AMA Fitting

Tooling to fit AMA parameters (ER, Fast, Slow) against real LP pool candle data. The optimizer writes result JSON by default and can explicitly export the chosen parameters into profiles/market_profiles.json for the market adapter.

Workflow Overview

1. fetch_lp_candles.ts   →   market_adapter/data/lp/<pairFolder>/lp_pool_<poolShort>_<interval>.json
2. optimizer_high_resolution.ts   →   optimization_results_*.json
                                   →   profiles/market_profiles.json  (only with --write-profiles)
3. scripts/generate_lp_chart.ts   →   market chart + comparison chart  (visual review)
   - LP chart: `npm run lp:chart -- --data <lp-export.json>`
   - Local LP comparison alias: `npm run ama:chart:lp-local -- --data <lp-export.json>`

Step 1 — Fetch LP Candles

fetch_lp_candles.ts fetches bidirectional LP swap data from Kibana and saves a full uncut candle file. Uses the same kibana_source as the market adapter bootstrap (gaps filled via candle_utils.fillCandleGaps), but without pruning.

Known asset details:

<ASSET_A>/<ASSET_B> pool (3 years):

node dist/analysis/ama_fitting/fetch_lp_candles.js \
  --pool 1.19.133 \
  --assetA <ASSET_A> --assetAId <asset_a_id> --assetAPrecision <n> \
  --assetB <ASSET_B>     --assetBId <asset_b_id>    --assetBPrecision <n> \
  --hours 26280

Output: market_adapter/data/lp/<pair_folder>/lp_pool_<poolShort>_<interval>.json (interval-dependent)

Options:

Step 2 — Run the Optimizer

optimizer_high_resolution.ts runs a parallel geometric grid search over ER × Fast × Slow combinations. Produces four AMA winners (AMA1–AMA4) using different distance-cap quantiles and writes results to a JSON file + auto-generates an interactive HTML chart.

By default this does not update runtime market-adapter profiles. Add --write-profiles when you intentionally want the fitted parameters exported to profiles/market_profiles.json.

Run on the fetched LP data:

node dist/analysis/ama_fitting/optimizer_high_resolution.js \
  --data market_adapter/data/lp/<pair>/lp_pool_<id>_<interval>.json

Export winners to the market adapter profile file:

node dist/analysis/ama_fitting/optimizer_high_resolution.js \
  --data market_adapter/data/lp/<pair>/lp_pool_<id>_<interval>.json \
  --write-profiles

Override ranges via CLI:

node dist/analysis/ama_fitting/optimizer_high_resolution.js \
  --data market_adapter/data/lp/<pair>/lp_pool_<id>_<interval>.json \
  --erMin 100 --erMax 600 \
  --slowMin 800 --slowMax 6000

AMA fitting caps:

These are the active defaults used by the optimizer:

--ama1Cap 0.25  --ama2Cap 0.30  --ama3Cap 0.35  --ama4Cap 0.40

AMA distance weights:

The distance penalty λ balances movement smoothness against price closeness (higher λ = AMA must stay tighter to price). Each AMA has a default weight; override any individually:

# Defaults (built-in)
--ama1Weight 0.0031  --ama2Weight 0.0025  --ama3Weight 0.00185  --ama4Weight 0.0013

# Override only AMA1 and AMA4, keeping AMA2/AMA3 defaults
node dist/analysis/ama_fitting/optimizer_high_resolution.js \
  --data market_adapter/data/lp/<pair>/lp_pool_<id>_<interval>.json \
  --ama1Weight 0.003 --ama4Weight 0.002

Inventory price range guidance:

Use an inventory range that sits above the fitted cap so the market maker has room to absorb normal noise without widening the book too much.

An optimized AMA plus the recommended buffer table is intended to provide a relatively safe operating range for extreme market conditions while still preserving reasonable inventory turnover.

• Safe buffer: +10% to +15% above the fitted cap
• Borderline: +20%
• Overkill: +25%+

Source: pool 133 IOB.XRP/BTS, 1h candles, 2023-04-12 00:00 UTC through 2026-04-11 23:00 UTC (the working-tree dataset analysis/ama_fitting/data/lp_pool_133_iob.xrp_bts_1h.json, regenerated by a local fetch_lp_candles.ts run).

