fix: calibrate from replay outcomes
This commit is contained in:
ImBenji
@@ -12,6 +12,7 @@ const { calculateOutcome } = require('../src/autonomy/outcomes');
|
||||
const { createOrderIntent } = require('../src/autonomy/orderIntents');
|
||||
const { enqueueCoordinatorEvent, reconcileArchiveBatch, reconcileLiveBatch } = require('../workers/autonomyWorker');
|
||||
const { scheduleNext } = require('../workers/replayWorker');
|
||||
const { refreshHistoricalCalibration, createDecisions } = require('../workers/calibrationWorker');
|
||||
|
||||
test('autonomy schema and leased jobs are restart-safe', () => {
|
||||
const db = new Database(':memory:');
|
||||
@@ -120,6 +121,53 @@ test('calibration and policy abstain on insufficient evidence', () => {
|
||||
assert.equal(decide({ direction: 'negative', probability: 0.8, expectedExcessReturn: -0.02, lowerReturn: -0.04, sampleSize: 40 }).action, 'SELL');
|
||||
});
|
||||
|
||||
test('historical replay outcomes create replay calibration snapshots', () => {
|
||||
const db = new Database(':memory:');
|
||||
initAutonomySchema(db);
|
||||
const proposal = db.prepare("INSERT INTO autonomy_proposals(payload, information_cutoff, status) VALUES ('{}', '2020-01-01T00:00:00Z', 'accepted')").run();
|
||||
const prediction = db.prepare(`
|
||||
INSERT INTO autonomy_predictions
|
||||
(proposal_id, instrument, direction, event_type, horizon_days, information_cutoff, evidence_article_ids,
|
||||
learning_eligible, strategy_version, origin, replay_run_id, status)
|
||||
VALUES (?, 'NVDA', 'positive', 'earnings', 10, '2020-01-01T00:00:00Z', '[1]', 0, 'test', 'replay', 7, 'resolved')
|
||||
`).run(proposal.lastInsertRowid);
|
||||
db.prepare(`
|
||||
INSERT INTO autonomy_outcomes(prediction_id, excess_return, direction_correct)
|
||||
VALUES (?, 0.04, 1)
|
||||
`).run(prediction.lastInsertRowid);
|
||||
|
||||
assert.equal(refreshHistoricalCalibration(db, 'test-cal'), 1);
|
||||
const snapshot = db.prepare("SELECT source, replay_run_id, sample_size, directional_probability FROM autonomy_calibration_snapshots").get();
|
||||
assert.equal(snapshot.source, 'replay');
|
||||
assert.equal(snapshot.replay_run_id, 7);
|
||||
assert.equal(snapshot.sample_size, 1);
|
||||
assert(snapshot.directional_probability > 0.5);
|
||||
});
|
||||
|
||||
test('live decisions map calibration snapshot fields into policy inputs', () => {
|
||||
const db = new Database(':memory:');
|
||||
initAutonomySchema(db);
|
||||
const proposal = db.prepare("INSERT INTO autonomy_proposals(payload, information_cutoff, status) VALUES ('{}', datetime('now'), 'accepted')").run();
|
||||
const prediction = db.prepare(`
|
||||
INSERT INTO autonomy_predictions
|
||||
(proposal_id, instrument, direction, event_type, horizon_days, information_cutoff, evidence_article_ids,
|
||||
learning_eligible, strategy_version, origin, status)
|
||||
VALUES (?, 'NVDA', 'positive', 'earnings', 10, datetime('now'), '[1]', 1, 'test', 'live', 'open')
|
||||
`).run(proposal.lastInsertRowid);
|
||||
db.prepare(`
|
||||
INSERT INTO autonomy_calibration_snapshots
|
||||
(cohort_key, sample_size, effective_sample_size, directional_probability, expected_excess_return,
|
||||
lower_return, upper_return, parent_cohort_key, version, source)
|
||||
VALUES ('unknown|earnings|10|positive', 40, 42, 0.7, 0.02, -0.01, 0.06, NULL, 'test-cal', 'replay')
|
||||
`).run();
|
||||
|
||||
