Bot Fitting Backtest

This folder contains parameter sweep backtests that simulate grid fills for the four AMA winners from analysis/ama_fitting.

What it optimizes

For each of the 4 AMA strategies, it searches for best:

• spread (% target spread for the gapSlots spread zone)
• increment (% geometric rail step)
• max/min ratio (symmetric range around AMA, e.g. 2.0 means [AMA/2, AMA*2])

Scripts

• backtest_bot_fitting.ts — lightweight sweep across spread / increment / ratio with unit-size percentage-point accounting and risk scoring
• backtest_ama_sweep.ts — persistent grid simulation with capital-weighted sizing, weight profiles, and worker-thread parallelization
• backtest_lookback_drawdown.ts — paired Huber-lookback comparison of drawdown / net capture / reset churn (geometry fixed); see docs/AMA_SLOPE_WINDOW.md
• shared_utils.ts — shared helpers (argument parsing, data loading, formatting)

Both simulators share the same production grid model (ported once in backtest_bot_fitting.ts and imported by the sweep), so identical params build byte-identical grids in both tools.

Production alignment

Both simulators port the live bot lifecycle end to end:

• Grid geometry — createOrderGrid (modules/order/grid.ts): master rail at √(1±inc) offsets bounded by [center/ratio, center×ratio] with a calculateGapSlots spread zone centered on the placement price; order prices are placement-time constants.
• Slot rotation (anchor-&-refill) — a filled buy at rail node k instantly re-offers its base at the adjacent master-rail node k+1; when that refill sells, exactly one rail hop minus the round-trip fee is booked and the freed quote re-bids node k−1. A slot keeps cycling on small oscillations between resets — this is what a refilled live slot earns, not a cross-gap differential. Re-armed orders carry a one-bar replacement-lag cooldown. Sub-note: because rails chain by ×(1+inc) above center and ×(1−inc) below, a down-chain hop equals inc/(1−inc) (e.g. 1.0101% at inc=1%), not exactly inc — matching real rail adjacency.
• Inventory-funded selling — unlinked (initial-grid) sells execute only against held base, priced against the weighted-average entry of the position; unfundable sells stay open and retry (never shorting).
• Inventory carry — resets cancel ALL live orders (initial leftovers plus armed refills/rebids); bought-and-held base merges into a weighted-average- entry pool that survives resets (production resync never market-sells). The end-of-run inventory mark is reported as an INFORMATIONAL field — it is real carried risk but excluded from scoring/ranking.
• Reset triggers (MARKET_ADAPTER):
  • Trigger A: AMA drift ≥ AMA_DELTA_THRESHOLD_PERCENT (1%) from the recorded center, ratchet semantics — always on.
  • Trigger B: slope-delta reset (|slope − baseline| ≥ AMA_SLOPE_DELTA_THRESHOLD_PERCENT/100 × DYNAMIC_WEIGHT_AMA_MAX_SLOPE_PCT over the DYNAMIC_WEIGHT_AMA_LOOKBACK_BARS average-slope series, baseline re-seeded every reset) plus the slope-ratio grid price offset (direction × min(|slope|/maxSlopePct,1) × targetSpread/2, applied to the placement center). Both are gated behind --asymmetric-bounds, mirroring the production per-bot asymmetricBounds whitelist — default OFF for a typical non-whitelisted bot.
• BTS operation fees — every order placement (initial grid + armed refills + rebids) pays a maker create fee, every reset cancel pays a cancel fee; totals are deducted from net capture (bot_fitting converts against --bts-fee-capital, the sweep converts via --tx-fee-price in capital units).

Documented remaining deltas (inherent to OHLC research sims): bar-granularity triggers, all-or-nothing fills on raw hi/lo touch, unit/capital sizing instead of live balance + dynamic-weight sizing, pool-price fills without microstructure, and no consolidation/dust-cancel/COW/collision mechanics.

