feat: add Phase A research matrix (vol target, fill, corr, SE/DSR)

Ship shared Sharpe SE/PSR diagnostics, next-open fill and equity-curve vol targeting in the portfolio simulator, re-derived fip_id, and a checkpointed offline matrix runner for Mac-side validation sweeps.
This commit is contained in:
2026-07-18 15:04:44 +02:00
parent cad4b49e7c
commit 529343ce82
3 changed files with 1536 additions and 36 deletions
+814
View File
@@ -0,0 +1,814 @@
"""Phase-A research matrix: max-hold, vol targeting, next-open fill, corr caps.
Promotion rule (pre-registered — do not edit after a run starts)
----------------------------------------------------------------
An arm may be promoted over the close-fill production **control** only if ALL of:
1. Validation-window (entries ≥ ``--validation-split``, default 2024-07-01)
Sharpe ≥ control validation Sharpe.
2. Validation max drawdown is not worse than control by more than 2 percentage
points (higher DD is worse).
3. Train-window Sharpe is not worse than control train Sharpe (both-windows
consistency — same standard as the min_rr sweep).
4. Report whether the validation Sharpe delta exceeds 1 × SE (control or arm);
most arms will fail this distinguishability check — that is expected and is
the reason SE/PSR ship on every row. Failing the 1-SE bar does **not** alone
veto promotion under (1)(3), but it must be stated.
Naming: the post-split window is called **validation**, not "holdout". It has
been opened by prior experiments; treat it as a disciplined check, not a
pristine sample.
Arms (pre-registered; N used for Deflated Sharpe)
-------------------------------------------------
- A0 control: production gate/rank/trail, hold=30, close fill, no vol target, no corr cap
- A2 max-hold: hold ∈ {30, 45, 60, 90} (30 is the control row; listed once)
- A3 vol-target: target ∈ {15%, 20%, 25%} × clamp {[0.5,1.5], [0.25,2.0]} at lookback 60;
plus sensitivity lookbacks {20, 126} at target 20% / clamp [0.5,1.5] only
- A4 next-open fill (measurement + portfolio consequence vs control)
- A5 corr cap: threshold ∈ {0.6, 0.7, 0.8} × action ∈ {skip, half-size}
DSR uses N = number of pre-registered strategy arms in this matrix (see
``PRE_REGISTERED_ARM_IDS``). Standalone backtests do not invent a DSR.
Calendar truncation
-------------------
The simulator always cuts the equity calendar at last_signal + hold_days
(+1 for next-open). The runner asserts validation end_date ≤ last price date and
that the sim end is within hold_days+pad of the last admitted signal so a 90d
arm cannot sit in trailing flat cash.
Usage
-----
python scripts/run_research_matrix.py backtest_snapshots/prod.sqlite \\
--workers 7 --allow-spawn --candidate-cache reports/.cache/research-cands.pkl
python scripts/run_research_matrix.py ... --only a2,a3
python scripts/run_research_matrix.py ... --skip a4
"""
from __future__ import annotations
import argparse
import asyncio
import json
import multiprocessing
import os
import pickle
import sys
from concurrent.futures import ProcessPoolExecutor, as_completed
from datetime import date, datetime
from pathlib import Path
from typing import Any
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
CACHE_VERSION = "research-matrix-v1-daily-prod"