assert.equal(createDecisions(db), 1);
|
||||
const decision = db.prepare('SELECT * FROM autonomy_decisions WHERE prediction_id=?').get(prediction.lastInsertRowid);
|
||||
assert.equal(decision.action, 'BUY');
|
||||
assert.equal(decision.calibrated_probability, 0.7);
|
||||
assert.equal(decision.expected_excess_return, 0.02);
|
||||
});
|
||||
|
||||
test('paper execution is allowlisted, bounded and idempotent', () => {
|
||||
const intent = validatePaperIntent({ decisionId: 12, instrument: 'NVDA', action: 'BUY', notional: 100 }, { tradable: true, maxNotional: 500 });
|
||||
const broker = createSimulator();
|
||||
|
||||
@@ -6,13 +6,26 @@ const { decide } = require('../src/autonomy/policy');
|
||||
|
||||
function sleep(ms) { return new Promise((resolve) => setTimeout(resolve, ms)); }
|
||||
|
||||
function refreshCalibration(db, version = `cal-${Date.now()}`) {
|
||||
function snapshotToDecisionInput(snapshot, direction) {
|
||||
return {
|
||||
direction,
|
||||
probability: snapshot.directional_probability,
|
||||
expectedExcessReturn: snapshot.expected_excess_return,
|
||||
lowerReturn: snapshot.lower_return,
|
||||
upperReturn: snapshot.upper_return,
|
||||
sampleSize: snapshot.sample_size,
|
||||
};
|
||||
}
|
||||
|
||||
function refreshCalibration(db, version = `cal-${Date.now()}`, { origin = 'live', source = origin, replayRunId = null } = {}) {
|
||||
const learningClause = origin === 'live' ? 'AND p.learning_eligible = 1' : '';
|
||||
const replayClause = replayRunId ? 'AND p.replay_run_id = @replayRunId' : '';
|
||||
const groups = db.prepare(`
|
||||
SELECT p.direction, p.event_type, p.horizon_days, o.*
|
||||
FROM autonomy_predictions p
|
||||
JOIN autonomy_outcomes o ON o.prediction_id = p.id
|
||||
WHERE p.status = 'resolved' AND p.learning_eligible = 1 AND p.origin = 'live'
|
||||
`).all().reduce((map, row) => {
|
||||
WHERE p.status = 'resolved' AND p.origin = @origin ${learningClause} ${replayClause}
|
||||
`).all({ origin, replayRunId }).reduce((map, row) => {
|
||||
const key = cohortKey({ direction: row.direction, eventType: row.event_type, horizonDays: row.horizon_days });
|
||||
if (!map.has(key)) map.set(key, []);
|
||||
map.get(key).push(row);
|
||||
@@ -22,20 +35,44 @@ function refreshCalibration(db, version = `cal-${Date.now()}`) {
|
||||
INSERT INTO autonomy_calibration_snapshots
|
||||
(cohort_key, sample_size, effective_sample_size, directional_probability,
|
||||
expected_excess_return, lower_return, upper_return, parent_cohort_key, version, source, replay_run_id)
|
||||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, 'live', NULL)
|
||||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
|
||||
`);
|
||||
const tx = db.transaction(() => {
|
||||
for (const [key, rows] of groups) {
|
||||
if (db.prepare("SELECT 1 FROM autonomy_calibration_snapshots WHERE cohort_key = ? AND version = ?").get(key, version)) continue;
|
||||
if (db.prepare(`
|
||||
SELECT 1 FROM autonomy_calibration_snapshots
|
||||
WHERE cohort_key = ? AND version = ? AND source = ? AND COALESCE(replay_run_id, 0) = COALESCE(?, 0)
|
||||
`).get(key, version, source, replayRunId)) continue;
|
||||
const result = calibrateOutcomes(rows);
|
||||
insert.run(key, result.sampleSize, result.effectiveSampleSize, result.directionalProbability,
|
||||
result.expectedExcessReturn, result.lowerReturn, result.upperReturn, null, version);
|
||||
