NeuroHealth1 / app.py
Mikecode123's picture
Update app.py
391ef4f verified
Raw
History Blame Contribute Delete
35.9 kB
#!/usr/bin/env python3
"""
100optimization.py — safe Hugging Face CRT strategy optimizer.
This script is designed for a Hugging Face Space. It:
1. Calls load_dataset("Mikecode123/volatility_100_index") with remote dataset
code disabled.
2. Locates volatility_100_index.zip in the dataset snapshot.
3. Inspects the archive and extracts CSV files only. Scripts, executables,
symlinks, traversal paths, and oversized archives are rejected.
4. Loads M1.csv, M5.csv, M15.csv, M30.csv, H1.csv, and H4.csv case-insensitively.
5. Implements a clearly documented Candle Range Theory (CRT) rule:
- use a completed H1 or H4 reference candle;
- wait for a sweep of its high/low;
- require a close back inside the reference range;
- require the next-bar confirmation through the sweep candle extreme;
- enter at confirmation close;
- target the opposite edge of the reference range and evaluate staged
ATR-based TP/SL outcomes.
6. Uses a bounded deterministic candidate search. The supervisor model can
select only from supplied candidate IDs; it cannot generate code or change
the evaluation rules.
7. Saves every new best result and milestone result (20%, 30%, 80%) to the
Space checkpoint directory, then creates a downloadable ZIP.
There is no guarantee that 80% accuracy is achievable. The target is a stopping
criterion, not a promise. Validation is used for optimization; the final test
period remains untouched until the end.
Supervisor model (pinned and approved):
google/flan-t5-small
revision=0fc9ddf78a1e988dac52e2dac162b0ede4fd74ab
trust_remote_code=False
Expected Space requirements:
datasets
huggingface_hub
transformers
torch
pandas
numpy
Research sources for the CRT specification:
https://innercircletrader.net/tutorials/candle-range-theory-crt/
https://tradingwyckoff.com/en/crt/
The public CRT descriptions are educational retail sources, not peer-reviewed
proof of profitability. The implementation therefore reports validation and
test results separately and records coverage beside accuracy.
"""
from __future__ import annotations
import gc
import itertools
import json
import math
import os
import random
import re
import shutil
import stat
import time
import zipfile
from dataclasses import asdict, dataclass
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Optional
import numpy as np
import pandas as pd
try:
from datasets import load_dataset
except ImportError as exc: # pragma: no cover - dependency supplied by the Space
raise RuntimeError(
"Install the Space requirements first: datasets, huggingface_hub, "
"transformers, torch, pandas, numpy"
) from exc
# ---------------------------------------------------------------------------
# Configuration
# ---------------------------------------------------------------------------
DATASET_ID = os.environ.get("HF_DATASET_ID", "Mikecode123/volatility_100_index")
SUPERVISOR_MODEL_ID = "google/flan-t5-small"
SUPERVISOR_REVISION = "0fc9ddf78a1e988dac52e2dac162b0ede4fd74ab"
TF_MINUTES = {
"M1": 1,
"M5": 5,
"M15": 15,
"M30": 30,
"H1": 60,
"H4": 240,
}
CLASS_NAMES = {
0: "NO_TRADE",
1: "SL",
2: "BE",
3: "TP1",
4: "TP2",
5: "TP3",
}
LOCAL_R_VALUE = np.array([0.0, -1.0, 0.0, 2.0, 4.0, 6.0], dtype=np.float32)
LOCAL_RANK = np.array([1, 0, 2, 3, 4, 5], dtype=np.int8)
@dataclass
class Config:
dataset_id: str = DATASET_ID
seed: int = int(os.environ.get("OPTIMIZER_SEED", "42"))
target_accuracy: float = float(os.environ.get("TARGET_ACCURACY", "0.80"))
max_trials: int = int(os.environ.get("MAX_TRIALS", "30"))
min_trades: int = int(os.environ.get("MIN_TRADES", "100"))