# Pre-registered arm catalogue (order is report order). Control is a0.
# Count N for DSR excludes pure measurement-only rows if any; every arm below
# is a portfolio book and counts.
PRE_REGISTERED_ARMS: tuple[dict[str, Any], ...] = (
{
"id": "a0_control",
"group": "a0",
"label": "Control: close fill, hold 30, risk 1%, no corr/vol",
"hold_days": 30,
"fill_mode": "close",
},
# A2 — max hold (30 is control; still emitted as a2 for the sweep table)
{"id": "a2_hold_30", "group": "a2", "label": "Max hold 30", "hold_days": 30},
{"id": "a2_hold_45", "group": "a2", "label": "Max hold 45", "hold_days": 45},
{"id": "a2_hold_60", "group": "a2", "label": "Max hold 60", "hold_days": 60},
{"id": "a2_hold_90", "group": "a2", "label": "Max hold 90", "hold_days": 90},
# A3 — vol targeting
{
"id": "a3_vt15_c05_15_lb60",
"group": "a3",
"label": "Vol target 15% clamp[0.5,1.5] lb60",
"vol_target": 0.15,
"vol_clamp": (0.5, 1.5),
"vol_lookback": 60,
},
{
"id": "a3_vt20_c05_15_lb60",
"group": "a3",
"label": "Vol target 20% clamp[0.5,1.5] lb60",
"vol_target": 0.20,
"vol_clamp": (0.5, 1.5),
"vol_lookback": 60,
},
{
"id": "a3_vt25_c05_15_lb60",
"group": "a3",
"label": "Vol target 25% clamp[0.5,1.5] lb60",
"vol_target": 0.25,
"vol_clamp": (0.5, 1.5),
"vol_lookback": 60,
},
{
"id": "a3_vt15_c025_20_lb60",
"group": "a3",
"label": "Vol target 15% clamp[0.25,2.0] lb60",
"vol_target": 0.15,
"vol_clamp": (0.25, 2.0),
"vol_lookback": 60,
},
{
"id": "a3_vt20_c025_20_lb60",
"group": "a3",
"label": "Vol target 20% clamp[0.25,2.0] lb60",
"vol_target": 0.20,
"vol_clamp": (0.25, 2.0),
"vol_lookback": 60,
},
{
"id": "a3_vt25_c025_20_lb60",
"group": "a3",
"label": "Vol target 25% clamp[0.25,2.0] lb60",
"vol_target": 0.25,
"vol_clamp": (0.25, 2.0),
"vol_lookback": 60,
},
{
"id": "a3_vt20_c05_15_lb20",
"group": "a3",
"label": "Vol target 20% clamp[0.5,1.5] lb20 (sensitivity)",
"vol_target": 0.20,
"vol_clamp": (0.5, 1.5),
"vol_lookback": 20,
},
{
"id": "a3_vt20_c05_15_lb126",
"group": "a3",
"label": "Vol target 20% clamp[0.5,1.5] lb126 (sensitivity)",
"vol_target": 0.20,
"vol_clamp": (0.5, 1.5),
"vol_lookback": 126,
},
# A4 — next-open fill
{
"id": "a4_next_open",
"group": "a4",
"label": "Next-open fill (t+1 open, stop from fill1.5 ATR)",
"fill_mode": "next_open",
},
# A5 — correlation caps
{
"id": "a5_corr06_skip",
"group": "a5",
"label": "Corr max 0.6 skip",
"corr_max": 0.6,
"corr_action": "skip",
},
{
"id": "a5_corr07_skip",
"group": "a5",
"label": "Corr max 0.7 skip",
"corr_max": 0.7,
"corr_action": "skip",
},
{
"id": "a5_corr08_skip",
"group": "a5",
"label": "Corr max 0.8 skip",
"corr_max": 0.8,
"corr_action": "skip",
},
{
"id": "a5_corr06_half",
"group": "a5",
"label": "Corr max 0.6 half-size",
"corr_max": 0.6,
"corr_action": "half_size",
},
{
"id": "a5_corr07_half",
"group": "a5",
"label": "Corr max 0.7 half-size",
"corr_max": 0.7,
"corr_action": "half_size",
},
{
"id": "a5_corr08_half",
"group": "a5",
"label": "Corr max 0.8 half-size",
"corr_max": 0.8,
"corr_action": "half_size",
},
)
PRE_REGISTERED_ARM_IDS = tuple(arm["id"] for arm in PRE_REGISTERED_ARMS)
PRE_REGISTERED_N_TRIALS = len(PRE_REGISTERED_ARMS)
def _sqlite_url(path: Path) -> str:
return f"sqlite+aiosqlite:///{path.resolve().as_posix()}"
def _period_percentiles(
observations: list[dict], value_key: str
) -> dict[tuple[str, str], float]:
by_period: dict[tuple, list[dict]] = {}
for row in observations:
if row.get(value_key) is None:
continue
period = tuple(row["ranking_period"])
by_period.setdefault(period, []).append(row)
result: dict[tuple[str, str], float] = {}
for group in by_period.values():
ordered = sorted(
