Files
signal-platform/app/services/outcome_service.py
T
dennisthiessen 7e9a6cd7ec
Deploy / lint (push) Successful in 8s
Deploy / test (push) Successful in 1m1s
Deploy / deploy (push) Successful in 36s
fix: only count matured setups in the live track record
The outcome stats were dominated by quick stop-outs: near stops resolve as losses
within days while far targets take weeks, so a young sample (mostly pending,
0 expired) skewed sharply negative (e.g. 13.8% hit / -0.46R vs the backtest's
35.8% / +0.18R) — a maturation artifact, not a real result.

get_performance_stats now counts only setups whose full ~30-day window has
elapsed (_MATURITY_DAYS), so winners had as long as losers (unbiased, and
comparable to the backtest). A new `maturing` count reports the younger setups
held back. The Track Record UI relabels "Evaluated" -> "Matured", shows the
maturing count, and explains the window in the empty state + methodology note.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-28 13:41:48 +02:00

263 lines
9.2 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""Trade setup outcome evaluation service.
Closes the feedback loop on R:R scanner setups: walks daily OHLCV bars
after detection and records whether the stop or the target was hit first.
Outcome semantics (entry is the close at detection time, i.e. market entry):
- target_hit: target reached before the stop
- stop_hit: stop reached before the target
- ambiguous: stop AND target both within the same daily bar — with daily
granularity the order is unknowable, counted as a loss in stats
- expired: neither level hit within ``max_bars`` trading days
- (NULL): not enough bars yet to decide — re-evaluated on the next run
"""
from __future__ import annotations
import logging
from dataclasses import dataclass
from datetime import date, datetime, timedelta, timezone
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession
from app.models.ohlcv import OHLCVRecord
from app.models.trade_setup import TradeSetup
from app.services.qualification import setup_qualifies
logger = logging.getLogger(__name__)
OUTCOME_TARGET_HIT = "target_hit"
OUTCOME_STOP_HIT = "stop_hit"
OUTCOME_AMBIGUOUS = "ambiguous"
OUTCOME_EXPIRED = "expired"
DEFAULT_MAX_BARS = 30
# A setup's outcome is only unbiased once its full evaluation window has elapsed:
# until then, near stops resolve as losses within days while far targets are still
# pending, so a young sample skews sharply negative. Only count setups detected at
# least this many CALENDAR days ago (~max_bars trading days, ×1.5 to cover
# weekends/holidays). Younger setups are reported separately as "maturing".
_MATURITY_DAYS = int(DEFAULT_MAX_BARS * 1.5)
# Confidence buckets for the performance breakdown
_CONFIDENCE_BUCKETS = [
("<50%", 0.0, 50.0),
("50-70%", 50.0, 70.0),
("≥70%", 70.0, 100.01),
]
@dataclass(frozen=True)
class Bar:
date: date
high: float
low: float
def evaluate_setup_against_bars(
direction: str,
stop_loss: float,
target: float,
bars: list[Bar],
max_bars: int = DEFAULT_MAX_BARS,
) -> tuple[str | None, date | None]:
"""Determine a setup's outcome from daily bars strictly after detection.
Returns (outcome, outcome_date); (None, None) while still undecided.
"""
for i, bar in enumerate(bars):
if i >= max_bars:
break
if direction == "long":
stop_hit = bar.low <= stop_loss
target_hit = bar.high >= target
else:
stop_hit = bar.high >= stop_loss
target_hit = bar.low <= target
if stop_hit and target_hit:
return OUTCOME_AMBIGUOUS, bar.date
if stop_hit:
return OUTCOME_STOP_HIT, bar.date
if target_hit:
return OUTCOME_TARGET_HIT, bar.date
if len(bars) >= max_bars:
return OUTCOME_EXPIRED, bars[max_bars - 1].date
return None, None
async def evaluate_pending_setups(
db: AsyncSession,
max_bars: int = DEFAULT_MAX_BARS,
) -> dict[str, int]:
"""Evaluate all unevaluated trade setups against stored OHLCV data.
Bars are fetched once per ticker. Setups that cannot be decided yet
remain NULL and are picked up on the next run.
"""
result = await db.execute(
select(TradeSetup).where(TradeSetup.actual_outcome.is_(None))
)
pending = list(result.scalars().all())
summary = {"evaluated": 0, "still_pending": 0, "by_outcome": {}}
if not pending:
return summary
by_ticker: dict[int, list[TradeSetup]] = {}
for setup in pending:
by_ticker.setdefault(setup.ticker_id, []).append(setup)
now = datetime.now(timezone.utc)
for ticker_id, setups in by_ticker.items():
earliest = min(s.detected_at for s in setups).date()
bars_result = await db.execute(
select(OHLCVRecord)
.where(
OHLCVRecord.ticker_id == ticker_id,
OHLCVRecord.date > earliest,
)
.order_by(OHLCVRecord.date.asc())
)
records = list(bars_result.scalars().all())
all_bars = [Bar(date=r.date, high=r.high, low=r.low) for r in records]
for setup in setups:
detected_date = setup.detected_at.date()
bars = [b for b in all_bars if b.date > detected_date]
outcome, outcome_date = evaluate_setup_against_bars(
setup.direction, setup.stop_loss, setup.target, bars, max_bars
)
if outcome is None:
summary["still_pending"] += 1
continue
setup.actual_outcome = outcome
setup.outcome_date = outcome_date
setup.evaluated_at = now
summary["evaluated"] += 1
summary["by_outcome"][outcome] = summary["by_outcome"].get(outcome, 0) + 1
await db.commit()
return summary
def _realized_r(setup: TradeSetup) -> float | None:
"""Realized result in R-multiples: win = +rr_ratio, loss = -1R, expired = 0R."""