result.expectedExcessReturn, result.lowerReturn, result.upperReturn, null, version, source, replayRunId);
|
||||
}
|
||||
});
|
||||
tx();
|
||||
return groups.size;
|
||||
}
|
||||
|
||||
function refreshHistoricalCalibration(db, version = `replay-cal-${Date.now()}`) {
|
||||
const runs = db.prepare(`
|
||||
SELECT DISTINCT replay_run_id AS replayRunId
|
||||
FROM autonomy_predictions
|
||||
WHERE origin = 'replay' AND replay_run_id IS NOT NULL
|
||||
ORDER BY replay_run_id
|
||||
`).all();
|
||||
let groups = 0;
|
||||
for (const run of runs) {
|
||||
groups += refreshCalibration(db, `${version}-run-${run.replayRunId}`, {
|
||||
origin: 'replay',
|
||||
source: 'replay',
|
||||
replayRunId: run.replayRunId,
|
||||
});
|
||||
}
|
||||
if (!runs.length) {
|
||||
groups += refreshCalibration(db, version, { origin: 'replay', source: 'replay' });
|
||||
}
|
||||
return groups;
|
||||
}
|
||||
|
||||
function createDecisions(db, strategyVersion = 'autonomy-1') {
|
||||
const predictions = db.prepare(`
|
||||
SELECT p.* FROM autonomy_predictions p
|
||||
@@ -57,7 +94,7 @@ function createDecisions(db, strategyVersion = 'autonomy-1') {
|
||||
const key = cohortKey({ direction: prediction.direction, eventType: prediction.event_type, horizonDays: prediction.horizon_days });
|
||||
const calibration = latest.get(key);
|
||||
const decision = calibration
|
||||
? decide({ ...calibration, direction: prediction.direction }, { minSampleSize: 30 })
|
||||
? decide(snapshotToDecisionInput(calibration, prediction.direction), { minSampleSize: 30 })
|
||||
: { action: 'ABSTAIN', rationale: 'calibration unavailable' };
|
||||
insert.run(prediction.id, decision.action, calibration?.directional_probability || null,
|
||||
calibration?.expected_excess_return || null, decision.rationale, strategyVersion);
|
||||
@@ -114,10 +151,12 @@ async function runCalibrationWorker({ intelligencePath, pollMs = 60000, workerId
|
||||
while (true) {
|
||||
try {
|
||||
const state = db.prepare('SELECT COUNT(*) AS count, COALESCE(MAX(prediction_id), 0) AS max_id FROM autonomy_outcomes').get();
|
||||
const groups = refreshCalibration(db, `cal-${state.count}-${state.max_id}`);
|
||||
const version = `cal-${state.count}-${state.max_id}`;
|
||||
const groups = refreshCalibration(db, version);
|
||||
const historicalGroups = refreshHistoricalCalibration(db, version);
|
||||
const decisions = createDecisions(db);
|
||||
const replayEvaluations = refreshReplayEvaluations(db);
|
||||
if (groups || decisions || replayEvaluations) console.log(`[${workerId}] calibration groups=${groups} decisions=${decisions} replay_evaluations=${replayEvaluations}`);
|
||||
if (groups || historicalGroups || decisions || replayEvaluations) console.log(`[${workerId}] calibration groups=${groups} historical_groups=${historicalGroups} decisions=${decisions} replay_evaluations=${replayEvaluations}`);
|
||||
} catch (error) {
|
||||
console.error(`[${workerId}] calibration error:`, error.message);
|
||||
}
|
||||
@@ -125,4 +164,4 @@ async function runCalibrationWorker({ intelligencePath, pollMs = 60000, workerId
|
||||
}
|
||||
}
|
||||
|
||||
module.exports = { refreshCalibration, createDecisions, refreshReplayEvaluations, runCalibrationWorker };
|
||||
module.exports = { refreshCalibration, refreshHistoricalCalibration, createDecisions, refreshReplayEvaluations, runCalibrationWorker };
|
||||
|
||||
Reference in New Issue
Block a user