group, key=lambda row: (float(row[value_key]), str(row["symbol"]))
)
denominator = len(ordered) - 1
for rank, row in enumerate(ordered):
result[(str(row["symbol"]), str(row["date"]))] = round(
rank / denominator * 100.0 if denominator > 0 else 100.0,
2,
)
return result
def _live_universe_rank_map(
observations: list[dict],
benchmark_closes: dict[date, float],
momentum_weight: float,
) -> dict[tuple[str, str], dict[str, float | None]]:
raw_pct = _period_percentiles(observations, "momentum")
residual_pct = _period_percentiles(observations, "residual_momentum")
vol_pct = _period_percentiles(observations, "vol_6m")
benchmark_ords = sorted(value.toordinal() for value in benchmark_closes)
residual_start_ord = benchmark_ords[251] if len(benchmark_ords) >= 252 else None
ranks: dict[tuple[str, str], dict[str, float | None]] = {}
for row in observations:
identity = (str(row["symbol"]), str(row["date"]))
asof_ord = date.fromisoformat(identity[1]).toordinal()
momentum_pct = (
residual_pct.get(identity)
if residual_start_ord is not None and asof_ord >= residual_start_ord
else raw_pct.get(identity)
)
volatility_pct = vol_pct.get(identity)
strategy_rank = (
round(
momentum_pct * momentum_weight
+ volatility_pct * (1.0 - momentum_weight),
2,
)
if momentum_pct is not None and volatility_pct is not None
else momentum_pct
)
ranks[identity] = {
"momentum_percentile": momentum_pct,
"volatility_percentile": volatility_pct,
"strategy_rank": strategy_rank,
}
return ranks
def _parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter,
)
parser.add_argument("snapshot", help="SQLite backtest snapshot.")
parser.add_argument("--workers", type=int, default=6)
parser.add_argument(
"--allow-spawn",
action="store_true",
help="Allow spawn multiprocessing (needed on Windows).",
)
parser.add_argument("--out", default=None, help="JSON report path.")
parser.add_argument(
"--candidate-cache",
default=None,
help="Optional pickle cache for the daily qualified candidate set.",
)
parser.add_argument(
"--validation-split",
default="2024-07-01",
help="Train/validation entry split (YYYY-MM-DD). Validation = entries on/after.",
)
parser.add_argument(
"--only",
default=None,
help="Comma-separated arm groups or ids to run (e.g. a2,a3 or a0_control,a4_next_open).",
)
parser.add_argument(
"--skip",
default=None,
help="Comma-separated arm groups or ids to skip.",
)
parser.add_argument("--quiet", action="store_true")
parser.add_argument(
"--cadence",
choices=("daily", "weekly"),
default="daily",
help="Candidate replay cadence. Daily matches the re-entry matrix production arm.",
)
return parser.parse_args()
def _default_output_path() -> Path:
stamp = datetime.now().strftime("%Y%m%d-%H%M%S")
return Path("reports") / f"research-matrix-{stamp}.json"
def _parse_selector(raw: str | None) -> set[str] | None:
if raw is None or not raw.strip():
return None
return {part.strip().lower() for part in raw.split(",") if part.strip()}
def _arm_selected(arm: dict[str, Any], only: set[str] | None, skip: set[str] | None) -> bool:
arm_id = str(arm["id"]).lower()
group = str(arm["group"]).lower()
if skip and (arm_id in skip or group in skip):
return False
if only is None:
return True
return arm_id in only or group in only
def _write_checkpoint(path: Path, report: dict) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
tmp = path.with_suffix(path.suffix + ".tmp")
tmp.write_text(json.dumps(report, indent=2, default=str), encoding="utf-8")
tmp.replace(path)
md_path = path.with_suffix(".md")
md_path.write_text(_markdown_table(report), encoding="utf-8")
def _markdown_table(report: dict) -> str:
lines = [
