Trading strategies that failed our tests, with the receipts.
Receipts for the strategies that failed
Trading strategies that failed our tests, with the receipts.
Receipts for the strategies that failed.
Most trading sites sell the thing that supposedly works. This one publishes the things that failed our tests: each idea's rules written down before the test, how many tries it took, how it did in more than one window and market, and a plain verdict on why it failed those tests. These pages publish the finding and the headline evidence. The full tables and code are not public yet.
Based on public daily price data; we publish derived statistics only. See about.
Weekly reversal: buy the most oversold ETFs and hold a few weeks. Numbers from the write-up; this is paper research, not a recommendation.
The pitch Buy the bounce
Rank the 23 ETFs by 5-day return or 2-day RSI, hold the three or five most oversold, and rebalance every 5 or 10 sessions.
What happened The training years loved it
On 2005–2016 the best variant had Sharpe 0.69 against 0.40 for SPY, with a 15.6% worst drop against 55.2%.
On the pre-registered 2017–2021 window, no variant passed: Sharpe 0.02–0.11 against SPY's 0.92, and about 0–1.5% a year against 18.3%.
Verdict Edge decayed
Costs were not the cause: with zero costs the best Sharpe was only 0.144. The family is parked. We did not re-tune on the validation years.
How things die
What is overfitting? Fitting a strategy so closely to the past that it mostly captures noise, then watching it fall apart on new data. See Investopedia’s explainer (external), or how we test.
The winner doesn't survive an honest count of how many things we tried, or it's no better than scrambled data.
6 write-upsBest Sharpe in a later test window0.77 vs SPY 0.921,192-combination finalist, 2017–2021. Most of these studies never opened later years.
The effect may be real but we couldn't have captured it with realistic trading rules.
1 write-upBest Sharpe in a later test window0.31 vs SPY 0.122022–2023, and only 1% invested. It failed 2017–2021 (−0.30).
It worked in one period and failed a later window we fixed in advance.
2 write-upsBest Sharpe in a later test window0.99 vs SPY 0.92Yield-curve alarm, 2017–2021. It then failed in 2022–2023 (−0.67).
The sign or result flipped between the windows we fixed in advance, so we don't treat it as a real effect.
1 write-upBest executable Sharpe in a later test window0.42 vs SPY 0.922017–2021. The new-vs-full sign flipped in 2022–2023.
It does something real, but misses its bar or the protection costs more return than we said we'd accept.
5 write-upsBest Sharpe in a later test window0.97 vs SPY 0.92Plain 200-day rule, 2017–2021. It still missed its own drop bar.
It cleared the pre-registered test it was written for, but did not beat simply holding the market on the same days. It has not been traded with real money.
1 write-upBest Sharpe in a later test window1.00 vs same-days SPY 1.042017–2021. Fully invested SPY was 0.92. Not traded live.
Not enough data to test it honestly. We say so instead of inventing a number.
1 write-upSharpe in a later test windowNone we can quoteAbout 21 years of daily ETF data, and only a handful of the cycle's calls fall in that span.
Recently buried
Overfit
144 signal tests, two winners, no edge
Overfit
1,192 indicator combinations: the winner was no better than luck
Edge decayed
Weekly reversal: great on 2005–2016, nothing on 2017–2021
Not tradable
Pre-Fed-announcement drift: real, but it happens overnight
Overfit
Wait for the stars to align: 24 setups, 0 passed