the cortex
the whole board, failures included.
Six hundred and thirty momentum configurations were searched against 36 Coinbase products over 2,799 days, 6 of them delisted and kept in the tape. 34 had a positive out-of-sample Sharpe. One beat buying and holding BTC. The in-sample column sits next to the out-of-sample column on purpose: the distance between them is the single most useful thing on this page.
configurations searched
630
positive out-of-sample
34
5.4%
beat buy-and-hold
1
benchmark +10%
PBO across the grid
0.60
reject above 0.50
| agent | status | horizon | OOS sharpe | IS sharpe | IS deflated | PBO | OOS ret | vs BTC | max DD |
|---|---|---|---|---|---|---|---|---|---|
| VESPER-9spot · 36 products · coinbase | probation | 10d | 0.35 | 0.43 | 0.00 | 0.45 | +12.4% | beat | -37.3% |
| KESTREL-2spot · 36 products · coinbase | probation | 5d | 0.28 | 0.50 | 0.00 | 0.45 | +8.0% | lost | -39.3% |
| MAGPIE-7spot · 36 products · coinbase | rejected | 20d | 0.25 | -0.03 | 0.00 | 0.45 | +6.5% | lost | -35.1% |
| ANNEAL-4spot · 36 products · coinbase | rejected | 20d | 0.22 | -0.37 | 0.00 | 0.50 | +3.8% | lost | -15.2% |
| SIEVE-1spot · 36 products · coinbase | rejected | 20d | 0.19 | 0.06 | 0.00 | 0.45 | +3.2% | lost | -37.1% |
| PARALLAX-3spot · 36 products · coinbase | rejected | 20d | 0.11 | 0.47 | 0.00 | 0.50 | −3.2% | lost | -44.0% |
| CINDER-5spot · 36 products · coinbase | rejected | 20d | 0.10 | 0.79 | 0.00 | 0.50 | −1.7% | lost | -38.2% |
| DRIFTER-6spot · 36 products · coinbase | rejected | 10d | -0.33 | 0.48 | 0.00 | 0.50 | −22.9% | lost | -45.4% |
| LATTICE-8spot · 36 products · coinbase | rejected | 3d | -0.96 | 0.86 | 0.00 | 0.35 | −47.8% | lost | -59.6% |
| VERGE-0spot · 36 products · coinbase | rejected | 1d | -3.69 | -2.85 | 0.00 | 0.65 | −73.8% | lost | -76.5% |
IS deflated is the in-sample Sharpe after the 630 configurations searched are charged against it. It is 0.00 for every row on this board, which is the honest reading of what the search found. vs BTC compares each out-of-sample return to buying and holding Bitcoin over the same window (+10%). Rejected specimens are kept here rather than deleted: a leaderboard that lists only winners is the survivorship bias it claims to have removed.
the honest result.
Of 630 configurations, the median out-of-sample Sharpe is -0.84. PBO across the grid is 0.60 — on three fifths of the cross-validation splits, the configuration that won in sample ranked below the median out of it. Every in-sample Sharpe here deflates to zero once the size of the search is charged against it.
That is the finding, and we are publishing it rather than searching until something wins. Searching until something wins is the thing this entire site is about not doing.
why rejected agents stay.
The specimens at the bottom of this board were selected on purpose: the configuration with the highest in-sample Sharpe that landed in the bottom half out of sample, the one that is positive before costs and negative after them, and the one whose result collapses when the delisted products are taken out of the universe. Deleting them would leave a page showing only winners, drawn from a process that produced almost entirely losers.
gate 02 · survivorship