feat: pattern calibration — progressive AI→observed expected_move blending
DB (database.py): - 3 new columns on custom_patterns: calibrated_expected_move, calibration_weight, observed_avg_win_pct - update_bayesian_posteriors() now also computes credibility blend w=n/(n+5): calibrated = (1-w)*ai_estimate + w*observed_avg_win_pct (only when wins exist) - log_trade_entries() prefers calibrated_expected_move when w>10% - get_calibration_summary() returns per-pattern state (source: pure_ai/early/mixed/data_driven) Backend (patterns.py, auto_cycle.py): - GET /api/patterns/calibration endpoint - calibration_report block in cycle report: counts by source, avg weight, per-pattern detail Frontend (PatternExplorer.tsx, RapportIA.tsx, useApi.ts): - MaturityBadge on each PatternCard: blend bar (AI→observed), win rate, AI estimate vs calibrated - usePatternCalibration hook - Cycle report: calibration section with global bar + per-pattern table (weight%, n_trades, WR, AI→calibrated) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
OpenSquared
@@ -86,6 +86,10 @@ def init_db():
|
||||
# Pattern Lab — backtest reliability tracking
|
||||
"ALTER TABLE custom_patterns ADD COLUMN backtest_hits INTEGER DEFAULT 0",
|
||||
"ALTER TABLE custom_patterns ADD COLUMN backtest_runs_count INTEGER DEFAULT 0",
|
||||
# Calibration — observed vs AI blending
|
||||
"ALTER TABLE custom_patterns ADD COLUMN calibrated_expected_move REAL",
|
||||
"ALTER TABLE custom_patterns ADD COLUMN calibration_weight REAL DEFAULT 0.0",
|
||||
"ALTER TABLE custom_patterns ADD COLUMN observed_avg_win_pct REAL",
|
||||
# Remove all built-in patterns (no proof of legitimacy)
|
||||
"DELETE FROM custom_patterns WHERE source = 'builtin'",
|
||||
# Regime / counter-scenario architecture
|
||||
@@ -1506,11 +1510,16 @@ def log_trade_entries(run_id: str, scored_patterns: List[Dict[str, Any]], quotes
|
||||
|
||||
delta = int(trade.get("score_delta") or 0)
|
||||
eff_score = max(0, min(100, base_score + delta))
|
||||
# Fallback chain: trade field → scored sp field → auto_cycle enrichment → original DB pattern
|
||||
# Fallback chain: trade field → scored sp field → calibrated DB (observed blend) → AI estimate DB
|
||||
_calib = _orig.get("calibrated_expected_move")
|
||||
_calib_w = float(_orig.get("calibration_weight") or 0)
|
||||
_ai_est = _orig.get("expected_move_pct")
|
||||
# Use calibrated value when credibility weight > 10% (at least ~1 mature trade)
|
||||
_db_move = _calib if (_calib and _calib_w > 0.1) else _ai_est
|
||||
exp_move = abs(float(
|
||||
trade.get("expected_move_pct") or
|
||||
sp.get("expected_move_pct") or
|
||||
_orig.get("expected_move_pct") or
|
||||
_db_move or
|
||||
0
|
||||
))
|
||||
if exp_move == 0:
|
||||
@@ -3328,18 +3337,31 @@ def get_risk_dashboard() -> Dict:
|
||||
|
||||
def update_bayesian_posteriors() -> int:
|
||||
"""
|
||||
Met à jour les posteriors Beta(α,β) de chaque pattern selon ses trades matures.
|
||||
Prior faible : α₀=1, β₀=1 (Laplace smoothing).
|
||||
Posterior : α = 1 + wins, β = 1 + losses → win_rate bayésien = α/(α+β)
|
||||
Retourne le nombre de patterns mis à jour.
|
||||
Met à jour les posteriors Beta(α,β) et la calibration de l'expected_move.
