallel tail-risk behavior.
• 8.3. Challenges to Rational Pricing — background on the broader predictability puzzle that momentum exemplifies.