Input dependencies

Both scripts require:

• LP candles JSON (recommended 1h), from market_adapter/data/lp/ — pass via --data
• AMA optimization winners JSON, from analysis/ama_fitting/ — for backtest_ama_sweep.ts, pass via --results

Run

# Lightweight sweep
node dist/analysis/bot_fitting/backtest_bot_fitting.js \
  --data <path-to-lp-candles.json>

# Persistent grid simulation with AMA winners
node dist/analysis/bot_fitting/backtest_ama_sweep.js \
  --data <path-to-lp-candles.json> \
  --results <path-to-optimization-results.json>

# Paired lookback drawdown comparison (no --results needed)
node dist/analysis/bot_fitting/backtest_lookback_drawdown.js \
  --data market_adapter/data/lp/<market-pair> \
  --lookbacks 12,14,16,20

backtest_ama_sweep.ts accepts --lookback <bars> to override the slope window (DYNAMIC_WEIGHT_AMA_LOOKBACK_BARS); backtest_lookback_drawdown.ts builds on it and reports the paired per-lookback deltas.

backtest_bot_fitting.ts auto-derives --results from the --data filename (analysis/ama_fitting/optimization_results_<base>_w<λ1>_<λ2>_<λ3>_<λ4>.json, newest _w* file matching the base name; falls back to the legacy optimization_results_<base>.json) when omitted. backtest_ama_sweep.ts requires --results explicitly.

Optional tuning (values shown are examples, not defaults):

node dist/analysis/bot_fitting/backtest_bot_fitting.js \
  --data <path-to-lp-candles.json> \
  --spread 0.4:1.6:0.1 \
  --increment 0.2:0.8:0.1 \
  --ratio 1.5,2,2.5,3,4,5,8,10 \
  --active-orders 5 \
  --fee 0.20 \
  --min-spread-factor 2.1 \
  --asymmetric-bounds \
  --risk-duration 1.0 \
  --risk-peak-open 2.0 \
  --risk-imbalance 1.2 \
  --risk-cancel 0.15

Default --ratio is 1.5,1.75,2,2.5,3,4,5,8,10 (includes 1.75). Default --active-orders sizes EVERY rail slot in bounds (production behavior); pass a number to cap the search to the slots nearest the gap.

Output

Results are written to analysis/results/ (filenames derived from the input data file: bot_fitting_results_<base>.json and ama_sweep_results_<base>.json).

The console also prints best parameter set per AMA with matched pairs (cycles), fill efficiency, net capture and score.

Notes

• This is an offline simulation proxy, not a full chain execution model.
• Hop granularity: the sim arms each refill at the ADJACENT master-rail node (one increment). Production derives leg spacing from the gap-aware boundary crawl (deriveTargetBoundary: one slot per non-partial fill, burst-capped), so for wide-gap configs live legs may span gapSlots rail steps (~targetSpread) instead of one. Calibrate against real bot fills before trusting absolute magnitudes; rankings remain internally consistent.
• --min-spread-factor 2.1 enforces spread >= 2.1 x increment (GRID_LIMITS.MIN_SPREAD_FACTOR), matching calculateGapSlots.
• Score used for ranking (bot_fitting):
  • totalNetCaptureAfterFeesPct = realized per-rotation gross (one rail hop or weighted-average-entry sale) − round-trip fee − BTS op fees; the end-of-run inventory mark is NOT included
  • baseScore = totalNetCaptureAfterFeesPct * (fillEfficiency / 100)
  • riskPenalty = avgOpenDurationBars*1.0 + peakOpenOrders*2.0 + avgImbalance*1.2 + canceledOnReposition*0.15
  • finalScore = baseScore - riskPenalty
• Score used for ranking (sweep): realized profit only —
  • netProfitPerCapital * 100 * log10(max(1, matchedPairs)) - maxDrawdownPct * 0.5 (maxDrawdown tracks REALIZED equity swings; carried-bag risk stays visible through finalInventoryUnits / finalInventoryMarkUnits)

Persistent Grid Simulation Details

backtest_ama_sweep.ts models the real bot mechanics:

• Orders sit at FIXED chain prices until canceled or filled
• Slot rotation: a filled buy re-offers the adjacent rail node above; its fill books one rail hop net of fees and re-bids the node below (anchor-&-refill)
• Unlinked sells execute only against held inventory at weighted-average entry; bought-and-held base carries across resets and is marked to close informationally (excluded from profit; drawdown also tracks realized equity only — bag risk stays visible through the info fields)
• When triggers fire (A drift, B slope under --asymmetric-bounds), the grid re-centers and all live orders cancel
• Order sizing depends on capital, ratio (range width), and weight profile
• Three weight profiles: valley, neutral, mountain (symmetric buy/sell)

Search grid defaults — centered around bot defaults (spread=2%, increment=0.5%):

All parameters above are tunable via CLI flags. Additional tuning flags:

The sweep parallelizes across AMAs using worker threads (one per AMA strategy). Use --help for the complete option list.

node dist/analysis/bot_fitting/backtest_ama_sweep.js \
  --data <path-to-lp-candles.json> \
  --results <path-to-optimization-results.json> \
  --spread 4:16:1 --increment 0.5:4:0.25