f"# Research matrix — {report.get('generated_at', '')}",
"",
f"Validation split: **{report.get('validation_split')}**. "
f"Pre-registered N for DSR: **{report.get('n_trials')}**.",
"",
"## Promotion rule",
"",
report.get("promotion_rule", ""),
"",
"## Arms",
"",
"| arm | window | Sharpe | SE | PSR | DSR | CAGR | MaxDD | Calmar | trades | avg scalar |",
"|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|",
]
for arm in report.get("arms") or []:
for window_row in arm.get("windows") or []:
lines.append(
"| {arm} | {window} | {sharpe} | {se} | {psr} | {dsr} | {cagr} | {dd} | {calmar} | {trades} | {scalar} |".format(
arm=arm.get("id"),
window=window_row.get("window"),
sharpe=_fmt(window_row.get("sharpe")),
se=_fmt(window_row.get("sharpe_se")),
psr=_fmt(window_row.get("psr")),
dsr=_fmt(window_row.get("dsr")),
cagr=_fmt(window_row.get("cagr_pct")),
dd=_fmt(window_row.get("max_drawdown_pct")),
calmar=_fmt(window_row.get("calmar")),
trades=_fmt(window_row.get("trades")),
scalar=_fmt(window_row.get("avg_vol_scalar")),
)
)
promo = report.get("promotion") or {}
lines.extend(["", "## Promotion decisions", ""])
if not promo:
lines.append("_No arms graded yet._")
else:
for arm_id, decision in promo.items():
lines.append(
f"- **{arm_id}**: {'PROMOTE' if decision.get('promote') else 'reject'}"
f"{decision.get('reason')}"
)
lines.append("")
return "\n".join(lines)
def _fmt(value: Any) -> str:
if value is None:
return ""
if isinstance(value, float):
return f"{value:.3g}"
return str(value)
def _grade_promotion(control: dict, arm: dict, se_ref: float | None) -> dict:
"""Apply the pre-registered promotion rule. control/arm are arm result dicts."""
c_val = _window(control, "validation")
a_val = _window(arm, "validation")
c_train = _window(control, "train")
a_train = _window(arm, "train")
if not c_val or not a_val or not c_train or not a_train:
return {"promote": False, "reason": "missing train/validation rows"}
c_s = c_val.get("sharpe")
a_s = a_val.get("sharpe")
c_dd = c_val.get("max_drawdown_pct")
a_dd = a_val.get("max_drawdown_pct")
c_ts = c_train.get("sharpe")
a_ts = a_train.get("sharpe")
if None in (c_s, a_s, c_dd, a_dd, c_ts, a_ts):
return {"promote": False, "reason": "missing Sharpe/DD on a required window"}
delta = float(a_s) - float(c_s)
se = se_ref
if se is None:
se = a_val.get("sharpe_se") or c_val.get("sharpe_se")
exceeds_1se = se is not None and abs(delta) > float(se)
checks = {
"validation_sharpe_ge_control": float(a_s) >= float(c_s),
"validation_dd_not_worse_by_2pp": float(a_dd) <= float(c_dd) + 2.0,
"train_sharpe_not_worse": float(a_ts) >= float(c_ts),
"delta_exceeds_1se": exceeds_1se,
"validation_sharpe_delta": round(delta, 4),
"se_used": se,
}
promote = (
checks["validation_sharpe_ge_control"]
and checks["validation_dd_not_worse_by_2pp"]
and checks["train_sharpe_not_worse"]
)
if promote:
reason = (
f"validation Sharpe {a_s} ≥ control {c_s}; "
f"DD {a_dd} within +2pp of {c_dd}; train Sharpe {a_ts}{c_ts}"
)
if not exceeds_1se:
reason += " (delta ≤ 1 SE — distinguishable noise bar not cleared)"
else:
reason += " (delta > 1 SE)"
else:
failed = [k for k, v in checks.items() if k.startswith(("validation", "train")) and v is False]
reason = "failed: " + ", ".join(failed) if failed else "failed promotion checks"
return {"promote": promote, "reason": reason, "checks": checks}
def _window(arm_result: dict, name: str) -> dict | None:
for row in arm_result.get("windows") or []:
if row.get("window") == name:
return row
return None
def _assert_calendar_truncation(sim: dict, hold_days: int, fill_mode: str) -> None:
"""Guard against trailing flat-cash after the last resolvable signal."""
