Gesamtempfehlung-Paket: 2 Rollouts, 3 Anzeige-Features, 2 Messungen (v=114)
- Entry-Raum-Gate 0,6→1,0 (Messung lag vor: beide Hälften besser) - Konfidenz-Kalibrierung gemessen (analyze_verdict_calibration.py): conf INVERTIERT zwischen Regimen → keine P(Erfolg)-Aufwertung, kein Konf-Sizing - verdict_votes-Logging (1×/min) für spätere Copilot/Elliott-Entscheidung - Order-Dialog zeigt eigenen Ausrichtungs-Split (36/40 Trades ohne Signal: −423€) - Live-Kosten-Chip (Spread/ATR) im Verdict - P(break)×Entry-Raum-Freigabe gemessen VERWORFEN (Flip in jeder Schwelle, backtest_entryroom_pbreak.py) — 13. verworfener Signal-Eingriff - News-Konflikt-Chip + recommendations.news_score wieder befüllt Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Axel Hocks
co-authored by
Claude Opus 4.8
Claude Opus 4.8
parent
c3c6a8ce5f
commit
d61c4a3634
@@ -0,0 +1,119 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Kalibrierung der Gesamtempfehlung (2026-07-24, Idee 2 „Verdict ans eigene
|
||||
Erfolgsrezept halten"): Ist die angezeigte Konfidenz (conf_pct) PRÄDIKTIV — sagt
|
||||
conf=80 mehr Erfolg voraus als conf=55? Und: sind die unbelegten Verdict-Stimmen
|
||||
(KI-Copilot ai_sentiment, news_score) prädiktiv?
|
||||
|
||||
Methodik wie beim P(break)-Modell (dem kalibrierten Erfolgsfall):
|
||||
- `recommendations` (LONG/SHORT) aus der DB, Sampling ≥15 min Abstand
|
||||
(Autokorrelation dämpfen — die Empfehlung ändert sich minütlich kaum).
|
||||
- Forward-Return = Signalrichtung × (Close[t+H] − Close[t0]) / ATR(M5, t0)
|
||||
für Horizonte 1 h und 2 h (M5-Bars von MT5; rec.timestamp = lokale Epoch,
|
||||
Bar-Zeit = Broker-Zeit → −3 h).
|
||||
- Bins über conf_pct → n / WR(fwd>0) / Ø-fwdR je Bin, in ZWEI Zeitraum-Hälften
|
||||
(Sign-Stabilität wie immer).
|
||||
- Copilot: ai_sentiment (LONG/bullish=+1, SHORT/bearish=−1) → Ø-fwdR in
|
||||
Copilot-Richtung (unabhängig vom Wave-Signal!). News: Vorzeichen news_score.
|
||||
Kalibriert = monoton steigende Ø-fwdR über die conf-Bins, in BEIDEN Hälften.
|
||||
"""
|
||||
import sqlite3, datetime as dt
|
||||
import MetaTrader5 as mt5
|
||||
|
||||
_BROKER_OFF = 3*3600
|
||||
_SAMPLE_S = 15*60 # Mindestabstand zwischen zwei gewerteten Empfehlungen
|
||||
_H1, _H2 = 12, 24 # Forward-Horizonte in M5-Bars (1 h / 2 h)
|
||||
_ATR_P = 14
|
||||
_BINS = [(0,45),(45,55),(55,65),(65,75),(75,85),(85,101)]
|
||||
|
||||
|
||||
def _atr_series(H, L, C, p=_ATR_P):
|
||||
t=[0.0]
|
||||
for i in range(1,len(C)): t.append(max(H[i]-L[i],abs(H[i]-C[i-1]),abs(L[i]-C[i-1])))
|
||||
return [(sum(t[max(1,i-p+1):i+1])/max(1,len(t[max(1,i-p+1):i+1]))) if i else None for i in range(len(C))]
|
||||
|
||||
|
||||
def _dirnum(s):
|
||||
if not s: return 0
|
||||
s = str(s).strip().upper()
|
||||
if s in ("LONG","BULLISH"): return 1
|
||||
if s in ("SHORT","BEARISH"): return -1
|
||||
return 0
|
||||
|
||||
|
||||
def line(lbl, rows, key):
|
||||
v=[r[key] for r in rows]
|
||||
if not v: return f" {lbl:<22} —"
|
||||
n=len(v); wr=100*sum(1 for x in v if x>0)/n
|
||||
return f" {lbl:<22} n={n:>5} WR={wr:>3.0f}% ØR={sum(v)/n:+.3f}"
|
||||
|
||||
|
||||
def main():
|
||||
mt5.initialize()
|
||||
