The long-only studies showed factors can't beat the index (a long book is ~90% market beta).
The real test of a factor is its dollar-neutral spread: long the top quintile, short the
bottom quintile, beta stripped, after turnover + borrow costs. If there's alpha, it shows up
here as a positive net Sharpe with beta ≈ 0. Survivorship-free CRSP, PIT S&P 500, 2005–2024
(ML combiner from 2012), monthly, 5 bps/side slippage + 50 bps/yr borrow. Reproduce:
scripts/crsp_longshort_study.py.
| signal | ann return | ann vol | Sharpe | max DD | market beta | turnover |
|---|---|---|---|---|---|---|
| earnings_yield | −1.88% | 11.3% | −0.17 | −44.6% | 0.18 | 0.46 |
| book_yield | −2.53% | 13.4% | −0.19 | −57.3% | 0.26 | 0.28 |
| roe (quality) | +1.07% | 10.0% | +0.11 | −27.6% | −0.18 | 0.28 |
| reversal_1m | −2.09% | 14.2% | −0.15 | −53.8% | 0.35 | 3.13 |
| momentum_12_1 | −4.53% | 19.5% | −0.23 | −75.2% | −0.51 | 0.95 |
| low_vol | −3.46% | 21.4% | −0.16 | −78.9% | −0.94 | 0.28 |
| candidate (val+rev) | −3.06% | 13.7% | −0.22 | −57.1% | 0.33 | 2.18 |
| ML combiner | −4.19% | 7.6% | −0.55 | −42.9% | 0.13 | 2.19 |
- Every factor's net long-short Sharpe is ≈0 or negative. Only quality (roe) is barely positive (+1.1%/yr, Sharpe 0.11) — economically negligible and not robust.
- The ML combiner is the worst (Sharpe −0.55) — combining edgeless factors just overfits.
- Betas aren't perfectly zero (low_vol −0.94, momentum −0.51): those factors carry structural market exposure by construction, so they aren't even cleanly neutral — and still no alpha.
- reversal's turnover is 313%/side/month → costs alone would bury any raw signal.
This is the correct, well-documented modern result: classic academic factors on large-cap US equities have been largely arbitraged away (especially post-~2003), and after realistic costs + borrow a retail trader cannot extract them. Long-only they're just beta; long-short they're flat-to-negative.
Across first_study → survivorship_study → multifactor_study → this: a rigorous, survivorship-free, out-of-sample, cost-aware, benchmarked, long-only AND market-neutral evaluation. The honest output is "no edge here" — which is the valuable result. The +372% we started with was a survivorship + look-ahead mirage; under rigor it vanishes. Most retail "edges" are exactly such artifacts; this pipeline is built to refuse them.
- Smaller / less-liquid universe (mid/micro caps): arbitrage capital can't fish there, so factor premia may survive — but borrow is hard/expensive and capacity is tiny. The CRSP lake can be widened beyond the S&P 500 to test this directly (the next clean experiment).
- Non-classic signals: alternative data, shorter horizons, cross-sectional interactions — validated with this same survivorship-free, cost-aware harness.
- Sober meta-conclusion: beating efficient large-cap US markets is genuinely hard. The durable contribution of plutus is the methodology that tells the truth, reusable for any future signal — not this (null) result on stale factors.