14 · Dynamic Exit Rules — Walk-Forward Backtest

Generated 2026-05-15. Raw results: /tmp/exit-rules-results.json. Source: scripts/backtest-exit-rules.mjs, helpers: src/lib/exit-rules.ts.

Verdict (honest)

NO exit rule combination beats the validation threshold of ΔSharpe_ann ≥ +0.20 AND max-DD reduction ≥ 30 % AND bootstrap p-value < 0.10 vs the T+90 fixed-hold baseline. The largest Sharpe improvement observed is ΔSR = +0.034 (time+insider+vol) — an order of magnitude below threshold and statistically indistinguishable from noise (p = 0.71).

STRATEGY_PROOF.exitRules was NOT updated. The live recommendations remain a pure entry signal with no documented exit playbook beyond T+90.

Setup

• Universe: 16 005 BUY-direction BacktestResult rows with full priceAtTrade → price90d coverage (2021-12 → 2026-04).
• Walk-forward: 24-month train / 12-month test, rolling. 2 folds (2023-12, 2024-12). σ priors computed on TRAIN only.
• Bucket σ: per (role × size) on TRAIN-window 90-day returns; fallback cascade role::size → role → __overall.
• Price path: reconstructed from {30, 60, 90, 160} day snapshots in BacktestResult. Sparse but no look-ahead. Linear interpolation for insider-sell event prices (rule 4).
• DD calc: trades binned into 16 chronological bins, averaged within bin, compounded — a realistic-portfolio proxy. Bin-level NAV.
• Bootstrap: 1000 iter, seed=42, two-sided pooled-resample p-value.

Rule definitions

#	Rule	Math
1	Stop-loss	exit at first t > 0 where (P_t − P_0)/P_0 ≤ −k·σ_bucket, k=2
2	Profit target	partial 50 % at +1σ, remainder at +2σ (else T+90 close)
3	Time-stop	exit at T+30 if `
4	Insider sell	exit at first t ≤ 90 with any new Cession* on company.id
5	Vol regime	exit if 30-day rolling annualized realized vol > 2·σ_ann (entry bucket)

Per-rule walk-forward results (vs baseline)

Baseline: SR_ann = −0.012, mean = 0.83 %, win = 46.9 %, max-DD = −17.89 %, hold = 90 d, n = 7 032.

Rule	SR_ann	ΔSR	Mean %	Win %	Max-DD %	DD red. %	Avg hold	p
stop	−0.036	−0.024	0.35	46.7	−22.21	−24.1	91.1 d	0.41
profit	−0.050	−0.038	0.31	47.4	−16.83	5.9	89 d	0.29
time	−0.012	0.000	0.62	45.9	−16.60	7.2	68.8 d	0.69
insider	−0.016	−0.004	0.57	47.2	−15.08	15.7	69.1 d	0.64
vol	+0.018	+0.030	1.60	46.8	−18.59	−3.9	91.4 d	0.26

Observations:

• Vol regime is the only single rule with positive ΔSharpe AND positive Δmean (+0.77 pp). p = 0.26 — not significant. Holds barely change (avg 91.4 d) → the rule fires on very few trades but those it exits are unusually bad.
• Time-stop cuts holding period to 68.8 d (free capital) with ΔSR = 0 — neutral on quality, marginally faster turnover. Plausible capital-efficiency play but not a return enhancer.
• Stop-loss underperforms baseline — the −2σ band crystallizes losses that mostly mean-revert by T+90. Confirms a well-known small-cap insider-buy pattern.
• Profit target suffers worst Sharpe drop (−0.038). +1σ is hit frequently then the partial leg locks in modest gains while the full +2σ is rarely reached → caps upside without trimming risk.
• Insider-sell reduces DD by 15.7 % but at the cost of ΔSR = −0.004 (insiders' SELL signal on FR mid-caps is noisy in this window).

Best combination

time+insider+vol:

• SR_ann = 0.022, ΔSR = +0.034
• Mean = 1.03 %, ΔMean = +0.20 pp
• Win = 45.9 %, ΔWin = −1.0 pp
• Max-DD = −10.29 %, DD reduction = +42.5 % vs baseline ✓
• Avg hold = 55.1 d (39 % faster capital recycling)
• p = 0.71 — not statistically significant ✗

This combination DOES clear the DD-reduction gate (≥30 %) but fails BOTH the ΔSharpe gate (+0.034 < +0.20) and significance (p = 0.71 ≫ 0.10). DD reduction is partly mechanical: shorter holding period → smaller bin returns → smaller compounded drawdowns. We refuse to claim it as alpha.

Sensitivity: stop-loss multiplier

k (σ mult.)	Mean %	SR_ann	Win %	Max-DD %	Avg hold
1.5	−0.29	−0.068	45.7	−26.69	92.3 d
2.0	0.35	−0.036	46.7	−22.21	91.1 d
2.5	0.62	−0.023	46.8	−19.51	90.5 d
3.0	0.69	−0.019	46.8	−18.99	90.1 d

Monotone in k: tighter stop, worse outcome. The −2σ band is below the optimum (the optimum is "no stop at all" on this universe). This argues against any tight stop on insider-buy small/mid-cap signals.

Open questions / next steps

1. Sparse price path: rules can only fire at 4 checkpoints (30, 60, 90, 160 d). A daily price store would let stop-loss fire intraday and is the single largest possible improvement. Hydrating a PriceHistory model from Yahoo daily closes is on the roadmap.
2. Bucket σ at entry-time only: we use TRAIN-window σ_90d; a more sophisticated overlay would use trailing realized vol per company.
3. Regime conditioning: rules may work in bear regimes only. Conditional walk-forward by RecoSnapshot.regime is a logical next pass.
4. Profit-target asymmetry: a one-sided variant (full @ +2σ only, no partial) would avoid clipping the tail. Worth a follow-up.
5. Insider-sell stratified: filter cessions to ≥1 % of market-cap and to dirigeant (vs board) roles — the 6 810-event index is dominated by small administrative cessions which dilute the signal.

Reproducibility

node --env-file=.env.local scripts/backtest-exit-rules.mjs

Seed = 42 throughout. Output: /tmp/exit-rules-results.json.

Files

• src/lib/exit-rules.ts — 5 pure helpers + earliestExit combiner.
• scripts/backtest-exit-rules.mjs — walk-forward harness, JS port of the TS kernels (kept logically identical).