2 The Predictable Tail
Lottery Characteristics Price Drawdown, Not Return
SSRN: abstract_id=6899979 Status: Live preprint
2.1 Abstract
We test whether payoff asymmetry (PA) — derived from the Hill estimator of individual equity return distributions — predicts forward equity returns when computed prospectively over 106,163 stock-month observations (715 U.S. equities, 2009–2025). Using distribution-free statistics appropriate for fat-tailed data, we find: (1) PA is a monotonic, symmetric, inverted predictor — low-PA names outperform with a 63.6% win rate and +4.90% 60-day median, while high-PA names underperform; (2) the same tail class (SUPER_FAT, α ≤ 2) produces 70.4% win at the bottom of PA and 34.8% at the top — tail class alone is insufficient, direction matters; (3) Fama-MacBeth regressions fail to detect the signal (t = 1.37) because it is nonlinear and tail-concentrated; (4) common backtests using current PA values are look-ahead-contaminated, overstating the apparent edge by 35+ percentage points. The mechanism is the equity lottery premium: high-PA stocks exhibit positive-skew, negative-drift payoff structures; low-PA stocks exhibit positive drift with occasional large left-tail events.
2.2 1. Introduction
The practitioner intuition is intuitive: high payoff asymmetry — a return distribution with large upside relative to downside — should predict positive future performance. A stock whose historical gains have consistently exceeded its historical losses in magnitude sounds like a stock worth holding.
The data says the opposite.
This chapter documents the inversion and its mechanism. It also documents a methodological trap: computing payoff asymmetry from current data and testing it against past outcomes — a form of look-ahead contamination that inflates apparent predictive power by 35 percentage points and flips the sign of the conclusion.
The inversion is not a curiosity. It is a consequence of the equity lottery premium (Boyer et al. 2010): stocks with lottery-like payoff structures (large right tail, positive skew) attract speculative demand that raises prices and depresses subsequent returns. Payoff asymmetry, correctly computed prospectively, identifies this structure in real time.
2.3 2. The Payoff Asymmetry Metric
Definition. For a return series \(\{r_t\}\) over a trailing \(T\)-day window:
\[PA = \frac{E[r \mid r > 0]}{E[|r| \mid r < 0]}\]
A stock with PA > 1 has average gains exceeding average losses in magnitude. A stock with PA < 1 has average losses exceeding average gains.
PA is computed from a 756-day (3-year) trailing window of daily returns, prospectively — using only data available before the evaluation date. The tail class (THIN, SUBEXP, FAT, SUPER_FAT) is derived from the Hill estimator applied to the same window.
2.4 3. Main Results
| PA Group | N | Win% | Median 60d Return |
|---|---|---|---|
| Bottom 2.5% | 423 | 66.7% | +6.14% |
| Bottom 10% | 1,694 | 63.6% | +4.90% |
| Baseline | — | 59.7% | — |
| Top 10% | 1,695 | 53.3% | +1.91% |
| Top 1% | 170 | 40.6% | −4.60% |
The bottom-10% 95% CI [61.2%, 65.9%] does not overlap with baseline [59.0%, 60.4%]. Mann-Whitney bottom 10% vs top 10%: p < 10⁻⁶.
Within SUPER_FAT (α ≤ 2): bottom 10% PA → 70.4% win (+7.82% median); top 10% PA → 34.8% win (−8.76% median). The same tail class, opposite outcomes.
2.5 4. Why Standard Methods Miss It
On the same 112,152 observations (T = 202 monthly Fama-MacBeth periods):
- Spearman rank correlation (median-like, robust): PA ↔︎ return = −0.045, p < 10⁻⁴
- Linear Fama-MacBeth (conditional mean, OLS): PA coefficient t = +1.74 → +0.13 in the full model (effectively zero)
OLS fits the conditional mean. Fat tails inflate the mean at the high-PA extreme (top 1%: median −5.04% but mean +8.72%). Rank methods neutralize the moonshots and reveal the typical outcome. This is Taleb (2025) §3.4 Consequences 2 and 7 demonstrated simultaneously on real equity data.
2.6 5. The Look-Ahead Contamination
Computing PA from current data and testing against past outcomes creates an artifact: a stock that has already risen shows high PA and high realized returns simultaneously. This is circular.
| Method | Win Rate (top decile) |
|---|---|
| Contaminated (current PA, past outcomes) | 76–100% |
| Prospective (PA before evaluation date) | 40.6% |
| Contamination | ~35 percentage points |
The contaminated result is what most practitioners compute. The prospective result is what actually occurs.
2.7 6. Conclusion
Payoff asymmetry predicts forward equity returns — in the opposite direction from naive expectation. The mechanism is the equity lottery premium. Standard linear factor models cannot detect the signal because it is nonlinear and tail-concentrated. Common backtest implementations are contaminated by look-ahead bias that inflates apparent win rates by 35+ percentage points and reverses the conclusion.
The implication for the taxonomy developed in subsequent chapters: measuring the tail exponent α alone is insufficient. The direction of payoff asymmetry within a tail class is the economically relevant variable.
→ Chapter 2 explores why this direction varies systematically across identifiable subgroups within the SUPER_FAT class.