The foundation under everything above. Most backtests quietly cheat twice: they test on today’s
index members (so every bankruptcy is excluded in advance), and they use fundamentals as later revised
(so the model knows numbers before they were filed). We built the boring fix: historical S&P 500
membership month by month, and SEC fundamentals as originally filed, usable only after their filing
date. Then we made the overnight run test everything we could think of.
What died
- Theme/industry ETF rotation (48 ETFs, every momentum flavor): mostly below random-pick baselines.
- Earnings and news-jump drift on big moves: roughly zero follow-through at 5–60 days.
- Pure price factors on large caps (momentum, 52-week high, low-vol, reversal, deep dips): nothing significant.
- Revenue acceleration: worked in-sample 2010–17, decayed to nothing out-of-sample. A textbook decayed factor.
What survived
One composite of value, quality and growth measured from as-filed data: +5.2%/yr over the equal-weight
universe (t 2.2) — and stronger out-of-sample (2018+: +8.0%/yr, t 2.1) than in-sample, the
opposite of an overfit signal. Against 1,000 random 20-stock portfolios it beat the 99th percentile.
Turnover is low; costs take ~0.3%/yr.
Verdict: adopted as the fund’s stock-selection engine. One survivor out of a dozen ideas is the
point — the graveyard is what makes the survivor believable.