~940 signals/year: how keel went from 2 trades per asset-year to a collectable evidence pipeline (Coinbase spot)

[eaitbrahim]

The number

~940 entry signals per year, pooled across 19 assets (~2.6/day). Six weeks ago the same engine, on the same venue — Coinbase spot, USD pairs, the only venue keel trades — produced 2.15 signals per asset per year.

The problem it fixes

keel's promotion gate demands n=100 trades per rule per product before paper evidence can argue for real money. At 2.15 signals/asset-yr that gate is 31–84 years away per asset. As the August status review put it: waiting is not a slower path; it is no path. The engine's honesty gates were fine; the clock was the defect.

How it was achieved — three moves, each measured

1. Trade the same rules on an hourly clock (issue #337). The daily turtle rules evaluated on ONE_HOUR bars fire 49.4 signals per asset-year (measured over 5 years of cached Coinbase candles; median n=268 per rule-product; range 48.5 LTC to 57.0 BTC) — a 23x increase with zero parameter changes. This shipped as a third deployment profile, paper-hourly: its own database, its own config, one cycle per UTC hour, kill-switch-gated.

2. Fix the warm-up so the confirmation series actually exists (issue #349, shipped in v0.9.1). The documented cache-warm step turned out to fetch only two of the profile's three candle series — keel fetch was ignoring config.market_data.granularities, so the FIFTEEN_MINUTE entry-gate confirmation series started with literally one bar per product. v0.9.1 makes every fetch path honor the config; the warm step now fills ~35,000 fifteen-minute bars per product.

3. Widen the universe 8 → 19, behind a measured health gate (issue #351, PR #358). Eleven Tier-2 assets (ZEC, NEAR, AVAX, UNI, FET, ICP, DOT, CRV, ALGO, BCH, DOGE) were admitted only after a 90-day FIFTEEN_MINUTE data-health screen on Coinbase: coverage >= 95.98%, zero zero-volume bars, worst gap 5 bars. They trade with target weights capped at 2% each (78/22 Tier-1/Tier-2 split) pending the spread-gate rail (#350).

49.4 signals/asset-yr x 19 assets = ~940/year. The n=100 pooled forward-trades review (#353) is now ~5-6 weeks of accrual away, not decades — it is scheduled for 2026-09-30.

The honest caveat (read this part)

The hourly configuration is measured NET-NEGATIVE at Coinbase's taker fees — 0 of 90 backtest cells profitable at every fee tier the venue offers (restated 2026-08-13 under the production-faithful engine). This pipeline exists to produce admissible forward evidence — rail vetoes, real outcomes, pending lifespans, intent divergence: the things a backtest cannot observe — not profit. A positive stretch would prove nothing: ~250 sequential trades inside one regime are not 250 independent draws. Daily-tuned parameters on an hourly clock is also, legitimately, a different strategy; forward evidence is the only thing that can settle it.

Status

The kill switch on the hourly profile was opened by the operator on 2026-08-18; the first 19-asset cycle ran clean (19 rules evaluated, correct no-signal verdicts, no errors). Progress tracks in milestone 14; the n=100 review lands in #353.