The graveyard
Every study here has its own page: the idea, what we tested, headline numbers, and why it died. Each cause of death is explained on how we test.
Didn't work
| Idea | Family | Cause of death | Headline |
|---|---|---|---|
| 144 signal tests | Diagnostics | Overfit | Two small "winners"; largest effect inside the noise (p = 0.22) |
| 1,192 indicator combinations | Models | Overfit | Best Sharpe 0.74; scrambled-data searches usually did better (p = 0.73) |
| Weekly reversal | Reversal | Edge decayed | Train Sharpe 0.69; validation Sharpe 0.02–0.11 vs SPY 0.92 |
| 32 trend variants | Trend | Overfit | Best Sharpe 0.70, drawdown 5%; overfitting probability 0.62 |
| Plain 200-day rule | Trend | Works but not worth it | Halved the drop in 2022–23; failed its bar in 2017–21 |
| Turn of the month | Calendar | Works but not worth it | Invested 12% of the time, 3.8% a year |
| Pre-Fed-announcement drift | Calendar | Not tradable | The gain was overnight, before our fill |
| Turn of month plus Fed days | Calendar | Works but not worth it | Failed both windows |
| Volatility-targeted SPY/VTI/VUG | Volatility | Works but not worth it | Sharpe 0.89 vs SPY 0.92 in 2017–21 |
| Yield-curve alarm | Macro | Edge decayed | 7% invested in 2022–23 |
| Moon phases | Esoteric | Edge decayed | New moon lost to full moon in 2017–21 |
| 24 "conditions align" setups | Setups | Overfit | 0 of 24 survived |
| 28 indicator families and pairs | Models | Overfit | Best Sharpe 0.62; deflated Sharpe 0.943 vs bar 0.95 |
| Machine-learning models on 24 indicators | Models | Overfit | No better than models trained on shuffled data |
| A take-or-skip judge | Models | Works but not worth it | Approved nearly every signal |
Passed, with heavy caveats
| Idea | Family | Label | Headline |
|---|---|---|---|
| The drawdown brake | Trend + volatility | Passed, unproven live | Passed 6 of 7 new market groups; gives up much of the return. Read the caveats first. |
Couldn't be tested
| Idea | Family | Why |
|---|---|---|
| The Benner cycle (1875) | Esoteric | About 21 years of daily data cannot test a chart that makes a handful of calls |
On the list
Not run yet: Wilder's Delta Phenomenon, as soon as we can code its rules as published. When it runs, it gets a page here whether it works or not.
Trial counts and rules are on each page. Everything is based on public data (see about).