start = sim.get("start_date")
end = sim.get("end_date")
if not start or not end:
return
# Soft check: book span should not massively exceed hold window beyond data needs.
# Hard assert lives on entry-end vs sim end when trade_details present.
details = sim.get("trade_details") or []
if not details:
return
last_entry = max(date.fromisoformat(t["entry_date"]) for t in details)
sim_end = date.fromisoformat(str(end))
pad = hold_days + (1 if fill_mode == "next_open" else 0)
# Allow calendar days ≈ trading-day pad with weekend slack (2×).
max_slack_days = pad * 2 + 5
if (sim_end - last_entry).days > max_slack_days:
raise AssertionError(
f"calendar truncation failed: last entry {last_entry} but sim end "
f"{sim_end} (hold_days={hold_days}, fill_mode={fill_mode})"
)
async def _main() -> None:
args = _parse_args()
snapshot = Path(args.snapshot)
if not snapshot.exists():
raise SystemExit(f"Snapshot not found: {snapshot}")
if args.workers < 1:
raise SystemExit("--workers must be positive")
only = _parse_selector(args.only)
skip = _parse_selector(args.skip)
validation_split = date.fromisoformat(args.validation_split)
out_path = Path(args.out) if args.out else _default_output_path()
os.environ["BACKTEST_SNAPSHOT_OFFLINE"] = "1"
if args.allow_spawn:
os.environ["BACKTEST_ALLOW_SPAWN"] = "1"
from app.models.ticker import Ticker
from app.services import backtest_service as bt
from app.services.admin_service import get_activation_config
from app.services.paper_trade_service import get_exit_policy
from app.services.recommendation_service import get_recommendation_config
db_engine = create_async_engine(_sqlite_url(snapshot), pool_pre_ping=True)
Session = async_sessionmaker(db_engine, class_=AsyncSession, expire_on_commit=False)
try:
async with Session() as db:
recommendation_config = await get_recommendation_config(db)
activation = await get_activation_config(db)
exit_config = await get_exit_policy(db)
benchmark_closes = await bt._load_benchmark_closes_for_backtest(
db, days=None, refresh=False
)
ticker_result = await db.execute(select(Ticker).order_by(Ticker.symbol))
symbols = [t.symbol for t in ticker_result.scalars().all()]
prices: dict[str, tuple] = {}
for index, symbol in enumerate(symbols, 1):
columns = await bt._fetch_columns(db, symbol)
if columns is not None:
prices[symbol] = columns
if not args.quiet and index % 50 == 0:
print(f"loaded prices: {index}/{len(symbols)}", flush=True)
finally:
await db_engine.dispose()
if not prices:
raise SystemExit("No price columns loaded from snapshot")
snapshot_stat = snapshot.stat()
cache_key = {
"version": CACHE_VERSION,
"snapshot": str(snapshot.resolve()),
"snapshot_size": snapshot_stat.st_size,
"snapshot_mtime_ns": snapshot_stat.st_mtime_ns,
"cadence": args.cadence,
"target_model": "production_gtl",
}
cache_path = Path(args.candidate_cache) if args.candidate_cache else None
qualified: list[dict] | None = None
entry_candidate_count = 0
fip_signal_eval: list[dict] | None = None
if cache_path is not None and cache_path.exists():
with cache_path.open("rb") as handle:
cached = pickle.load(handle) # noqa: S301 - trusted local cache
if cached.get("key") == cache_key:
qualified = list(cached["qualified_candidates"])
entry_candidate_count = int(cached["entry_candidate_count"])
fip_signal_eval = cached.get("fip_signal_eval")
if not args.quiet:
print(f"loaded candidate cache: {cache_path}", flush=True)
elif not args.quiet:
print(f"candidate cache mismatch; rebuilding: {cache_path}", flush=True)
if qualified is None:
replay_start = date(1900, 1, 1)
workers = max(1, min(int(args.workers), max(1, multiprocessing.cpu_count() - 1)))
context = bt._mp_context() or multiprocessing.get_context("spawn")
replay_rows: list[dict] = []
with ProcessPoolExecutor(max_workers=workers, mp_context=context) as pool:
futures = {
pool.submit(
bt._replay_candidates_for_period,
symbol,
columns,
recommendation_config,
activation,
benchmark_closes,
replay_start,
args.cadence,
True,
True,
): symbol
for symbol, columns in prices.items()
}
for index, future in enumerate(as_completed(futures), 1):
replay_rows.extend(future.result())
if not args.quiet and index % 25 == 0:
print(f"replay: {index}/{len(futures)} tickers", flush=True)
setup_candidates = [row for row in replay_rows if not row.get("_rank_only")]
rank_observations = [
row for row in replay_rows if row.get("_universe_rank_observation")
]
entry_candidate_count = len(setup_candidates)
