import argparse
import json
import sys
import time
from pathlib import Path
sys.path.insert(0, str(Path(__file__).parent))
# Force UTF-8 output on Windows so emoji/em-dash in log lines don't crash the terminal
if hasattr(sys.stdout, "reconfigure"):
sys.stdout.reconfigure(encoding="utf-8", errors="replace")
if hasattr(sys.stderr, "reconfigure"):
sys.stderr.reconfigure(encoding="utf-8", errors="replace")
# Load .env for Telegram credentials (silently ignored if python-dotenv not installed)
try:
from dotenv import load_dotenv
load_dotenv()
except ImportError:
pass
from src.logger import setup_logger, reconfigure_for_tf
from src.processor import process_pipeline, TF_CONFIG, PROCESSED_PATH, save_feature_scaler, load_feature_scaler
from src.engine_hmm import fit_hmm, save_model as save_hmm, load_model as load_hmm, get_model_path as hmm_model_path
from src.engine_xgb import (
prepare_features, train_xgb, get_predictions,
export_onnx, save_xgb, load_xgb, ONNX_PATH, FEATURE_COLS,
train_xgb_ensemble, get_predictions_ensemble,
save_xgb_ensemble, load_xgb_ensemble, export_onnx_ensemble, ENSEMBLE_PKL_PATH,
get_ensemble_path, TF_TRAIN_RATIO, compute_regime_stats,
)
from src.optimizer import (
run_optimization, get_best_params, _score_result as _calc_score,
extract_consensus_params, run_wfa as optimizer_run_wfa,
WFO_PARAMS, WFO_PARAMS_FAST,
)
from src.backtester import vectorized_backtest, format_payout
from src.visualizer import generate_full_report
from src.risk_manager import AdaptiveRiskManager
logger = setup_logger("main")
def _resolve_balance(args) -> float:
"""--balance takes precedence over --min_cap."""
return args.balance if args.balance is not None else args.min_cap
_M5_EXPIRY_HOURS = 120 # 5 days
# Staleness thresholds per TF (days). If the saved model is older than this
# the live gate aborts with a warning and Telegram alert.
# M5 is tightest (14d) — microstructure regimes shift weekly.
# H1/M15 are more stable but should still re-optimise monthly.
_MODEL_STALE_DAYS = {"M5": 14, "M15": 30, "H1": 30}
def _m5_meta_path(broker: str) -> Path:
return Path(f"models/m5_meta_{broker}.json")
def _check_m5_readiness(tf: str, broker: str = "headway_cent") -> bool:
"""Return True if the M5 model is fresh enough for live trading.
The M5 timeframe is sensitive to microstructure changes, so the model
must have been optimised within the last 5 days. A meta.json timestamp
is written by cmd_optimize after each successful M5 study.
"""
if tf.upper() != "M5":
return True
meta_path = _m5_meta_path(broker)
if not meta_path.exists():
# Legacy fallback: files created before per-broker naming used models/m5_meta.json
_legacy = Path("models/m5_meta.json")
if _legacy.exists():
logger.info(
"Migrating legacy m5_meta.json -> %s for broker=%s.", meta_path, broker
)
meta_path.write_text(_legacy.read_text())
else:
print(
"\nERROR: M5 model meta-data not found.\n"
"Run python main.py --mode optimize --tf M5 before live trading."