min_coverage: float = float(os.environ.get("MIN_COVERAGE", "0.001"))
max_rows: int = int(os.environ.get("MAX_ROWS", "0"))
reference_timeframe: str = os.environ.get("CRT_REFERENCE_TF", "H1").upper()
confirm_bars: int = int(os.environ.get("CRT_CONFIRM_BARS", "1"))
atr_period: int = int(os.environ.get("ATR_PERIOD", "14"))
default_sl_atr_mult: float = float(os.environ.get("SL_ATR_MULT", "1.0"))
default_horizon: int = int(os.environ.get("MAX_HORIZON", "60"))
checkpoint_dir: Path = Path(
os.environ.get("CHECKPOINT_DIR", "/data/100optimization_checkpoints")
)
data_dir: Path = Path(
os.environ.get("DATA_EXTRACT_DIR", "/tmp/volatility_100_index_data")
)
CFG = Config()
# ---------------------------------------------------------------------------
# Logging and filesystem helpers
# ---------------------------------------------------------------------------
def log(message: str, *args: Any, level: str = "INFO") -> None:
if args:
message = message.format(*args)
stamp = datetime.now().strftime("%H:%M:%S")
print(f"[{stamp}] [{level}] {message}", flush=True)
def utc_now() -> str:
return datetime.now(timezone.utc).isoformat()
def ensure_checkpoint_dir() -> Path:
try:
CFG.checkpoint_dir.mkdir(parents=True, exist_ok=True)
return CFG.checkpoint_dir
except PermissionError:
fallback = Path("./100optimization_checkpoints")
fallback.mkdir(parents=True, exist_ok=True)
log(
"Cannot write to {}; using {} instead",
CFG.checkpoint_dir,
fallback,
level="WARN",
)
CFG.checkpoint_dir = fallback
return fallback
# ---------------------------------------------------------------------------
# Hugging Face dataset discovery and safe archive handling
# ---------------------------------------------------------------------------
def _snapshot_dataset_files() -> Path:
"""Download only dataset files, never executable model or code files."""
from huggingface_hub import snapshot_download
return Path(
snapshot_download(
repo_id=CFG.dataset_id,
repo_type="dataset",
allow_patterns=["*.zip", "*.ZIP", "*.csv", "*.CSV"],
)
)
def locate_zip(snapshot_dir: Path) -> Path:
candidates = sorted(
p for p in snapshot_dir.rglob("*")
if p.is_file() and p.suffix.lower() == ".zip"
)
if not candidates:
raise FileNotFoundError(
f"No ZIP file was found in the dataset snapshot {snapshot_dir}."
)
exact = [p for p in candidates if p.name.lower() == "volatility_100_index.zip"]
if exact:
return exact[0]
if len(candidates) == 1:
log("Using the only ZIP in the dataset snapshot: {}", candidates[0], level="WARN")
return candidates[0]
raise FileNotFoundError(
"Multiple ZIP files were found and none was named "
f"volatility_100_index.zip: {candidates}"
)
def inspect_zip(zip_path: Path) -> list[zipfile.ZipInfo]:
"""Inspect the archive before extraction; allow data CSVs only."""
max_entries = 2_000
max_file_size = 25 * 1024 * 1024
max_total_size = 250 * 1024 * 1024
with zipfile.ZipFile(zip_path, "r") as archive:
infos = archive.infolist()
if len(infos) > max_entries:
raise ValueError(f"Archive contains too many entries: {len(infos)}")
total_size = 0
files: list[zipfile.ZipInfo] = []
log("Inspecting archive {}", zip_path)
for info in infos:
name = Path(info.filename.replace("\\", "/"))
if name.is_absolute() or ".." in name.parts:
raise ValueError(f"Unsafe archive path rejected: {info.filename}")
file_mode = (info.external_attr >> 16) & 0o170000
if file_mode == stat.S_IFLNK:
raise ValueError(f"Symlink rejected: {info.filename}")
if info.is_dir():
log(" [directory] {}", info.filename)
continue
if name.suffix.lower() != ".csv":
raise ValueError(
f"Non-CSV/script/configuration member rejected: {info.filename}. "
"This optimizer accepts a data-only archive."