|
||||
|
||||
Bayesian: α = 1 + wins, β = 1 + losses → bayesian_win_rate = α/(α+β)
|
||||
|
||||
Calibration (credibility blending):
|
||||
k = 5 → w = n / (n + 5) (50% credibility at 5 trades, 80% at 20)
|
||||
observed_avg_win_pct = mean(pnl_pct for wins)
|
||||
calibrated_expected_move = (1-w) × ai_estimate + w × observed_avg_win_pct
|
||||
(only blends when at least 1 win exists; pure AI until then)
|
||||
"""
|
||||
import math as _math
|
||||
from datetime import date as _date
|
||||
|
||||
_CREDIBILITY_K = 5 # trades needed to reach 50% observed weight
|
||||
|
||||
conn = get_conn()
|
||||
rows = conn.execute(
|
||||
"SELECT pattern_id, pnl_pct, entry_date, horizon_days FROM trade_entry_prices WHERE pnl_pct IS NOT NULL"
|
||||
).fetchall()
|
||||
# Also load ai estimates for blending
|
||||
ai_estimates = {
|
||||
r["id"]: float(r["expected_move_pct"] or 0)
|
||||
for r in conn.execute("SELECT id, expected_move_pct FROM custom_patterns").fetchall()
|
||||
if r["expected_move_pct"]
|
||||
}
|
||||
conn.close()
|
||||
|
||||
today = _date.today()
|
||||
@@ -3353,7 +3375,7 @@ def update_bayesian_posteriors() -> int:
|
||||
continue
|
||||
horizon = r.get("horizon_days") or 30
|
||||
if days_held / max(horizon, 1) < 0.35:
|
||||
continue # trades immatures exclus
|
||||
continue
|
||||
by_pattern.setdefault(r["pattern_id"], []).append(float(r["pnl_pct"] or 0))
|
||||
|
||||
if not by_pattern:
|
||||
@@ -3365,17 +3387,36 @@ def update_bayesian_posteriors() -> int:
|
||||
now_iso = datetime.utcnow().isoformat()
|
||||
for pid, pnls in by_pattern.items():
|
||||
n = len(pnls)
|
||||
wins = sum(1 for p in pnls if p > 0)
|
||||
wins_pnl = [p for p in pnls if p > 0]
|
||||
wins = len(wins_pnl)
|
||||
losses = n - wins
|
||||
alpha = 1.0 + wins # posterior alpha
|
||||
beta = 1.0 + losses # posterior beta
|
||||
|
||||
# Bayesian posteriors
|
||||
alpha = 1.0 + wins
|
||||
beta = 1.0 + losses
|
||||
bayes_wr = alpha / (alpha + beta)
|
||||
|
||||
# Credibility blending
|
||||
w = n / (n + _CREDIBILITY_K)
|
||||
ai_est = ai_estimates.get(pid)
|
||||
observed_avg_win = round(sum(wins_pnl) / len(wins_pnl), 2) if wins_pnl else None
|
||||
if ai_est and observed_avg_win is not None:
|
||||
calibrated = round((1 - w) * ai_est + w * observed_avg_win, 2)
|
||||
else:
|
||||
calibrated = None # not enough data — keep pure AI in log_trade_entries
|
||||
|
||||
c.execute("""
|
||||
UPDATE custom_patterns
|
||||
SET bayesian_alpha=?, bayesian_beta=?, bayesian_win_rate=?,
|
||||
bayesian_updated_at=?, bayesian_sample_size=?
|
||||
bayesian_updated_at=?, bayesian_sample_size=?,
|
||||
calibration_weight=?, observed_avg_win_pct=?, calibrated_expected_move=?
|
||||
WHERE id=?
|
||||
""", (round(alpha, 1), round(beta, 1), round(bayes_wr, 4), now_iso, n, pid))
|
||||
""", (
|
||||
round(alpha, 1), round(beta, 1), round(bayes_wr, 4),
|
||||
now_iso, n,
|
||||
round(w, 4), observed_avg_win, calibrated,
|
||||
pid,
|
||||
))
|
||||
if c.rowcount:
|
||||
updated += 1
|
||||
conn.commit()
|
||||
@@ -3383,6 +3424,53 @@ def update_bayesian_posteriors() -> int:
|
||||
return updated
|
||||
|
||||
|
||||
def get_calibration_summary() -> List[Dict]:
|
||||
"""
|
||||
Returns per-pattern calibration state for the cycle report and UI.
|
||||
Includes: ai_estimate, observed_avg_win_pct, calibration_weight,
|
||||
calibrated_expected_move, n_mature_trades, win_rate.
|
||||
"""
|
||||
conn = get_conn()
|
||||
rows = conn.execute("""
|
||||
SELECT cp.id, cp.name, cp.asset_class,
|
||||
cp.expected_move_pct AS ai_estimate,
|
||||
cp.calibrated_expected_move AS calibrated,
|
||||
cp.calibration_weight AS weight,
|
||||
cp.observed_avg_win_pct AS observed_win,
|
||||
cp.bayesian_sample_size AS n_trades,
|
||||
cp.bayesian_win_rate AS bayes_wr
|
||||
FROM custom_patterns cp
|
||||
WHERE cp.is_active = 1
|
||||
ORDER BY cp.calibration_weight DESC NULLS LAST, cp.bayesian_sample_size DESC
|
||||
""").fetchall()
|
||||
conn.close()
|
||||
|
||||
result = []
|
||||
for r in rows:
|
||||
d = dict(r)
|
||||
w = d.get("weight") or 0.0
|
||||
n = d.get("n_trades") or 0
|
||||
result.append({
|
||||
"pattern_id": d["id"],
|
||||
"pattern_name": d["name"],
|
||||
"asset_class": d.get("asset_class", ""),
|
||||
"ai_estimate": d.get("ai_estimate"),
|
||||
"observed_avg_win_pct": d.get("observed_win"),
|
||||
"calibrated_expected_move": d.get("calibrated"),
|
||||
"calibration_weight": round(w, 4),
|
||||
"calibration_weight_pct": round(w * 100, 1),
|
||||
"n_mature_trades": n,
|
||||
"bayes_win_rate": round((d.get("bayes_wr") or 0), 3),
|
||||
"source": (
|
||||
"pure_ai" if w < 0.1 else
|
||||
"early" if w < 0.4 else
|
||||
"mixed" if w < 0.75 else
|
||||
"data_driven"
|
||||
),
|
||||
})
|
||||
return result
|
||||
|
||||
|
||||
def get_bayesian_posteriors() -> List[Dict]:
|
||||
"""
|
||||
Retourne tous les patterns avec leurs posteriors bayésiens + intervalle de crédibilité 95%.
|
||||
|
||||
Reference in New Issue
Block a user