Files
Apex_AI_MT5_EA_Optimizer/validation/gate.py
T

217 lines
7.9 KiB
Python
Raw Normal View History

2026-04-13 02:28:09 +00:00
"""
validation/gate.py
IS / Walk-Forward / OOS validation pipeline.
"""
from __future__ import annotations
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Optional
import uuid
import yaml
from loguru import logger
from data.models import GateResult, RunMetrics
@dataclass
class WFVResult:
passed: bool
oos_is_ratio: float
fold_results: list[dict]
details: dict
class ValidationGate:
"""
Three-phase validation pipeline:
1. IS check — minimum thresholds on in-sample metrics
2. Walk-Forward Validation — split training period, test metric consistency
3. OOS test — held-out period, called explicitly by orchestrator
"""
def __init__(self, config_path: str | Path = "config.yaml"):
with open(config_path) as f:
cfg = yaml.safe_load(f)
self.thresh = cfg["thresholds"]
self.per = cfg["periods"]
# ── Phase 1: IS Check ─────────────────────────────────────────────────────
def run_is_check(
self,
metrics: RunMetrics,
baseline_score: float = 0.0,
) -> GateResult:
2026-04-13 02:28:09 +00:00
"""
Two-tier IS check:
- MUST: enough trades for statistical confidence
- MUST: composite_score is better than baseline (or meets abs thresholds)
Absolute calmar/PF thresholds are logged as warnings but are NOT blockers
when the hypothesis shows clear improvement over baseline.
2026-04-13 02:28:09 +00:00
"""
min_trades_ok = metrics.total_trades >= self.thresh["min_trades"]
abs_pf_ok = metrics.profit_factor >= self.thresh["min_profit_factor"]
abs_calmar_ok = metrics.calmar_ratio >= self.thresh["min_calmar"]
# Score-based relative pass: hypothesis is better than baseline
score_improve_ok = metrics.composite_score > baseline_score * 1.0 + 0.01
# Gate passes if:
# a) Enough trades AND (abs thresholds met OR clearly better than baseline)
passed = min_trades_ok and (
(abs_pf_ok and abs_calmar_ok) # standard absolute pass
or score_improve_ok # OR better than baseline
)
2026-04-13 02:28:09 +00:00
checks = {
"min_trades": min_trades_ok,
"min_pf": abs_pf_ok,
"min_calmar": abs_calmar_ok,
"score_improve": score_improve_ok,
2026-04-13 02:28:09 +00:00
}
reason = None
if not passed:
failed = [k for k, v in checks.items() if not v]
reason = f"Failed gates: {', '.join(failed)}"
logger.info(f"IS check {'PASSED' if passed else 'FAILED'}: {checks}")
return GateResult(passed=passed, details=checks, reason=reason)
# ── Phase 2: Walk-Forward Validation ─────────────────────────────────────
def run_walk_forward(
self,
2026-04-13 03:05:40 +00:00
params: dict[str, Any],
executor, # callable(params, start_str, end_str, fold_id) -> Optional[RunMetrics]
n_folds: int = 2,
2026-04-13 02:28:09 +00:00
) -> WFVResult:
"""
Split training period into n_folds sub-periods.
Run the same params on each sub-period.
Accept if mean OOS Calmar >= min_wfv_ratio × IS Calmar.
In MVP we use n_folds=2 (first half / second half).
v2 will use full rolling window WFV.
"""
from datetime import datetime, timedelta
train_start = datetime.strptime(self.per["train_start"], "%Y.%m.%d")
train_end = datetime.strptime(self.per["train_end"], "%Y.%m.%d")
total_days = (train_end - train_start).days
fold_days = total_days // n_folds
fold_metrics: list[RunMetrics] = []
for i in range(n_folds):
fold_start = train_start + timedelta(days=i * fold_days)
fold_end = fold_start + timedelta(days=fold_days)
if i == n_folds - 1:
fold_end = train_end # last fold gets remainder
logger.info(f"WFV fold {i+1}/{n_folds}: {fold_start.date()}{fold_end.date()}")
2026-04-13 03:05:40 +00:00
fold_id = f"wfv_fold{i+1}_{uuid.uuid4().hex[:6]}"
start_s = fold_start.strftime("%Y.%m.%d")
end_s = fold_end.strftime("%Y.%m.%d")
fm = executor(params, start_s, end_s, fold_id)
2026-04-13 02:28:09 +00:00
if fm:
fold_metrics.append(fm)
if not fold_metrics:
return WFVResult(passed=False, oos_is_ratio=0.0, fold_results=[], details={})
fold_calmars = [fm.calmar_ratio for fm in fold_metrics]
mean_oos_calmar = sum(fold_calmars) / len(fold_calmars)
# Get IS calmar (best score so far) for comparison
is_calmar = max((fm.calmar_ratio for fm in fold_metrics), default=0)
if is_calmar <= 0:
ratio = 0.0
else:
ratio = mean_oos_calmar / is_calmar
passed = ratio >= self.thresh["min_wfv_ratio"]
return WFVResult(
passed=passed,
oos_is_ratio=ratio,
fold_results=[
{"fold": i+1, "calmar": fm.calmar_ratio, "trades": fm.total_trades}
for i, fm in enumerate(fold_metrics)
],
details={
"mean_fold_calmar": round(mean_oos_calmar, 4),
"ratio": round(ratio, 4),
"threshold": self.thresh["min_wfv_ratio"],
},
)
# ── Parameter Sensitivity Check ───────────────────────────────────────────
def check_sensitivity(
self,
params: dict[str, Any],
key_params: list[str],
cfg: dict,
store,
builder,
runner,
parser,
log_rdr,
analyzers,
scorer,
perturbation: float = 0.10,
) -> tuple[bool, dict]:
"""
For each key parameter, perturb by ±10% and measure Calmar change.
Reject if any parameter causes > tolerance% degradation.
In MVP this is optional — add to v2 workflow.
"""
from main import execute_run
import uuid
tolerance = self.thresh["sensitivity_tolerance"]
degradations = {}
base_metrics, _ = execute_run(
run_id=f"sens_base_{uuid.uuid4().hex[:6]}",
params=params,
period_start=self.per["train_start"],
period_end=self.per["train_end"],
phase="validate",
hypothesis_id=None,
cfg=cfg, store=store, builder=builder, runner=runner,
parser=parser, log_rdr=log_rdr, analyzers=analyzers, scorer=scorer,
)
if base_metrics is None or base_metrics.calmar_ratio <= 0:
return True, {} # can't test sensitivity — skip
for param in key_params:
if param not in params:
continue
base_val = params[param]
if not isinstance(base_val, (int, float)):
continue
perturbed = {**params, param: base_val * (1 + perturbation)}
pm, _ = execute_run(
run_id=f"sens_{param[:8]}_{uuid.uuid4().hex[:6]}",
params=perturbed,
period_start=self.per["train_start"],
period_end=self.per["train_end"],
phase="validate",
hypothesis_id=None,
cfg=cfg, store=store, builder=builder, runner=runner,
parser=parser, log_rdr=log_rdr, analyzers=analyzers, scorer=scorer,
)
if pm and base_metrics.calmar_ratio > 0:
deg = (base_metrics.calmar_ratio - pm.calmar_ratio) / base_metrics.calmar_ratio
degradations[param] = round(deg, 4)
max_deg = max(degradations.values(), default=0)
passed = max_deg <= tolerance
return passed, degradations