32 trend variants: the timing was real, picking the winner was not
Failure type: overfit. Verdict: not credible under the rule we set in advance. No validation run.
The idea
Hold each ETF while its trend is up (price above its 200-day average, or positive 12-month momentum), otherwise cash. Variants add a market-wide regime gate, a per-fund volatility target, dual momentum (top-k by past return, with cash if it doesn't beat T-bills), and two rebalance speeds. 32 variants.
What happened
| Training years 2006–2016 | Sharpe | Yearly return | Worst drop |
|---|---|---|---|
| Best variant (200-day, vol target 10%, monthly) | 0.70 | 4.35% | 5.3% |
| SPY buy and hold | 0.41 | 7.50% | 55.2% |
| SPY at the strategy's own day-by-day invested level | 0.63 | 4.90% | 10.2% |
The best variant beat every benchmark at default and doubled costs. Costs barely mattered (0.03–0.24% a year).
Why it failed
The rule we wrote beforehand said the best variant must also keep the family's probability of backtest overfitting at or below 0.50. It came in at 0.619: in 62% of re-splits, the in-sample winner landed below the out-of-sample median. The deflated Sharpe was 0.945 against a 0.95 bar.
The honest reading: trend timing is real in-sample (SPY held on the same days reaches a Sharpe of 0.63 with a 10% worst drop), but the variants are close substitutes, and choosing the "best" one among them is mostly noise. Per the rule, nothing went on to validation. The plain, untuned version of this idea has its own page.
Receipts: the strategy family, the sweep, and the credibility rule were written down before the run. 32 trials plus 64 cost-sensitivity runs.
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.