Experiment: 5-year backtest of the hourly Turtle for a per-asset sample (N > 100), at the fees actually paid

[eaitbrahim]

The daily Turtle yields about 2 trades per product-year, so no per-asset sample of 100 exists even in 5 years of history. Hourly bars fire far more often: #427's reconstruction measured about 49 signals per asset-year.

Plan

Run keel simulate --years 5 over the hourly paper rule set (config.paper-hourly.yaml and the hourly rule rows) on a copy of the database, and report N, win rate and R-expectancy per product and pooled.

Fee guardrail

Fees must be the ones actually paid, not a hopeful tier:

• The simulator charges 1.20% taker per leg plus per-product slippage (5 to 101 bp).
• The forward trades on this deployment recorded at least 1.136% per leg (pooled-review preview, 2026-09-26).
• A 0.40–0.60% per-leg run is reported only as a labelled sensitivity, never as the headline. Higher-frequency breakouts pay fees and whipsaw on every round trip, and understating fees is the flattering direction.

Depends on the R-multiple fix, so the pooled expectancy is meaningful.

[eaitbrahim]

Result: 5-year backtest of the hourly Turtle, 2026-09-26

Setup: main at 438e78f, which includes the R-multiple fix (#826) and the DCA fix (#825). The rules are the 19 hourly turtle_breakout rows from keel-paperhourly.db, with --years 5 --no-fetch. Both runs used identical cached candles (2021-07-18 to 2026-09-26), on copies of the database, with --no-trial-record. Nothing in ~/keel changed.

	1.20% taker (headline, the fee actually paid)	0.60% taker (sensitivity only)
pooled N	4,871	4,871
per-asset N	235–289 (PAXG 78, shorter history)	same
pooled win rate	15.1%	18.0%
pooled expectancy	−1.178 R	−0.696 R
avg win / avg loss	+2.62 R / −1.85 R	+2.64 R / −1.43 R
profit factor (R)	0.25	0.41
longest losing streak (pooled, chronological)	70	57
assets with positive expectancy	0 of 19	0 of 19
best asset	SOL −0.80 R	SOL −0.33 R
account (63 trades, capital- and allowance-limited)	−2.3% vs DCA benchmark +80.2%	−1.1% vs +81.3%

Slippage is per product (5–101 bp per leg) in both runs.

What it says

• The per-asset sample is now well past 100, and the verdict doesn't depend on the fee assumption. Every asset loses at both fee levels. Halving the fee is worth about 0.48 R per trade and still leaves the pool at −0.70 R.
• The average loss is −1.85 R, not −1 R: fees and slippage on a tight hourly stop are a large fraction of the risk. Avg MAE is about 1.0 R and avg MFE about 1.46 R, so trades rarely run far enough to make up for it.
• Straight-line extrapolation of the two fee points to 0% still gives about −0.2 R before slippage. There is no sign of a gross edge that fees are merely eating.
• This agrees with the forward pool (24 trades, 12.5% win rate against a 32.7% break-even) and with the live/paper significance results already recorded in the config comments.

Full reports are kept locally for reference. They aren't committed.

[eaitbrahim]

Correction (appended, not edited): the zero-fee figure above ("about −0.2 R") was a straight-line extrapolation from the two fee arms, not a measurement. The measured 0% arm gives −0.215 R pooled, with 3 of 19 products positive: BTC +0.011, ETH +0.071, SOL +0.136. Those are the three with 5–11 bp slippage; every other product loses before any fee. This also reconciles with 2026-08-12's "profitable at zero fee", which was priced at a flat 5 bp. Full record and per-trade data: docs/experiments/2026-09-26-hourly-turtle-5year-backtest.md.