)
return False
meta = json.loads(meta_path.read_text())
age_h = (time.time() - meta.get("timestamp", 0)) / 3600
if age_h > _M5_EXPIRY_HOURS:
print(
f"\nWARNING: M5 model is {age_h:.0f} hours old (limit: {_M5_EXPIRY_HOURS}h / 5 days).\n"
"REQUIRED ACTION: Run python main.py --mode optimize --tf M5 to refresh before live.\n"
"\n Pre-flight checklist:\n"
" 1. python main.py --mode sync_validate --period 3m --tf M5\n"
" 2. python main.py --mode optimize --tf M5 --trials 500 --balance 15\n"
" 3. python main.py --mode train --tf M5\n"
)
return False
return True
def _train_for_tf(tf: str, balance: float, broker: str, params: dict):
"""Shared train logic for a single timeframe. Returns (result, model_hmm, states, X, metrics)."""
df = process_pipeline(
obs_cov=params.get("obs_cov"),
trans_cov=params.get("trans_cov"),
save=True,
tf=tf,
save_models=True,
broker=broker,
)
model_hmm, states, state_map = fit_hmm(
df, n_states=params.get("n_states", TF_CONFIG[tf].get("n_states_default", 3))
)
X, y, df_aligned, feature_scaler = prepare_features(df, states, tf=tf)
save_feature_scaler(feature_scaler, tf=tf, broker=broker)
models_ensemble, thresholds, metrics = train_xgb_ensemble(
X, y,
train_ratio=1.0, # full-data training — WFO validation happened in optimize step
max_depth=params.get("max_depth", 4),
learning_rate=params.get("learning_rate", 0.1),
n_estimators=params.get("n_estimators", 200),
subsample=params.get("subsample", 0.8),
colsample_bytree=params.get("colsample_bytree", 0.8),
min_child_weight=params.get("min_child_weight", 5),
gamma=params.get("gamma", 1.0),
reg_alpha=params.get("reg_alpha", 0.1),
reg_lambda=params.get("reg_lambda", 1.0),
scale_pos_weight=params.get("scale_pos_weight", 1.0),
)
_, probabilities = get_predictions_ensemble(models_ensemble, thresholds, X)
states_aligned = states[df.index.isin(df_aligned.index)]
metrics["regime_stats"] = compute_regime_stats(models_ensemble, thresholds, X, states_aligned)
result = vectorized_backtest(
df_aligned, probabilities, states_aligned,
split_idx=None, # full-data model — no IS/OOS split; CPCV scores used for validation
account_size=balance,
broker=broker,
tf=tf,
)
return result, model_hmm, state_map, models_ensemble, thresholds, metrics, df_aligned, states_aligned, X, probabilities
def _check_model_staleness(tf: str, broker: str, args) -> None:
"""Abort live/demo start if the saved model exceeds the staleness threshold.
Sends a Telegram alert and calls ``sys.exit(1)`` when the model is too old.
Pass ``--skip_stale_check`` on the CLI to bypass this gate (e.g. for demo
testing when you intentionally don't want to re-optimise).
Does nothing for unknown TFs or when ``--skip_stale_check`` is set.
"""
if getattr(args, "skip_stale_check", False):
logger.info("Staleness gate bypassed (--skip_stale_check).")
return
from src.validator import check_model_age
from src.notifier import send_telegram_msg
max_age = _MODEL_STALE_DAYS.get(tf.upper(), 30)
age_days = check_model_age(tf=tf, broker=broker)
if age_days <= max_age:
logger.info(
"Model freshness OK: %s/%s is %.1f days old (limit %d).",
tf, broker, age_days, max_age,
)
return
age_str = f"{age_days:.0f}" if age_days != float("inf") else "unknown (file missing)"
msg = (
f"⚠️ Market Drift/Staleness detected. Pausing trade loop — "
f"[{tf}] model is {age_str} days old (limit: {max_age} days).\n"
f"Re-optimise before going live:\n"
f" python main.py --mode optimize --tf {tf} --broker {broker} --trials 500\n"
f" python main.py --mode train --tf {tf} --broker {broker}\n"
f"Add --skip_stale_check to the live command to bypass this gate."
)
logger.warning("STALE MODEL [%s/%s]: %.0f days old — aborting.", tf, broker, age_days)
send_telegram_msg(f"{msg}")
print(f"\n{msg}")
sys.exit(1)
def cmd_wfa(args):
"""Walk-Forward Analysis: re-fit HMM+XGBoost on each rolling IS window.