)
if info.file_size > max_file_size:
raise ValueError(f"Archive member is too large: {info.filename}")
total_size += info.file_size
if total_size > max_total_size:
raise ValueError("Archive expanded size exceeds safety limit")
log(" [csv] {} ({:,} bytes)", info.filename, info.file_size)
files.append(info)
if not files:
raise ValueError("Archive contains no CSV files")
return files
def safe_extract_zip(zip_path: Path, destination: Path) -> Path:
infos = inspect_zip(zip_path)
if destination.exists():
shutil.rmtree(destination)
destination.mkdir(parents=True, exist_ok=True)
root = destination.resolve()
with zipfile.ZipFile(zip_path, "r") as archive:
for info in infos:
target = (destination / info.filename).resolve()
if target != root and root not in target.parents:
raise ValueError(f"Extraction escaped destination: {info.filename}")
target.parent.mkdir(parents=True, exist_ok=True)
with archive.open(info, "r") as source, target.open("wb") as sink:
shutil.copyfileobj(source, sink, length=1024 * 1024)
return destination
def find_timeframe_file(directory: Path, timeframe: str) -> Optional[Path]:
wanted = f"{timeframe}.csv".lower()
for candidate in directory.iterdir():
if candidate.is_file() and candidate.name.lower() == wanted:
return candidate
return None
def find_data_directory(root: Path) -> Path:
for m1 in sorted(root.rglob("*.csv")):
if m1.is_file() and m1.stem.lower() == "m1":
required = ("M1", "M5", "M15", "M30", "H1", "H4")
missing = [tf for tf in required if find_timeframe_file(m1.parent, tf) is None]
if not missing:
return m1.parent
log("Ignoring {} because it is missing {}", m1.parent, missing, level="WARN")
raise FileNotFoundError(
f"Could not find a directory containing M1/M5/M15/M30/H1/H4 CSVs below {root}"
)
def load_dataset_archive() -> Path:
"""Call load_dataset as requested, then use the ZIP snapshot safely."""
try:
loaded = load_dataset(CFG.dataset_id, trust_remote_code=False)
if hasattr(loaded, "keys"):
log("load_dataset opened splits: {}", list(loaded.keys()))
else:
log("load_dataset opened {} rows", len(loaded))
except Exception as exc:
# A repository containing only a ZIP may not have a Datasets builder.
# This fallback downloads data files only and still inspects the ZIP
# before extraction.
log(
"load_dataset could not parse the ZIP-only repository: {}. "
"Using the data-file snapshot fallback.",
exc,
level="WARN",
)
snapshot = _snapshot_dataset_files()
archive = locate_zip(snapshot)
log("Dataset archive: {}", archive)
return archive
# ---------------------------------------------------------------------------
# OHLC loading and CRT feature construction
# ---------------------------------------------------------------------------
COLUMN_ALIASES = {
"timestamp": ["timestamp", "time", "date", "datetime", "open_time"],
"open": ["open", "o"],
"high": ["high", "h"],
"low": ["low", "l"],
"close": ["close", "c", "adj_close"],
"volume": ["volume", "vol", "v", "tick_volume"],
}
def resolve_column(columns: list[str], aliases: list[str]) -> Optional[str]:
mapping = {column.lower(): column for column in columns}
for alias in aliases:
if alias in mapping:
return mapping[alias]
return None
def load_ohlcv(path: Path) -> pd.DataFrame:
header = pd.read_csv(path, nrows=0)
columns = list(header.columns)
resolved = {
key: resolve_column(columns, aliases)
for key, aliases in COLUMN_ALIASES.items()
}
required = ("timestamp", "open", "high", "low", "close")
missing = [key for key in required if resolved[key] is None]
if missing:
raise ValueError(f"{path} is missing required columns: {missing}")
