Thirty-One Signals, Six Bets · part 4 of 10
Twelve configurations per universe (lookback 60/120/250 × hold 21/63/126/252). Turnover falls from 22–30× to 0.6–3.6×, and in small caps the spread drag collapses from 43 points to under 2. The binding constraint is removed.
What it reveals is a random field. Excess over equal-weight ranges from −228 to +225 points with a standard deviation of ~129, and the largest swing from a single parameter step is 302 points. A real effect produces a gradient; this does not. The standout cell (+225%, or +4.82%/yr) was chosen from twelve draws whose expected maximum under a pure-noise null sits about 210 points above the mean — the observed 225 is almost exactly what noise looks like once sorted.
The standout cell was then tested as a single named configuration, on windows it had not been chosen on.
| Statistic | Large cap | Small cap |
|---|---|---|
| Years positive | 5 / 9 | 4 / 9 |
| Mean excess | +5.5%/yr | +3.9%/yr |
| t-statistic | 1.07 | 1.32 |
The large-cap +225% comes almost entirely from three years — 2020, 2024 and 2026. Six of nine years are flat or negative.
Every result so far carried the market factor: one giant correlated bet whose variance swamps everything else. The stated rationale was explicit and falsifiable — at a 4%/yr excess, 16% volatility gives IR 0.25 and needs 64 years to reach t = 2, while 6% volatility gives IR 0.67 and needs 9. Same alpha; removing variance nobody is paying for was supposed to be the highest-leverage change available.
| Book | Return | Vol | Sharpe | Beta | Max DD | Margin calls |
|---|---|---|---|---|---|---|
| Long-only momentum | +519.0% | 22.3% | 1.04 | 0.96 | 30.1% | 0 |
| L/S momentum 12−1 | −49.9% | 22.1% | −0.25 | −0.14 | 55.9% | 1 |
| L/S reversal 5d | +0.9% | 17.1% | 0.09 | 0.08 | 32.0% | 0 |
| L/S low volatility | −45.2% | 17.3% | −0.31 | −0.31 | 51.2% | 1 |
| L/S combined | −40.9% | 20.1% | −0.20 | 0.05 | 55.8% | 1 |
| Buy-and-hold SPY | +242.7% | — | — | 1.00 | — | — |
Beta collapsed from 0.96 to roughly zero — the construction does what it says. The variance reduction did not follow. Volatility stayed near 20% instead of falling toward 6%, because the short leg cancels the market factor while contributing its own idiosyncratic risk: you replace one correlated bet with forty individual ones. The arithmetic that justified the whole exercise does not survive its own test.
The alpha is also negative — shorting the bottom of a momentum ranking through a decade-long bull market loses money — and three of five books hit margin calls.
One constraint is worth recording because it makes this permanently a measurement rather than a strategy: short selling is not available in an IRA, the natural home for an account like this. The long/short book exists to answer "is there alpha here at all" with the statistical power to see it, not to be deployed.