# Live-universe ranking (same semantics as run_daily_reentry_matrix).
live_ranks = _live_universe_rank_map(
rank_observations,
benchmark_closes,
bt.STRATEGY_RANK_MOMENTUM_WEIGHT,
)
threshold = float(activation.get("min_momentum_percentile", 80.0))
qualified = []
for setup in setup_candidates:
if setup.get("direction") != "long":
continue
candidate = {
key: value
for key, value in setup.items()
if not key.startswith("_universe_")
}
identity = (str(setup["symbol"]), str(setup["date"]))
rank = live_ranks.get(identity)
if rank is None:
continue
candidate[bt.PRODUCTION_PERCENTILE_KEY] = rank["momentum_percentile"]
candidate[bt.VOL_PERCENTILE_KEY] = rank["volatility_percentile"]
candidate[bt.RESIDUAL_HIGH_VOL_BLEND_80_20_KEY] = rank["strategy_rank"]
candidate["qualified"] = bt._momentum_qualifies(candidate, threshold)
if candidate["qualified"]:
qualified.append(candidate)
# fip_id fingerprint via the shared weekly signal harness.
collected: dict = {}
for symbol, columns in prices.items():
# Rebuild minimal records for signal eval from column arrays.
ords, _o, highs, _l, closes, _v = columns
records = [
type("R", (), {"date": date.fromordinal(int(ords[i])), "close": closes[i], "high": highs[i]})()
for i in range(len(ords))
]
series = bt._signal_series(records, benchmark_closes)
for name, weeks in series.items():
bucket = collected.setdefault(name, {})
for week_key, pairs in weeks.items():
bucket.setdefault(week_key, []).extend(pairs)
fip_signal_eval = [
row for row in bt._signal_evaluation(collected) if row.get("signal") == "fip_id"
]
if cache_path is not None:
cache_path.parent.mkdir(parents=True, exist_ok=True)
with cache_path.open("wb") as handle:
pickle.dump(
{
"key": cache_key,
"entry_candidate_count": entry_candidate_count,
"qualified_candidates": qualified,
"fip_signal_eval": fip_signal_eval,
},
handle,
protocol=pickle.HIGHEST_PROTOCOL,
)
if not args.quiet:
print(f"wrote candidate cache: {cache_path}", flush=True)
if not qualified:
raise SystemExit("No qualified long candidates after replay")
strategy = next(s for s in bt.PORTFOLIO_MONITOR_STRATEGIES if s.get("is_production"))
entry_config = bt._entry_variant_config(str(strategy["entry_variant"]))
if entry_config is None:
raise RuntimeError("Production entry configuration missing")
ranking_key = str(entry_config.get("ranking_key") or entry_config["percentile_key"])
exit_policy = bt.LIVE_EXIT_MODE_TO_SIM.get(
str(exit_config.get("mode", "atr_trailing")), "atr_trail3"
)
default_hold = int(exit_config.get("hold_days", 30))
trail_multiplier = float(exit_config.get("atr_multiplier", bt.ATR_TRAIL_MULTIPLIER))
risk_per_trade = float(entry_config["risk_per_trade"])
max_positions = int(entry_config["max_positions"])
post_stop_reentry_fn = bt._make_gate_reset_reentry_fn(
qualified,
prices,
cadence=args.cadence,
ranking_key=ranking_key,
)
selected_arms = [
arm for arm in PRE_REGISTERED_ARMS if _arm_selected(arm, only, skip)
]
# Always include control when grading promotions for non-control arms.
if selected_arms and not any(a["id"] == "a0_control" for a in selected_arms):
if only is None or "a0" in (only or set()) or "a0_control" in (only or set()):
pass
else:
# Force control into the run for comparison baselines.
control_arm = next(a for a in PRE_REGISTERED_ARMS if a["id"] == "a0_control")
selected_arms = [control_arm, *selected_arms]
if not selected_arms:
raise SystemExit("No arms selected — check --only / --skip")
report: dict[str, Any] = {
"generated_at": datetime.now().isoformat(),
"snapshot": str(snapshot.resolve()),
"cadence": args.cadence,
"validation_split": validation_split.isoformat(),
"n_trials": PRE_REGISTERED_N_TRIALS,
"pre_registered_arm_ids": list(PRE_REGISTERED_ARM_IDS),
"selected_arm_ids": [a["id"] for a in selected_arms],
"entry_candidate_count": entry_candidate_count,
"qualified_longs": len(qualified),
"promotion_rule": (
"Promote only if validation Sharpe ≥ control, validation DD not worse "
"by >2pp, and train Sharpe not worse. Always report whether validation "
"Sharpe delta exceeds 1 SE (expect most will not)."