usecols = [resolved[key] for key in required]
if resolved["volume"] is not None:
usecols.append(resolved["volume"])
dtype = {
resolved[key]: np.float32
for key in ("open", "high", "low", "close")
}
if resolved["volume"] is not None:
dtype[resolved["volume"]] = np.float32
raw = pd.read_csv(
path,
usecols=list(dict.fromkeys(usecols)),
dtype=dtype,
parse_dates=[resolved["timestamp"]],
)
out = pd.DataFrame({
"timestamp": raw[resolved["timestamp"]],
"open": raw[resolved["open"]],
"high": raw[resolved["high"]],
"low": raw[resolved["low"]],
"close": raw[resolved["close"]],
})
out["volume"] = (
raw[resolved["volume"]]
if resolved["volume"] is not None
else np.float32(0.0)
)
out = (
out.dropna(subset=["timestamp", "open", "high", "low", "close"])
.sort_values("timestamp")
.drop_duplicates("timestamp", keep="last")
.reset_index(drop=True)
)
invalid = (
(out["high"] < out["low"])
| (out[["open", "high", "low", "close"]] <= 0).any(axis=1)
)
if invalid.any():
log("{}: dropping {} invalid rows", path.name, int(invalid.sum()), level="WARN")
out = out.loc[~invalid].reset_index(drop=True)
return out
def compute_atr(df: pd.DataFrame, period: int) -> pd.Series:
previous_close = df["close"].shift(1)
true_range = pd.concat(
[
df["high"] - df["low"],
(df["high"] - previous_close).abs(),
(df["low"] - previous_close).abs(),
],
axis=1,
).max(axis=1)
return true_range.rolling(period, min_periods=period).mean()
def closed_reference_features(
base: pd.DataFrame,
reference: pd.DataFrame,
timeframe: str,
) -> pd.DataFrame:
"""Attach only fully closed reference-candle levels to M1 rows."""
ref = reference[["timestamp", "open", "high", "low", "close"]].copy()
ref["timestamp"] = ref["timestamp"] + pd.Timedelta(
minutes=TF_MINUTES[timeframe]
)
ref = ref.rename(columns={
"open": "crt_open",
"high": "crt_high",
"low": "crt_low",
"close": "crt_close",
})
return pd.merge_asof(
base.sort_values("timestamp"),
ref.sort_values("timestamp"),
on="timestamp",
direction="backward",
)
# ---------------------------------------------------------------------------
# Staged outcomes used for CRT backtesting
# ---------------------------------------------------------------------------
def simulate_direction(
close: np.ndarray,
high: np.ndarray,
low: np.ndarray,
risk: np.ndarray,
direction: int,
horizon: int,
) -> tuple[np.ndarray, np.ndarray]:
"""Simulate SL -> BE -> TP1-lock -> TP3 for one direction."""
n = len(close)
sign = 1.0 if direction > 0 else -1.0
entry = close
stop = entry - sign * risk
tp1 = entry + sign * risk * 2.0
tp2 = entry + sign * risk * 4.0
tp3 = entry + sign * risk * 6.0
lock = tp1
stage = np.zeros(n, dtype=np.int8)
resolved = np.zeros(n, dtype=bool)
outcome = np.zeros(n, dtype=np.int8)
valid = np.isfinite(risk)
for step in range(1, horizon + 1):
owners = np.arange(n)
future = owners + step
active = (~resolved) & valid & (future < n)
if not active.any():
continue
owners = owners[active]
future = future[active]
snapshot = stage[owners].copy()
for stage_id, favorable, adverse in (
(0, tp1, stop),
(1, tp2, entry),
(2, tp3, lock),
):
mask = snapshot == stage_id
if not mask.any():
continue
rows = owners[mask]
hi = high[future[mask]]
lo = low[future[mask]]
if direction > 0:
touched_favorable = hi >= favorable[rows]
touched_adverse = lo <= adverse[rows]
else:
touched_favorable = lo <= favorable[rows]
touched_adverse = hi >= adverse[rows]
adverse_rows = rows[touched_adverse]
favorable_rows = rows[touched_favorable & ~touched_adverse]
if stage_id == 0:
outcome[adverse_rows] = 1
resolved[adverse_rows] = True
stage[favorable_rows] = 1
elif stage_id == 1:
outcome[adverse_rows] = 2
