AMA-Slope Window Decision — 16h + Persistence Gate

Status: applied (window 20 → 16, persistence gate enabled by default). Related: docs/GRID_RECALCULATION.md §4, market_adapter/README.md, analysis/trend_detection/README.md, analysis/tradingview/README.md.

Decision

• MARKET_ADAPTER.DYNAMIC_WEIGHT_AMA_LOOKBACK_BARS: 20 → 16 (Huber window).
• MARKET_ADAPTER.AMA_SLOPE_PERSIST_ENABLED: false → true; gate length AMA_SLOPE_PERSIST_BARS = 3.

No estimator change. The canonical slope stays computeHuberWindowSlopePct (Huber-robust regression of ln(AMA)).

Why

The window and the gate are coupled. A shorter window is more responsive but noisier; the gate removes the noise-driven resets while adding at most K−1 bars to the surviving slope resets (the independent price/Drift trigger is unaffected), so together they dominate either knob alone.

Measured on a live 1h market-pair pool (AMA3, fixed centred reference half-window 24 bars, via backtest_ama_slope_huber.ts --truth-window 24):

16h + K=3 vs the old 20h-ungated:

• Churn collapses: resets 1467 → 962 (~35% fewer), resets/day 1.45 → 0.95, whipsaw 52% → 17%.
• Lag improves: AMA group delay 10 → 8 bars; reversal-confirmation lag 22 → 20.
• Range tilt points wrong less often: 13.8% → 12.6% of forward windows.
• Cost: slope wobble +8%, zero-crossings +7%. That is the noise the gate absorbs at the decision layer (it never changes the slope value itself).

16h is the knee; below it, returns diminish (14h buys 1 more bar of lag for +1.6% resets and +4% wobble).

Reproduce:

node dist/analysis/trend_detection/backtest_ama_slope_huber.js \
  --data market_adapter/data/lp/<market-pair> \
  --lookback 20:8:2 --truth-window 24   # add --slope-persist 1 for the ungated row

amaLag is robust to the reference half-window; revLag scales with it, so compare revLag across runs only at a fixed --truth-window. The LP shards are refreshed by live collection, so reset counts can drift by ±1 from the table.

Huber scale estimate — why C stays 1.345

The robust scale is a plug-in 1.4826 * MAD of the fit's own residuals; with two fitted parameters those are shrunk, so it reads low by ≈ 1/(bars − 1) and the effective Huber constant is ~1.25 at 16 bars (nominal 1.345). Compensating it (a sqrt(n/(n-2)) correction, or a proposal-2 M-scale) is decision-neutral — wobble, lag, reset counts and wrong-way all within noise — and gives up more robustness (~0.8% RMS under 10% contamination) than it recovers in efficiency (~0.4% clean). So the detune is left in place: it makes short windows more conservative, the safe direction for a trend filter. The outlier diagnostic is the only metric that moves.

node dist/analysis/trend_detection/backtest_ama_slope_huber.js \
  --data market_adapter/data/lp/<market-pair> --scale-mode none|df|mscale

(none == production; see analysis/trend_detection/huber_scale_variants.ts.)

Fill-model drawdown — the window effect is pool-dependent

Paired persistent-grid runs (simulatePersistentGrid, geometry fixed, asymmetricBounds on) show the window does move fill-model economics, but with no stable sign across pools: on a long-lived liquid pair 16h beat 12h on realized drawdown (−8.9 pts, lower in 108/108 geometries) and net capture (+26 pts) at equal activity, while on a shorter pair the ordering reversed. So it supports keeping the shipped 16h without being a general economic proof. Trust the paired delta, not the absolute level (realized-equity DD is model-shaped and can exceed capital).

node dist/analysis/bot_fitting/backtest_lookback_drawdown.js \
  --data market_adapter/data/lp/<market-pair> --lookbacks 12,16

Rejected alternative — Kalman slope estimator

A constant-velocity Kalman filter on ln(AMA) was implemented, unit-tested (bar-for-bar parity with the live KalmanFilter class) and benchmarked. It cut lag ~2 bars at matched churn, but:

• it required a stateful estimator, chart re-embedding, adapter restructuring, a persisted estimator id + baseline re-seed, and a flag rollout — a real migration/correctness surface for a latency-only gain;
• the fill model could not turn the lag win into a profit win (estimator differences within model noise, marginally favouring the incumbent);
• range wrong-way was estimator-independent (~13.5% for both).

So it was removed completely in favour of the zero-migration window change.

Interactions updated

• modules/constants.ts — the two defaults + comments.
• profiles/general.settings.json — the local (gitignored) override set to 16 so the running config agrees with the code default.
• market_adapter/core/strategies/dynamic_weight_series.ts — comments.
• analysis/tradingview/grid_reset_config.ts — resolves slopePersistBars (per-bot override → global enable/value → legacy 1) and exports it in toGridSimPayload, so the embedded chart replay fires the same Δs resets the live adapter does. The panel prints persist K.
• analysis/tradingview/tradingview_uplot_chart_generator.ts — spreads the gate into simulateGridResetSeries (via gridSimCfg) and shows it in the panel.
• analysis/tradingview/grid_reset_sim.ts — the persistence gate itself.
• market_adapter/core/market_adapter_service.ts — adapter gate (resolveAmaSlopePersistBars / advanceAmaSlopePersistence, state persisted in botState.amaSlopePersistCount / amaSlopePersistDir).
• analysis/bot_fitting/backtest_bot_fitting.ts — gate + bandTilt fill-model hooks (used by the comparison harness).
• Docs: docs/GRID_RECALCULATION.md, market_adapter/README.md, analysis/trend_detection/README.md, analysis/tradingview/README.md.
• Research (new): analysis/trend_detection/huber_scale_variants.ts and analysis/bot_fitting/backtest_lookback_drawdown.ts, with --scale-mode on backtest_ama_slope_huber.ts, --lookback on backtest_ama_sweep.ts, and tests/test_huber_scale_variants.ts.

Operational notes

• Live behaviour change. Both edits are active by default. The window change moves readyBars = erPeriod + lookbackBars by −4 bars (faster post-restart convergence).
• Restart transient. The gate counters persist across restarts; a pending count is cleared by any successful reset.
• Baseline transient. The persisted gridRangeScalingAmaSlope baseline was written by the 20h window; the first post-deploy cycle compares a 16h reading against it. Magnitudes are close (the delta is usually below the reset gate), but a bot whose slope sits near the threshold can cost one extra recenter, once — the same class of effect as the Huber swap, with no version marker to suppress it.
• Rollback. Window: set DYNAMIC_WEIGHT_AMA_LOOKBACK_BARS back to 20. Gate: AMA_SLOPE_PERSIST_ENABLED = false (or amaSlope.persistEnabled: false per bot). No schema change.
• Caveat. Proxy-scored. The fill model lacks queue-position loss and adverse selection, so the economic benefit is not conclusively demonstrated; validate in shadow mode before relying on it. The churn/lag/wrong-way wins are robust across six markets. The fill-model drawdown comparison (above) found a real per-pool window effect but with a pool-dependent sign, so it does not convert this into a general economic proof.