Weekly reversal: great on 2005–2016, nothing on 2017–2021
Failure type: edge decayed. Verdict: parked, no re-tuning.
The idea
Short-term reversal: ETFs that fell hardest over the last few days tend to bounce. Rank by 5-day return or 2-day RSI, hold the three or five most oversold, rebalance every 5 or 10 sessions.
What we did
- Training sweep (2005–2016): 32 variants, with realistic costs and settled-cash rules.
- Fair comparison: against SPY held at the same invested level and at the same volatility, because a strategy that is only invested part of the time looks safer for free.
- Pre-registered validation (2017–2021): three variants chosen and written down before the run. 9 runs.
What happened
On training years the best variant had Sharpe 0.689 against 0.404 for SPY, with a 15.6% worst drop against 55.2%. It also beat both matched SPY mixes in 7 of 8 comparison rows. Its deflated Sharpe, though, was 0.893 (bar 0.95), and most of the gain came from lower risk at about 57% invested, not higher return.
Raising per-fund caps made things worse: more trading, more drawdown, lower Sharpe (0.688 down to 0.643 for the best variant).
On validation, no variant passed:
| Variant | Sharpe (default costs) | Yearly return | Worst drop | SPY buy and hold |
|---|---|---|---|---|
| A | 0.074 | 0.96% | 32.2% | Sharpe 0.923, 18.3% |
| B | 0.108 | 1.52% | 32.2% | same |
| C | 0.021 | 0.15% | 35.5% | same |
Why it failed
Costs were not the cause: with zero costs the best Sharpe was only 0.144. The edge simply was not there in a strong trending market. We did not re-tune on the validation years, as the pre-registration says.
Receipts: the training sweep came first; the validation rules were written down and committed before any validation run. 32 + 9 official trials, plus 64 cost-sensitivity and 12 fair-test runs recorded.
Research and education only. These are paper results from historical data on a limited set of funds. Not investment advice, not a recommendation, not a prediction. See Legal.