sym = next((c for c in ("SpotCrude","USOIL","WTI","XTIUSD") if mt5.symbol_info(c)), None)
|
||||
bars = None
|
||||
for req in (100000, 80000, 60000, 40000):
|
||||
bars = mt5.copy_rates_from_pos(sym, mt5.TIMEFRAME_M5, 0, req)
|
||||
if bars is not None and len(bars) > 2000: break
|
||||
mt5.shutdown()
|
||||
H=[float(b["high"]) for b in bars]; L=[float(b["low"]) for b in bars]
|
||||
C=[float(b["close"]) for b in bars]
|
||||
T=[int(b["time"])-_BROKER_OFF for b in bars] # → echte UTC-Epoch
|
||||
A=_atr_series(H,L,C)
|
||||
t2i = {t:i for i,t in enumerate(T)}
|
||||
|
||||
con = sqlite3.connect("oil_widget_history.db"); con.row_factory=sqlite3.Row
|
||||
recs = con.execute(
|
||||
"SELECT timestamp, signal, conf_pct, ai_sentiment, news_score "
|
||||
"FROM recommendations WHERE signal IN ('LONG','SHORT') ORDER BY timestamp"
|
||||
).fetchall()
|
||||
|
||||
ev=[]; last_ts=0
|
||||
for r in recs:
|
||||
ts=r["timestamp"]
|
||||
if ts-last_ts < _SAMPLE_S: continue
|
||||
bar_t = (ts//300)*300 # auf M5-Raster runden
|
||||
i = t2i.get(bar_t)
|
||||
if i is None or i+_H2 >= len(C) or not A[i]: continue
|
||||
last_ts=ts
|
||||
d = 1 if r["signal"]=="LONG" else -1
|
||||
atr=max(A[i],0.06)
|
||||
ev.append(dict(ts=ts, conf=r["conf_pct"] or 0, d=d,
|
||||
ai=_dirnum(r["ai_sentiment"]),
|
||||
news=(r["news_score"] if r["news_score"] is not None else None),
|
||||
f1=(C[i+_H1]-C[i])*d/atr, f2=(C[i+_H2]-C[i])*d/atr))
|
||||
if not ev:
|
||||
print("Keine matchbaren Empfehlungen (History-Überlappung prüfen)."); return
|
||||
mid_ts = ev[len(ev)//2]["ts"]
|
||||
halves=[("H1 (alt)",[e for e in ev if e["ts"]<mid_ts]),
|
||||
("H2 (neu)",[e for e in ev if e["ts"]>=mid_ts])]
|
||||
def span(rows):
|
||||
return (f"{dt.datetime.fromtimestamp(rows[0]['ts']):%d.%m.%y}–"
|
||||
f"{dt.datetime.fromtimestamp(rows[-1]['ts']):%d.%m.%y}")
|
||||
print("="*84)
|
||||
print(f" Verdict-Kalibrierung — {sym}, {len(ev)} gesampelte Signale (≥15 min Abstand)")
|
||||
print(f" fwd-Return in Signalrichtung, ×ATR(M5) · Horizonte 1 h/2 h")
|
||||
print("="*84)
|
||||
|
||||
for lbl, rows in halves:
|
||||
print(f"\n{lbl} ({span(rows)}, n={len(rows)}):")
|
||||
print(" conf_pct-Bins (Horizont 2 h) — kalibriert = ØR steigt monoton:")
|
||||
for lo,hi in _BINS:
|
||||
sub=[e for e in rows if lo<=e["conf"]<hi]
|
||||
print(line(f"conf {lo}–{hi-1}", sub, "f2"))
|
||||
print(" KI-Copilot (eigene Richtung, unabhängig vom Wave-Signal):")
|
||||
for tag,dd in (("Copilot LONG",1),("Copilot SHORT",-1)):
|
||||
sub=[dict(f2=e["f2"]*e["d"]*dd) for e in rows if e["ai"]==dd]
|
||||
print(line(tag, sub, "f2"))
|
||||
print(" News-Score (Richtung des Scores, |score|≥0.3):")
|
||||
for tag,dd in (("News bullisch",1),("News bärisch",-1)):
|
||||
sub=[dict(f2=e["f2"]*e["d"]*dd) for e in rows
|
||||
if e["news"] is not None and (e["news"]>=0.3 if dd>0 else e["news"]<=-0.3)]
|
||||
print(line(tag, sub, "f2"))
|
||||
|
||||
print("\n Lesart: conf kalibriert → höhere Bins klar bessere ØR in BEIDEN Hälften.")
|
||||
print(" Copilot/News prädiktiv → 'LONG'-Zeile positiv UND 'SHORT'-Zeile positiv")
|
||||
print(" (jeweils fwd in der EIGENEN Richtung gemessen).")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in New Issue
Block a user