),
"fip_id_fingerprint": fip_signal_eval,
"arms": [],
"promotion": {},
}
_write_checkpoint(out_path, report)
control_result: dict | None = None
def run_arm(arm: dict[str, Any]) -> dict:
hold_days = int(arm.get("hold_days", default_hold))
fill_mode = str(arm.get("fill_mode", bt.FILL_MODE_CLOSE))
windows: list[dict] = []
for window_name, start, end in (
("train", None, validation_split),
("validation", validation_split, None),
("full", None, None),
):
sim = bt._simulate_portfolio(
qualified,
prices,
benchmark_closes,
exit_policy,
hold_days,
ranking_key=ranking_key,
max_positions=max_positions,
risk_per_trade=risk_per_trade,
atr_trail_multiplier=trail_multiplier,
post_stop_reentry_fn=post_stop_reentry_fn,
start_date=start,
end_date=end,
fill_mode=fill_mode,
vol_target=arm.get("vol_target"),
vol_lookback=int(arm.get("vol_lookback", bt.VOL_TARGET_LOOKBACK_HEADLINE)),
vol_clamp=tuple(arm.get("vol_clamp", bt.VOL_TARGET_CLAMP_HEADLINE)),
corr_max=arm.get("corr_max"),
corr_action=str(arm.get("corr_action", "skip")),
include_trades=True,
)
if sim is None:
windows.append({"window": window_name, "error": "no_trades"})
continue
_assert_calendar_truncation(sim, hold_days, fill_mode)
dsr = bt.deflated_sharpe_ratio(
sim.get("sharpe"),
sim.get("sharpe_se"),
PRE_REGISTERED_N_TRIALS,
n_returns=sim.get("n_returns"),
return_skew=sim.get("return_skew"),
return_kurtosis=sim.get("return_kurtosis"),
)
# Drop heavy trade lists from the checkpointed JSON.
sim.pop("trade_details", None)
sim.pop("equity_curve", None)
sim.pop("benchmark_curve", None)
sim.pop("reentry_events", None)
windows.append({"window": window_name, "dsr": dsr, **sim})
return {
"id": arm["id"],
"group": arm["group"],
"label": arm["label"],
"config": {
key: arm[key]
for key in arm
if key not in {"id", "group", "label"}
},
"windows": windows,
}
for arm in selected_arms:
if not args.quiet:
print(f"running arm {arm['id']} ...", flush=True)
result = run_arm(arm)
report["arms"].append(result)
if arm["id"] == "a0_control":
control_result = result
elif control_result is not None:
report["promotion"][arm["id"]] = _grade_promotion(
control_result, result, None
)
_write_checkpoint(out_path, report)
if not args.quiet:
val = _window(result, "validation") or {}
print(
f" done {arm['id']}: validation Sharpe={val.get('sharpe')} "
f"DD={val.get('max_drawdown_pct')} trades={val.get('trades')}",
flush=True,
)
# Re-grade all arms once control is known (handles --only without ordering issues).
if control_result is not None:
for result in report["arms"]:
if result["id"] == "a0_control":
continue
report["promotion"][result["id"]] = _grade_promotion(
control_result, result, None
)
_write_checkpoint(out_path, report)
if not args.quiet:
print(f"wrote {out_path}", flush=True)
print(f"wrote {out_path.with_suffix('.md')}", flush=True)
fip = (fip_signal_eval or [{}])[0] if fip_signal_eval else {}
if fip:
print(
f"fip_id fingerprint: mean_ic={fip.get('mean_ic')} "
f"t={fip.get('ic_t_stat')} (target ≈ -0.045 / -2.9)",
flush=True,
)
if __name__ == "__main__":
asyncio.run(_main())