Uses the best Optuna params from the study DB to run the same rolling
IS/OOS evaluation that scored those params during optimization. Prints
per-window OOS diagnostics and the Walk-Forward Efficiency (WFE) ratio.
WFE target: > 0.50 (OOS performance is at least 50% of IS performance).
"""
import pandas as pd
from src.notifier import send_telegram_msg
balance = _resolve_balance(args)
broker = args.broker
tf = args.tf.upper()
wfo_mode = "fast" if getattr(args, "fast_wfo", False) else "standard"
reconfigure_for_tf(tf)
processed_path = TF_CONFIG[tf]["processed_path"]
if not processed_path.exists():
print(f"\nERROR: Run --mode process --tf {tf} first.")
sys.exit(1)
df = pd.read_parquet(processed_path)
print(f"\n=== Walk-Forward Analysis [{tf} / {broker}] (mode: {wfo_mode}) ===")
print(f" Dataset: {len(df)} bars ({df.index[0].date()} – {df.index[-1].date()})")
print(" Running rolling IS/OOS windows with best Optuna params...\n")
wfa = optimizer_run_wfa(
df=df, tf=tf, broker=broker,
account_size=balance, wfo_mode=wfo_mode,
)
n_win = wfa["n_windows"]
n_valid = wfa["n_valid_windows"]
wfe = wfa["wfe_ratio"]
scores = wfa["window_scores"]
verdict = "ROBUST" if wfe >= 0.50 else "FRAGILE — re-optimise recommended"
print(f" Total windows : {n_win}")
print(f" Valid windows : {n_valid} (OOS trades above hard floor, DD < 20%)")
print(f" Median OOS score : {float(__import__('numpy').median(scores)) if scores else 0.0:+.3f}")
print(f" Std OOS Sharpe : {wfa['std_sharpe']:.3f}")
print(f" Median OOS trades : {wfa['median_trades']}")
print(f" WFE ratio : {wfe:.3f} [{verdict}]")
print(f" WFO score : {wfa['wfo_score']:+.3f}")
if scores:
print("\n Per-window scores:")
for i, s in enumerate(scores):
bar = "#" * max(0, int((s + 1) * 5))
print(f" Window {i+1:>2}: {s:+.3f} {bar}")
send_telegram_msg(
f"Walk-Forward Analysis [{tf}]\n"
f"Valid windows: {n_valid}/{n_win} | WFE: {wfe:.2f}\n"
f"WFO score: {wfa['wfo_score']:+.3f} "
+ ("Robust" if wfe >= 0.50 else "Fragile — re-optimise")
)
def cmd_process(args):
tfs = [t.strip().upper() for t in args.tf.split(",")]
for tf in tfs:
if tf not in TF_CONFIG:
logger.error("Unknown timeframe '%s'. Valid: %s", tf, list(TF_CONFIG))
continue
reconfigure_for_tf(tf)
try:
df = process_pipeline(save=True, tf=tf, save_models=True, broker=args.broker)
logger.info(
"[%s] Done: %d bars | Range: %s to %s",
tf, len(df), df.index.min(), df.index.max(),
)
except FileNotFoundError as e:
logger.error(str(e))
def cmd_optimize(args):
import pandas as pd
balance = _resolve_balance(args)
broker = args.broker
tfs = [t.strip().upper() for t in args.tf.split(",")]
wfo_mode = "fast" if getattr(args, "fast_wfo", False) else "standard"
for tf in tfs:
reconfigure_for_tf(tf)
logger.info(
"Optimizing [%s] broker=%s balance=$%.0f trials=%d wfo_mode=%s",
tf, broker, balance, args.trials, wfo_mode,
)
# Pre-load parquet once — make_objective only calls kalman_smooth per trial
processed_path = TF_CONFIG[tf]["processed_path"]
if not processed_path.exists():
print(
f"\nERROR: Processed parquet not found for {tf}.\n"
f"Run python main.py --mode process --tf {tf} first."