resolved[adverse_rows] = True
stage[favorable_rows] = 2
else:
outcome[adverse_rows] = 3
resolved[adverse_rows] = True
final_rows = rows[touched_favorable & ~touched_adverse]
outcome[final_rows] = 5
resolved[final_rows] = True
open_rows = (~resolved) & valid
outcome[open_rows & (stage == 0)] = 0
outcome[open_rows & (stage == 1)] = 3
outcome[open_rows & (stage == 2)] = 4
return outcome, LOCAL_R_VALUE[outcome]
def build_outcomes(
m1: pd.DataFrame,
atr_period: int,
sl_atr_mult: float,
horizon: int,
) -> dict[str, np.ndarray]:
atr = compute_atr(m1, atr_period).to_numpy(dtype=np.float32)
risk = np.float32(sl_atr_mult) * atr
close = m1["close"].to_numpy(dtype=np.float32)
high = m1["high"].to_numpy(dtype=np.float32)
low = m1["low"].to_numpy(dtype=np.float32)
long_outcome, long_r = simulate_direction(close, high, low, risk, 1, horizon)
short_outcome, short_r = simulate_direction(close, high, low, risk, -1, horizon)
long_rank = LOCAL_RANK[long_outcome]
short_rank = LOCAL_RANK[short_outcome]
choose_long = (long_rank > short_rank) | (
(long_rank == short_rank) & (long_r >= short_r)
)
chosen_outcome = np.where(choose_long, long_outcome, short_outcome).astype(np.int8)
chosen_direction = np.where(
chosen_outcome == 0,
0,
np.where(choose_long, 1, -1),
).astype(np.int8)
return {
"long_outcome": long_outcome,
"short_outcome": short_outcome,
"long_r": long_r,
"short_r": short_r,
"chosen_outcome": chosen_outcome,
"chosen_direction": chosen_direction,
}
# ---------------------------------------------------------------------------
# CRT signals and metrics
# ---------------------------------------------------------------------------
@dataclass(frozen=True)
class Candidate:
candidate_id: int
reference_timeframe: str
min_sweep_atr: float
use_reference_bias: bool
confirmation_window: int
cooldown_bars: int
sl_atr_mult: float
horizon: int
def make_candidates(seed: int) -> list[Candidate]:
rng = random.Random(seed)
candidates: list[Candidate] = []
candidate_id = 0
for values in itertools.product(
("H1", "H4"),
(0.00, 0.05, 0.10, 0.25),
(False, True),
(1, 2),
(0, 5, 15),
(0.75, 1.00, 1.50),
(30, 60),
):
candidates.append(Candidate(candidate_id=candidate_id, **dict(zip(
(
"reference_timeframe",
"min_sweep_atr",
"use_reference_bias",
"confirmation_window",
"cooldown_bars",
"sl_atr_mult",
"horizon",
),
values,
))))
candidate_id += 1
rng.shuffle(candidates)
return candidates
def generate_crt_signals(
m1: pd.DataFrame,
reference: pd.DataFrame,
atr: pd.Series,
candidate: Candidate,
) -> np.ndarray:
base = m1[["timestamp", "open", "high", "low", "close"]].copy()
merged = closed_reference_features(base, reference, candidate.reference_timeframe)
atr_values = atr.to_numpy(dtype=np.float32)
bullish_sweep = (
(merged["low"].to_numpy() < merged["crt_low"].to_numpy())
& (merged["close"].to_numpy() > merged["crt_low"].to_numpy())
& (
(merged["crt_low"].to_numpy() - merged["low"].to_numpy())
>= np.float32(candidate.min_sweep_atr) * atr_values
)
)
bearish_sweep = (
(merged["high"].to_numpy() > merged["crt_high"].to_numpy())
& (merged["close"].to_numpy() < merged["crt_high"].to_numpy())
& (
(merged["high"].to_numpy() - merged["crt_high"].to_numpy())
>= np.float32(candidate.min_sweep_atr) * atr_values
)
)
ref_open = merged["crt_open"].to_numpy()
ref_close = merged["crt_close"].to_numpy()
if candidate.use_reference_bias:
bullish_sweep &= ref_close >= ref_open
bearish_sweep &= ref_close <= ref_open
highs = merged["high"].to_numpy()
lows = merged["low"].to_numpy()
closes = merged["close"].to_numpy()
signals = np.zeros(len(merged), dtype=np.int8)
