Experiment: trailing vs static exits

Fee note (added 2026-09-27, #836). This record describes a fee-free regime inside a venue’s monthly allowance. For keel’s orders that regime is hypothetical: they go through Coinbase Advanced Trade, which has no fee-waiver tier. On 2026-09-27 the account reported the Intro tier (0.9% taker, 0.5% maker, has_promo_fee: false), and every live fill had paid a fee; Coinbase One’s zero trading fees apply on the main Coinbase platform, not Advanced Trade. Figures priced at the taker fee are the realistic ones; figures priced at zero fee describe a regime keel cannot reach. Nothing on this page has been rewritten.

Cost note (added 2026-09-02). The figures below are priced at the flat 5bp slippage floor. the per-product restatement later measured that no asset in keel’s universe reaches that floor — the range is 1.1× to 36.8× — so every profit factor here is optimistic by roughly 0.09 at the median. The verdict is unaffected: the correction only ever moves a number down, and every result here was already negative. Nothing on this page has been rewritten; records are appended to, not revised.

Issue #442. Driver: 2026-08-22-trailing-vs-static-exits.py (pre-registered in its docstring before the run). Artifact: every number below is a row of 2026-08-22-trailing-vs-static-exits.jsonl — 240 cells (2 families x 30 products x 4 arms) plus 8 pooled rows.

Headline, stated first because it is the point: the pre-registered expectation held — trailing and break-even rolls do NOT improve either family at the 120 bp taker fee, and the trailing arm makes both families clearly worse. The constructor defaults therefore stay OFF (trail_atr_mult=None, be_roll_rr=None): the capability is wired and measured, and the measurement says not to turn it on.

Why this run exists

#442 found executor.trail_stop_atr / roll_to_break_even / scale_out implemented, unit-tested, and never called. The PR that ships alongside this record wires the exit POLICY (ratchet-only ATR trailing + break-even roll) into the engines that drive exits per bar (strategy.backtest, sim.portfolio_sim, via strategy/exit_policy.py and the per-family trail_atr_mult/be_roll_rr params), with turtle deliberately excluded. This run measures what the wiring changes, before any default is flipped.

Method (frozen in the driver’s docstring before the run)

• Same engine as the #475 significance reconstruction: keel.strategy.backtest.backtest over every product’s full cached ONE_HOUR history (2021-07-18 to 2026-08-21 UTC; 30 products with

  = 2,000 hourly bars, 5,902 to 44,623 bars each), next-bar-open market fills, 5 bp per-leg slippage, and the 120 bp taker fee per leg (backtest.TAKER_FEE_PCT) — the rate outside the fee-free allowance, because the question is whether exit management changes the family’s real-cost economics.

• Four pre-registered arms, constructor defaults for everything except the exit knobs: static (no knobs — the shipped default and the pre-#442 behavior), trail_1_5 (trail_atr_mult=1.5 — the live primitive’s own default multiplier), be_1 (be_roll_rr=1), trail_1_5_be_1 (both).
• turtle_breakout is out of scope BY DESIGN: its real exit is the Donchian channel and the family carries no exit-policy knobs (#442 hypothesis 3) — there is no trailing arm to measure.
• Read-only against the deployment cache (file:...?mode=ro); the only writes are the JSONL artifact and stdout.

Results — pooled per (family, arm), the pre-registered question

family	arm	n (pooled)	win rate	expectancy (1-unit notional)	profit factor	avg R
pullback_continuation	static	3,261	0.1435	−68.00	0.0075	−2.438
pullback_continuation	trail_1_5	3,697	0.0933	−61.89	0.0017	−2.195
pullback_continuation	be_1	3,303	0.1150	−69.19	0.0016	−2.424
pullback_continuation	trail_1_5_be_1	3,704	0.0886	−62.03	0.0016	−2.196
rsi_meanrev	static	1,201	0.2714	−108.89	0.0776	−1.394
rsi_meanrev	trail_1_5	1,213	0.1665	−111.40	0.0197	−1.377
rsi_meanrev	be_1	1,207	0.2030	−118.30	0.0336	−1.409
rsi_meanrev	trail_1_5_be_1	1,215	0.1597	−110.33	0.0198	−1.375

Wall clock 5,189 s for 240 cells. Per-product direction (arm expectancy vs static expectancy, 30 products each cell):

family	arm	products better / worse / same
pullback_continuation	trail_1_5	17 / 13 / 0
pullback_continuation	be_1	16 / 13 / 1
pullback_continuation	trail_1_5_be_1	16 / 14 / 0
rsi_meanrev	trail_1_5	14 / 16 / 0
rsi_meanrev	be_1	13 / 15 / 2
rsi_meanrev	trail_1_5_be_1	13 / 17 / 0

What the numbers say

• Every arm of every family is deeply negative at 120 bp/leg — the same wall #475 hit; exit management does not move it. This run was never going to find an edge that survives a fee the families do not survive, and it does not.
• The trailing arm is clearly worse, not differently-bad. Win rate collapses (pullback 14.4% → 9.3%; rsi 27.1% → 16.7%) and profit factor drops ~4x in both families: the trail cuts winners before the fixed-R target they were sized to reach — exactly the pre-registered mechanism, and exactly why turtle (whose entire payoff is the long tail) was excluded by design.
• rsi_meanrev’s expectancy is strictly worse in every arm (−108.89 → −110.33..−118.30): a mean-reverter’s edge is the quick full move; every form of early exit gives back part of it and pays the same round-trip fee.
• pullback_continuation’s expectancy moves −68.00 → −61.89..−69.19 — a less-bad loss on the same sign, not an improvement. The trailing arms’ n grows 3,261 → 3,697 (+436 trades, +13%): earlier exits free the slot sooner, each re-entry paying the full round-trip friction again. The trail also halves what little profit factor the static arm retains.
• The per-product counts are a coin flip (best cell 17/30 better, worst 13/30) — there is no consistent per-product direction to appeal to either.

Decision the evidence supports

• Constructor defaults stay OFF for both families (trail_atr_mult=None, be_roll_rr=None). The knobs remain available for research sweeps and for any future rule row that wants to carry them, but nothing shipped turns them on: at the fee actually paid, the wiring’s measured effect is negative-to-neutral, exactly the direction the pre-registration predicted and for the reasons it predicted.
• The wiring itself is NOT reverted — the capability was the point of #442, and a default-off policy whose knobs measurably do what they say (ratchet, exit earlier) is now tested behavior rather than a dormant claim.
• Live stop management remains issue #502 (the broker port has no bracket/OCO OrderSpec kind), independent of this result.
• A stated limitation: the paper trader (strategy/paper.py) is NOT wired. It is signal-driven — it holds no Rule instance to read params from, and its on_candle receives one candle at a time, so the trailing ATR has no series to compute from. With the defaults OFF no paper row can carry the knobs into a divergence today; a paper row that sets them would backtest with management and paper-trade without it, which the knobs’ PARAM_DOCS state outright (“Sim/backtest engines only”). Paper should follow #502.