)
sys.exit(1)
df = pd.read_parquet(processed_path)
study = run_optimization(
df=df,
tf=tf,
broker=broker,
account_size=balance,
n_trials=args.trials,
wfo_mode=wfo_mode,
n_jobs=args.n_jobs,
)
print(f"\n=== Best Result [{tf}] ===")
print(f"Score: {study.best_value:.3f}")
print(f"Broker: {broker} | Balance: ${balance:.0f} | WFO mode: {wfo_mode}")
print("Best Params:")
for k, v in study.best_params.items():
print(f" {k}: {v}")
if tf == "M5":
meta_path = _m5_meta_path(broker)
meta_path.parent.mkdir(parents=True, exist_ok=True)
meta_path.write_text(json.dumps({
"timestamp": time.time(),
"tf": "M5",
"best_score": study.best_value,
}))
print(
"\nOptimization Complete. HMM States and XGBoost Weights have been updated. "
"You are now cleared for M5 Live Trading for the next 5 days."
)
def cmd_train(args):
balance = _resolve_balance(args)
broker = args.broker
tf = args.tf.upper()
reconfigure_for_tf(tf)
try:
params = get_best_params(balance=balance, broker=broker, tf=tf)
logger.info("Using Optuna best params [%s/%s]: %s", tf, broker, params)
except Exception:
logger.warning("No Optuna study found for tf=%s broker=%s — using defaults", tf, broker)
params = {}
result, model_hmm, state_map, models_ensemble, thresholds, metrics, df_aligned, states_aligned, X, probabilities = _train_for_tf(tf, balance, broker, params)
# Guard: refuse to save a degenerate HMM — identical state means + near-zero persistence
# indicate the Kalman/HMM params were bad (usually because Optuna params weren't loaded).
_min_persist = min(model_hmm.transmat_[i, i] for i in range(model_hmm.n_components))
if _min_persist < 0.70:
logger.critical(
"TRAINING ABORTED: degenerate HMM (min persistence=%.4f). "
"Model NOT saved. Usually caused by missing Optuna study for tf=%s broker=%s. "
"Ensure --mode optimize has completed and re-run --mode train.",
_min_persist, tf, broker,
)
print(
f"\nERROR: Degenerate HMM detected (min state persistence={_min_persist:.4f} < 0.70).\n"
f"The model was NOT saved — saving it would only produce garbage signals.\n"
f"\nFix: python main.py --mode optimize --tf {tf} --broker {broker} --trials 300\n"
f"Then: python main.py --mode train --tf {tf} --broker {broker}\n"
)
sys.exit(1)
save_hmm(model_hmm, hmm_model_path(tf, broker))
save_xgb_ensemble(models_ensemble, thresholds, metrics, get_ensemble_path(tf, broker))
arm = AdaptiveRiskManager(balance, broker=broker)
cpcv_score = None
try:
from src.optimizer import get_best_params as _gbp
_study = __import__("optuna").load_study(
study_name=None,
storage=f"sqlite:///models/study_{broker}.db",
)
cpcv_score = _study.best_value
except Exception:
pass
print(f"\n=== Training Results [{tf}] ===")
print(f"Broker: {broker} | Balance: ${balance:.0f} | Tier: {'small' if arm.is_small_account else 'growth'}")
print(f"Full-data model trained on 100% of available bars (CPCV validation via optimize step).")
print(f"Full-period Sharpe: {result['sharpe_ratio']:.3f} | MaxDD: {result['max_drawdown']*100:.1f}%"
f" | WR: {result['win_rate']*100:.1f}% | Trades: {result['n_trades']}")
if cpcv_score is not None:
print(f"CPCV Validation Score (best trial): {cpcv_score:.3f}")
print(f"\nModels saved. Run --mode export to generate ONNX.")
def cmd_extract_consensus(args):
"""Extract consensus hyperparameters from top-N trials (median per param).