Outputs:

• analysis/ama_fitting/optimization_results_<datafile>_w<λ1>_<λ2>_<λ3>_<λ4>.json — full results (e.g. _w0_0031_0_0025_0_00185_0_0013)
• analysis/charts/optimization_chart_<datafile>_w<λ1>_<λ2>_<λ3>_<λ4>.html — interactive AMA overlay chart (auto-generated)
• profiles/market_profiles.json — updated with new AMA parameters per pair only when --write-profiles is used

Boundary check: If a winner lands on the edge of the search range, the optimizer warns you. Widen the affected range and re-run.

Step 3 — Visual Review

A chart is auto-generated during Step 2 at analysis/charts/optimization_chart_<datafile>_w<λ1>_<λ2>_<λ3>_<λ4>.html. Open it in a browser to compare all four optimized AMA overlays against the candlestick price.

For a standalone chart without re-running the optimizer (uses current defaults from modules/constants.ts rather than optimized results):

npm run lp:chart -- \
  --data market_adapter/data/lp/<pair>/lp_pool_<id>_<interval>.json

This generates both:

• analysis/charts/lp_AMA_chart_pool_133.html — market-adapter style LP chart
• analysis/charts/lp_chart_pool_133.comparison.html — AMA comparison chart (derived from LP metadata)

The unified comparison chart (*_UNIFIED_COMPARISON.html) is produced by npm run ama:chart:lp-local (see workflow overview above).

How market_profiles.json is used

When the optimizer is run with --write-profiles, profiles/market_profiles.json is updated with the new AMA1–AMA4 parameters for the pair. The market adapter reads this file at startup and on each cycle via _resetCycleCache() / findAmaProfileForBot() in market_adapter/market_adapter.ts (lines 127 and 346 respectively). No restart required — takes effect on the next market adapter cycle.

Auxiliary Tools

calibrate_convergence_er.ts

Computes the implied AMA_CONVERGENCE_ER_AVG for modules/constants.ts from real LP candle data. Accounts for Jensen's inequality: the average smoothing constant is not the smoothing constant of the average ER.

node dist/analysis/ama_fitting/calibrate_convergence_er.js --data <lp-file.json> --amas AMA3

analyze_lambda_vs_slow.ts

Fixes ER and Fast at the AMA1 defaults (781 / 5.2), then scans λ (distance weight) over a range to find the optimal Slow period for each λ. Produces a 3-panel chart: λ→Slow, λ→Movement, and Slow→Movement (all cached slow values from 10 to maxSlow).

Useful for understanding how λ shapes the optimal Slow independent of ER/Fast variation. The annotations on the λ→Slow chart mark where the four default AMA λ values land on the curve — differences of ±1–2 slow units vs the 3-D optimizer are expected since the optimizer also tunes ER and Fast simultaneously.

node dist/analysis/ama_fitting/analyze_lambda_vs_slow.js \
  --data <lp-file.json> --maxSlow 250 --lambdaEnd 0.0045 --lambdaSteps 50

analyze_ama_price_changes.ts

Simulates AMA_DELTA_THRESHOLD_PERCENT grid-reposition frequency for all four AMA series on LP candle data. Reports reposition counts and inter-reposition step distributions.

node dist/analysis/ama_fitting/analyze_ama_price_changes.js \
  --data <lp-file.json> --results <optimization-results.json>

Data file format

{
  "meta": {
    "pool": "1.19.133",
    "assetA": { "id": "<asset_a_id>", "precision": <n>, "symbol": "<ASSET_A>" },
    "assetB": { "id": "<asset_b_id>",    "precision": <n>, "symbol": "<ASSET_B>" },
    "intervalSeconds": 3600,
    "lookbackHours": 26280,
    "candleCount": 26280
  },
  "candles": [
    [timestamp_ms, open, high, low, close, volume_A],
    ...
  ]
}