Stop guessing. Start building.

This is what happens when you stop averaging 33.5 years into one number. 40 buy-the-dip entries on SPY. Each one held four days, measured in three periods with the same number of bars against that period's own four-day drift, and sorted by what the three results did. Stable, 9: positive in all three, the latest within half to 1.5 times the earliest. Strengthening, 7: positive in all three, the latest more than 1.5 times the earliest. Decaying, 11: positive in all three, but the latest under half the earliest. Unstable, 10: the edge changed sign in at least one period. Thin sample, 3: under 20 trades in a period, not judgeable. 16 of 40 held up. That is the honest headline. Now hold it against the ranking. Over the whole window, 35 of the 37 rankable entries beat drift. Of the top 10 in that ranking, 6 are Decaying and only 4 held up. The ranking is not wrong. Every edge in it is real, measured over the whole window. It is just silent about whether the thing it measured is still happening. And the top of the list is where that costs the most, because the top is what you act on. A single blended number cannot show you which of these five an entry is. Three periods can. The stability verdict for every entry, and what it does to the ranking: https://statoasis.com/overfit/research/mean-reversion-buy-the-dip-ranked

Everyone tells you to test out of sample. The part that gets left out: you only get to do it once. Look at the held-back data, change a rule, look again, and it is not out of sample any more. You have fitted to it. So the useful question is which check you can run before that one honest look, and whether it predicts the result. I tested 11 of them on 5,472 strategies across 8 markets. Every check was computed on the first 75% of history. The final 25% stayed sealed until the end. With no checks at all, 63.7% of the strategies were still profitable in that sealed quarter. That is the bar. One check took it to 91.7%: are the parameter settings either side of yours also profitable? It passed 53.1% of candidates. Lists picked on a separate selection panel added more checks. Two reached 92.4%. Four reached 93.1%. The pass rate fell to 39.0%. One check did 95.4% of the four-check list's gain. Check the neighbours. Then sit the exam once. Every table, the selection panel and every limit: https://statoasis.com/overfit/research/is-your-backtest-overfit-validation-checklist

A 74.6% win rate is not an edge if sitting still pays more. Three rules for shorting VIX futures get published together, and the number that makes them look good is the win rate. I rebuilt all three on 4,924 daily bars of the Cboe contract and swept 2,520 variants around them. The measured win rate is 74.6%. Time in market is 39.62%. The average trade across twenty years is $173. 2009 was $1,298. 2021 was $1,267. Those are the only two years whose average trade cleared $1,000. Across the same twenty years, one trade lost $26,060. Simply staying short the contract across the same 4,924 bars earned $187,070, with a $77,420 worst drawdown. The timed version earned $64,720, with a $54,620 worst drawdown. That is 34.6% of what sitting still paid. Holding twenty days made $122,790 against $33,140 for a same-day exit, while the worst drawdown goes from $59,210 to $63,800 and the trade count falls from 1,216 to 186. The two-day rung in between lost $17,970. The extra money is days in the market, not more trades. A short VIX futures position is a flat you rent out. These rules kept it empty three days in five. Check your benchmark before you believe your win rate. Full tables and the hold ladder: https://statoasis.com/overfit/research/master-volatility-futures-vix-trading-strategies