# Confirmation is a close through the sweep candle's opposite extreme.
for delay in range(1, candidate.confirmation_window + 1):
prior_bull = np.zeros(len(merged), dtype=bool)
prior_bear = np.zeros(len(merged), dtype=bool)
if delay < len(merged):
prior_bull[delay:] = bullish_sweep[:-delay]
prior_bear[delay:] = bearish_sweep[:-delay]
signals[delay:][prior_bull[delay:] & (closes[delay:] > highs[:-delay])] = 1
signals[delay:][prior_bear[delay:] & (closes[delay:] < lows[:-delay])] = -1
if candidate.cooldown_bars > 0:
last_signal = -candidate.cooldown_bars - 1
for i in range(len(signals)):
if signals[i] != 0:
if i - last_signal <= candidate.cooldown_bars:
signals[i] = 0
else:
last_signal = i
return signals
def metrics_for_slice(
signals: np.ndarray,
outcomes: dict[str, np.ndarray],
start: int,
end: int,
min_trades: int,
min_coverage: float,
) -> dict[str, Any]:
signal = signals[start:end]
chosen_direction = outcomes["chosen_direction"][start:end]
long_r = outcomes["long_r"][start:end]
short_r = outcomes["short_r"][start:end]
long_outcome = outcomes["long_outcome"][start:end]
short_outcome = outcomes["short_outcome"][start:end]
trades = signal != 0
n_rows = len(signal)
n_trades = int(trades.sum())
coverage = float(n_trades / n_rows) if n_rows else 0.0
if n_trades:
realized_r = np.where(signal > 0, long_r, short_r)
realized_outcome = np.where(signal > 0, long_outcome, short_outcome)
trade_accuracy = float((realized_r[trades] > 0).mean())
direction_accuracy = float(
(signal[trades] == chosen_direction[trades]).mean()
)
tp1_rate = float(np.isin(realized_outcome[trades], [3, 4, 5]).mean())
tp2_rate = float(np.isin(realized_outcome[trades], [4, 5]).mean())
tp3_rate = float((realized_outcome[trades] == 5).mean())
sl_rate = float((realized_outcome[trades] == 1).mean())
expectancy = float(realized_r[trades].mean())
total_r = float(realized_r[trades].sum())
positive = float(realized_r[trades][realized_r[trades] > 0].sum())
negative = float(-realized_r[trades][realized_r[trades] < 0].sum())
profit_factor = positive / negative if negative > 0 else math.inf
else:
trade_accuracy = direction_accuracy = 0.0
tp1_rate = tp2_rate = tp3_rate = sl_rate = 0.0
expectancy = total_r = 0.0
profit_factor = 0.0
eligible = n_trades >= min_trades and coverage >= min_coverage
return {
"rows": n_rows,
"trades": n_trades,
"coverage": coverage,
"eligible": eligible,
"trade_accuracy": trade_accuracy,
"direction_accuracy": direction_accuracy,
"tp1_or_better_rate": tp1_rate,
"tp2_or_better_rate": tp2_rate,
"tp3_rate": tp3_rate,
"sl_rate": sl_rate,
"expectancy_R": expectancy,
"profit_factor": profit_factor,
"total_R": total_r,
}
def objective(metrics: dict[str, Any]) -> tuple[float, float, float]:
if not metrics["eligible"]:
return (-1.0, metrics["coverage"], metrics["expectancy_R"])
return (
metrics["trade_accuracy"],
metrics["coverage"],
metrics["expectancy_R"],
)
# ---------------------------------------------------------------------------
# Pinned FLAN-T5 supervisor
# ---------------------------------------------------------------------------
class Supervisor:
def __init__(self) -> None:
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import torch
self.torch = torch
log(
"Loading pinned supervisor {} at revision {}",
SUPERVISOR_MODEL_ID,
SUPERVISOR_REVISION,
)
self.tokenizer = AutoTokenizer.from_pretrained(
SUPERVISOR_MODEL_ID,
revision=SUPERVISOR_REVISION,
trust_remote_code=False,
)
self.model = AutoModelForSeq2SeqLM.from_pretrained(
SUPERVISOR_MODEL_ID,
revision=SUPERVISOR_REVISION,
trust_remote_code=False,
)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.model.to(self.device)
self.model.eval()
def choose(
self,
candidates: list[Candidate],
history: list[dict[str, Any]],
) -> tuple[Optional[int], str]:
if not candidates:
return None, "no candidates"
candidate_text = "\n".join(
f"ID {c.candidate_id}: ref={c.reference_timeframe}, "
f"sweep_atr={c.min_sweep_atr}, bias={c.use_reference_bias}, "
f"confirm={c.confirmation_window}, cooldown={c.cooldown_bars}, "
f"sl={c.sl_atr_mult}, horizon={c.horizon}"
for c in candidates[:24]
)
history_text = "\n".join(
f"trial {row['trial']}: candidate={row['candidate_id']}, "
f"accuracy={row.get('validation', {}).get('trade_accuracy', 0):.4f}, "
f"coverage={row.get('validation', {}).get('coverage', 0):.4f}"
for row in history[-8:]
) or "No previous trials."
