2 The Predictable Tail
Historical Drawdown Predictability in a Survivorship-Conditioned Equity Panel
SSRN: abstract_id=6899979 Status: Corrected working-paper chapter
The earlier chapter presented payoff asymmetry as an independent return predictor. An exact- 60-session sector reanalysis does not support that conclusion: the aggregate relation disappears when Healthcare is excluded. This version retains the historical drawdown finding, corrects the reported IC ratio from sixtyfold to approximately 43-fold, and states the panel’s survivorship limitation explicitly.
2.1 Abstract
This chapter compares the historical cross-sectional predictability of forward returns and subsequent maximum drawdown in 715 U.S. equities observed from 2009 through 2025. The universe is survivorship-conditioned; only feature construction is point-in-time. Total volatility has a rank information coefficient near \(-0.01\) for forward 60-session return and near \(-0.43\) for realized 252-session maximum drawdown, an approximately 43-fold difference in absolute magnitude. Payoff asymmetry, idiosyncratic skewness, and MAX show the same qualitative inversion. The earlier payoff-asymmetry return result is composition dependent: +6.08% per formation period (\(t_{NW}=2.65\)) in the full sample and \(-0.27\%\) (\(t_{NW}=-0.09\)) ex-Healthcare. The supported conclusion is that distributional characteristics warrant study as historical drawdown-risk descriptors. No independent return premium, causal sector mechanism, tradable implementation, or prospective efficacy result is established.
2.2 1. Question and evidence boundary
The narrow question is whether trailing distributional characteristics contain more historical rank information about future path loss than about future terminal return. This is a measurement question, not a claim that one statistic forecasts alpha or that a tail exponent is causal.
Features use only observations available by each formation date. The membership universe was assembled from securities available in the later data environment, however. It is therefore a survivorship-conditioned historical panel, not a point-in-time reconstruction of the investable universe. Delisted firms and failures may be underrepresented. Results describe the observed panel and cannot establish market-wide prevalence.
Repeated monthly formations for one company also overlap. Raw stock-month rows are not independent experimental units. Dependence-aware inference and an explicit distinction between historical discovery and prospective confirmation are required.
2.3 2. Outcomes and principal result
The return outcome is forward 60-session return. The risk outcome is maximum peak-to-trough drawdown over 252 sessions. They answer different questions and are evaluated separately.
| Predictor | Forward-return rank IC | 252-session drawdown rank IC | Interpretation |
|---|---|---|---|
| Total volatility | approximately \(-0.01\) | approximately \(-0.43\) | Much stronger historical drawdown association |
| Lottery/distributional axis | weak and unstable | materially stronger | Candidate risk descriptor, not return alpha |
The previous manuscript called the first comparison sixtyfold. With the rounded reported values, \(0.43/0.01\) is approximately 43. Because the return coefficient is close to zero, the ratio is sensitive to rounding; both component coefficients should always accompany it.
Maximum drawdown is a path outcome rather than a terminal-return moment. The stronger association is economically interesting, but it does not prove incremental value beyond every conventional risk measure, persistence in new regimes, or a capital-allocation threshold.
2.4 3. The return-anomaly claim fails a sector gate
Payoff asymmetry is the ratio of average positive daily return to average absolute negative daily return over a trailing window:
\[ PA = \frac{E[r\mid r>0]}{E[|r|\mid r<0]}. \]
The timing-clean aggregate return relation initially appeared to survive several checks. A later exact-60-session sector analysis provides a decisive limitation:
| Signal | All sectors spread / NW \(t\) | Ex-Healthcare spread / NW \(t\) |
|---|---|---|
| Payoff asymmetry | +6.08% / +2.65 | -0.27% / -0.09 |
| MAX | +4.59% / +3.18 | -0.19% / -0.11 |
In equal-size random ticker exclusions, the MAX spread usually retains part of its magnitude; loss of observations alone is therefore not a complete explanation. Healthcare-only estimates are imprecise, so the evidence supports composition dependence rather than a causal Healthcare mechanism. Static/current sector membership and incomplete delisting treatment are additional limitations.
2.5 4. Measurement and validation implications
A clean formation clock and Newey-West standard errors (Newey and West 1987) are necessary controls, but they do not address sector composition, survivor selection, multiple search paths, or implementation. Likewise, a Probability of Backtest Overfitting estimate (Bailey et al. 2017) is conditional on the evaluated candidates, splits, purge, embargo, and loss. A point estimate of zero does not prove that the true overfitting probability is zero.
The earlier terminal-return cap is not an executable stop-loss backtest. It omits intraperiod paths, gap-through losses, borrow and recalls, spreads, impact, option prices, and portfolio capital constraints. Profitability and tradability claims based on that cap are withdrawn.
2.6 5. Prospective test
The forward design will count canonical episodes rather than raw detections, use identical realized security and benchmark clocks, retain delisting and permanent-unresolved states, and evaluate 20-, 60-, and 120-session outcomes separately. Sector, liquidity, listing age, and regime are predeclared strata. Historical rows remain discovery evidence and never enter the prospective sample size.
2.7 6. Conclusion
In this survivorship-conditioned panel, distributional characteristics are much more informative about subsequent drawdown rank than about subsequent return rank. That asymmetry is the supported contribution. The independent payoff-asymmetry return premium, a causal Healthcare mechanism, a terminal-cap trading implementation, and universe-wide generalization are not supported.
→ Chapter 2 turns the rejected hard taxonomy into a registered event-attribution question.