if setup.actual_outcome == OUTCOME_TARGET_HIT:
return setup.rr_ratio
if setup.actual_outcome in (OUTCOME_STOP_HIT, OUTCOME_AMBIGUOUS):
return -1.0
if setup.actual_outcome == OUTCOME_EXPIRED:
return 0.0
return None
def _bucket_stats(setups: list[TradeSetup]) -> dict:
wins = sum(1 for s in setups if s.actual_outcome == OUTCOME_TARGET_HIT)
losses = sum(
1 for s in setups if s.actual_outcome in (OUTCOME_STOP_HIT, OUTCOME_AMBIGUOUS)
)
expired = sum(1 for s in setups if s.actual_outcome == OUTCOME_EXPIRED)
decided = wins + losses
realized = [r for s in setups if (r := _realized_r(s)) is not None]
return {
"total": len(setups),
"wins": wins,
"losses": losses,
"expired": expired,
"hit_rate": round(wins / decided * 100, 1) if decided else None,
"avg_r": round(sum(realized) / len(realized), 3) if realized else None,
"total_r": round(sum(realized), 2) if realized else None,
}
def _confidence_bucket(score: float | None) -> str | None:
if score is None:
return None
for label, lo, hi in _CONFIDENCE_BUCKETS:
if lo <= score < hi:
return label
return None
async def get_performance_stats(
db: AsyncSession,
config: dict | None = None,
) -> dict:
"""Aggregate outcome statistics over the *matured* evaluated trade setups.
Only setups whose full evaluation window has elapsed (see ``_MATURITY_DAYS``)
are counted, so the headline isn't dominated by quick stop-outs while slower
winners are still in flight. ``maturing`` reports how many are excluded for
being too young.
avg_r is the expectancy per trade in R-multiples (win = +rr_ratio,
loss = -1R, expired = 0R). A positive avg_r means the signals have
been profitable on a risk-adjusted basis.
When ``config`` (an activation-gate dict) is supplied, the overall,
direction and action breakdowns cover only qualified setups. The
confidence breakdown deliberately stays unfiltered: it is the
instrument for validating the gate itself.
"""
result = await db.execute(
select(TradeSetup).where(TradeSetup.actual_outcome.is_not(None))
)
evaluated_all = list(result.scalars().all())
# Matured cohort only — see _MATURITY_DAYS. Setups whose window hasn't fully
# elapsed are excluded so quick stop-outs can't drag the headline negative
# while their slower-to-resolve winners are still in flight.
cutoff_date = (datetime.now(timezone.utc) - timedelta(days=_MATURITY_DAYS)).date()
evaluated = [s for s in evaluated_all if s.detected_at.date() <= cutoff_date]
pending_result = await db.execute(
select(TradeSetup.id).where(TradeSetup.actual_outcome.is_(None))
)
pending_count = len(pending_result.scalars().all())
# Still inside their measurement window (excluded above so they can't bias the
# stats): young setups that already resolved + everything still pending.
maturing_count = (len(evaluated_all) - len(evaluated)) + pending_count
if config is not None:
qualified = [s for s in evaluated if setup_qualifies(s, config)]
else:
qualified = evaluated
by_direction: dict[str, list[TradeSetup]] = {}
by_action: dict[str, list[TradeSetup]] = {}
by_confidence: dict[str, list[TradeSetup]] = {}
for setup in qualified:
by_direction.setdefault(setup.direction, []).append(setup)
action = setup.recommended_action or "NONE"
by_action.setdefault(action, []).append(setup)
# Confidence buckets always cover the full evaluated population
for setup in evaluated:
bucket = _confidence_bucket(setup.confidence_score)
if bucket is not None:
by_confidence.setdefault(bucket, []).append(setup)
bucket_order = [label for label, _, _ in _CONFIDENCE_BUCKETS]
return {
"overall": _bucket_stats(qualified),
"pending": pending_count,
"maturing": maturing_count,
"by_direction": {k: _bucket_stats(v) for k, v in sorted(by_direction.items())},
"by_action": {k: _bucket_stats(v) for k, v in sorted(by_action.items())},
"by_confidence": {
label: _bucket_stats(by_confidence[label])
for label in bucket_order
if label in by_confidence
},
}