prompt = (
"You are a constrained trading-strategy supervisor. Choose one "
"candidate ID from the list. Do not invent an ID. Prefer enough "
"coverage and realistic validation accuracy. Reply exactly as "
"CANDIDATE_ID=<integer> followed by one short reason.\n\n"
f"Candidates:\n{candidate_text}\n\nHistory:\n{history_text}"
)
inputs = self.tokenizer(
prompt,
return_tensors="pt",
truncation=True,
max_length=768,
).to(self.device)
with self.torch.no_grad():
output = self.model.generate(**inputs, max_new_tokens=48)
reply = self.tokenizer.decode(output[0], skip_special_tokens=True)
match = re.search(r"CANDIDATE_ID\s*=\s*(\d+)", reply)
selected = int(match.group(1)) if match else None
valid_ids = {candidate.candidate_id for candidate in candidates}
if selected not in valid_ids:
selected = None
return selected, reply
# ---------------------------------------------------------------------------
# Checkpointing and optimization loop
# ---------------------------------------------------------------------------
def save_json(path: Path, value: Any) -> None:
path.write_text(json.dumps(value, indent=2, default=str), encoding="utf-8")
def save_checkpoint(
checkpoint_dir: Path,
candidate: Candidate,
result: dict[str, Any],
milestone: Optional[int] = None,
) -> None:
prefix = "best" if milestone is None else f"milestone_{milestone:02d}pct"
payload = {
"saved_at": utc_now(),
"candidate": asdict(candidate),
"result": result,
"supervisor_model": {
"id": SUPERVISOR_MODEL_ID,
"revision": SUPERVISOR_REVISION,
"trust_remote_code": False,
},
"crt_sources": [
"https://innercircletrader.net/tutorials/candle-range-theory-crt/",
"https://tradingwyckoff.com/en/crt/",
],
}
save_json(checkpoint_dir / f"{prefix}_checkpoint.json", payload)
log("Saved {} checkpoint at validation accuracy {:.2%}", prefix, result["validation"]["trade_accuracy"])
def package_checkpoints(checkpoint_dir: Path) -> Path:
archive_path = checkpoint_dir.parent / "100optimization_checkpoints.zip"
if archive_path.exists():
archive_path.unlink()
with zipfile.ZipFile(archive_path, "w", zipfile.ZIP_DEFLATED) as archive:
for path in sorted(checkpoint_dir.rglob("*")):
if path.is_file():
archive.write(path, arcname=f"{checkpoint_dir.name}/{path.relative_to(checkpoint_dir)}")
log("Checkpoint archive: {}", archive_path)
return archive_path
def load_market_data() -> tuple[dict[str, pd.DataFrame], Path]:
archive = load_dataset_archive()
extracted = safe_extract_zip(archive, CFG.data_dir)
data_dir = find_data_directory(extracted)
raw: dict[str, pd.DataFrame] = {}
for timeframe in ("M1", "M5", "M15", "M30", "H1", "H4"):
path = find_timeframe_file(data_dir, timeframe)
if path is None:
raise FileNotFoundError(f"Missing {timeframe}.csv in {data_dir}")
raw[timeframe] = load_ohlcv(path)
log("Loaded {}: {:,} rows", timeframe, len(raw[timeframe]))
return raw, data_dir
def run() -> dict[str, Any]:
checkpoint_dir = ensure_checkpoint_dir()
random.seed(CFG.seed)
np.random.seed(CFG.seed)
raw, data_dir = load_market_data()
m1 = raw["M1"]
if CFG.max_rows and len(m1) > CFG.max_rows:
m1 = m1.tail(CFG.max_rows).reset_index(drop=True)
log("Using the last {:,} M1 rows because MAX_ROWS is set", len(m1), level="WARN")
atr = compute_atr(m1, CFG.atr_period)
candidates = make_candidates(CFG.seed)
outcome_cache: dict[tuple[float, int], dict[str, np.ndarray]] = {}
signal_cache: dict[int, np.ndarray] = {}
history: list[dict[str, Any]] = []
evaluated: set[int] = set()
best_result: Optional[dict[str, Any]] = None
best_candidate: Optional[Candidate] = None
milestones_saved: set[int] = set()
n = len(m1)
train_end = int(n * 0.60)
validation_end = int(n * 0.80)
purge = max(CFG.default_horizon, 60)
validation_start = min(n, train_end + purge)
test_start = min(n, validation_end + purge)
supervisor: Optional[Supervisor]
try:
supervisor = Supervisor()
except Exception as exc:
log("Supervisor unavailable: {}. Continuing deterministically.", exc, level="WARN")
supervisor = None
log(
"CRT optimization rows={} | train={} | validation={} | test={}",
n,
train_end,
validation_end - validation_start,
n - test_start,
)
for trial in range(CFG.max_trials):
remaining = [candidate for candidate in candidates if candidate.candidate_id not in evaluated]
if not remaining:
break
selected_id: Optional[int] = None
supervisor_reply = ""
if supervisor is not None and history:
selected_id, supervisor_reply = supervisor.choose(remaining, history)
if selected_id is None:
# Deterministic fallback: evaluate candidates in the seeded order.
selected_id = remaining[0].candidate_id
candidate = next(c for c in candidates if c.candidate_id == selected_id)
evaluated.add(candidate.candidate_id)
key = (candidate.sl_atr_mult, candidate.horizon)
if key not in outcome_cache:
outcome_cache[key] = build_outcomes(
m1,
CFG.atr_period,
candidate.sl_atr_mult,
candidate.horizon,
)
outcomes = outcome_cache[key]
if candidate.candidate_id not in signal_cache:
signal_cache[candidate.candidate_id] = generate_crt_signals(
m1,
raw[candidate.reference_timeframe],
atr,
candidate,
)
signals = signal_cache[candidate.candidate_id]
validation = metrics_for_slice(
signals,
outcomes,
validation_start,
validation_end,
CFG.min_trades,
CFG.min_coverage,
)
test = metrics_for_slice(
signals,
outcomes,
test_start,
n,
CFG.min_trades,
CFG.min_coverage,
)
result = {
"trial": trial + 1,
"candidate_id": candidate.candidate_id,
"candidate": asdict(candidate),
"validation": validation,
"test_preview": test,
"supervisor_reply": supervisor_reply,
}
history.append(result)
save_json(checkpoint_dir / "trials.json", history)
log(
"Trial {} candidate={} validation accuracy={:.2%} coverage={:.2%} "
"trades={} expectancy={:.3f}R test accuracy={:.2%}",
trial + 1,
candidate.candidate_id,
validation["trade_accuracy"],
validation["coverage"],
validation["trades"],
validation["expectancy_R"],
test["trade_accuracy"],
)
if best_result is None or objective(validation) > objective(best_result["validation"]):
best_result = result
best_candidate = candidate
save_checkpoint(checkpoint_dir, candidate, result)
achieved = validation["eligible"] and validation["trade_accuracy"] >= CFG.target_accuracy
for milestone in (20, 30, 80):
if (
milestone not in milestones_saved
and validation["eligible"]
and validation["trade_accuracy"] >= milestone / 100.0
):
save_checkpoint(checkpoint_dir, candidate, result, milestone=milestone)
milestones_saved.add(milestone)
if achieved:
log(
"Target validation accuracy reached: {:.2%}. "
"No further optimization trials will run.",
validation["trade_accuracy"],
)
break