More robust than the single best trial. Prints the consensus params and
saves them nowhere — use with --mode train to apply the consensus params.
"""
tf = args.tf.upper()
broker = args.broker
top_n = getattr(args, "top_n", 10)
min_wfe = getattr(args, "min_wfe", 0.0)
reconfigure_for_tf(tf)
try:
consensus = extract_consensus_params(tf=tf, broker=broker, top_n=top_n, min_wfe=min_wfe)
except Exception as exc:
print(f"\nERROR: {exc}")
sys.exit(1)
meta = consensus.pop("meta", {})
print(f"\n=== Consensus Params [{tf} / {broker}] ===")
print(f" Study: {meta.get('study_name', '?')}")
print(f" Top-N used: {meta.get('top_n_actual', top_n)}")
print(f" Score range: [{meta.get('min_score', 0.0):.3f}, {meta.get('max_score', 0.0):.3f}]")
print(f" Mean WFE: {meta.get('mean_wfe', 0.0):.3f}")
print("\n Consensus parameters (median of top-N):")
for k, v in sorted(consensus.items()):
print(f" {k}: {v}")
def cmd_compare(args):
"""Train and backtest the specified timeframes, then show a side-by-side comparison."""
balance = _resolve_balance(args)
broker = args.broker
tfs = [t.strip().upper() for t in args.tf.split(",")]
results = {}
for tf in tfs:
reconfigure_for_tf(tf)
try:
params = get_best_params(balance=balance, broker=broker, tf=tf)
except Exception:
params = {}
try:
result, *_ = _train_for_tf(tf, balance, broker, params)
if "oos_sharpe_ratio" in result:
_oos_fdd = result.get("oos_floating_max_drawdown", result.get("oos_max_drawdown", 0.0))
result["oos_score"] = _calc_score({
"total_return": result.get("oos_total_return", 0.0),
"floating_max_drawdown": _oos_fdd,
"sharpe_ratio": result.get("oos_sharpe_ratio", 0.0),
"profit_factor": result.get("oos_profit_factor", 1.0),
})
results[tf] = result
except FileNotFoundError as e:
logger.error("Skipping %s: %s", tf, e)
continue
if len(results) < 2:
avail = ", ".join(tfs)
print(f"\nOnly {len(results)} timeframe(s) available. Run --mode process --tf {avail} first.")
if results:
tf, r = next(iter(results.items()))
print(f"\n[{tf}] Sharpe={r['sharpe_ratio']:.3f} | OOS={r.get('oos_sharpe_ratio',0):.3f} "
f"| DD={r['max_drawdown']*100:.1f}% | Trades={r['n_trades']}")
return
def _oos_score(r):
return r.get("oos_score", r.get("oos_sharpe_ratio", r.get("sharpe_ratio", 0.0)))
winner = max(results, key=lambda k: _oos_score(results[k]))
col_w = 12
tf_list = list(results.keys())
header = f"{'Metric':<20}" + "".join(f" {tf:>{col_w}}" for tf in tf_list)
print(f"\n{'='*50}")
print(f" TIMEFRAME COMPARISON | Balance: ${balance:.0f} | Broker: {broker}")
print(f"{'='*50}")
print(header)
print("-" * len(header))
metrics_to_show = [
("OOS Score", "oos_score", ".2f"),
("OOS Sharpe", "oos_sharpe_ratio", ".3f"),
("OOS RF", "oos_recovery_factor", ".2f"),
("OOS PF", "oos_profit_factor", ".2f"),
("IS Sharpe", "is_sharpe_ratio", ".3f"),
("OOS Float DD", "oos_floating_max_drawdown", ".1%"),
("OOS Win Rate", "oos_win_rate", ".1%"),
("OOS Trades", "oos_n_trades", "d"),
("Full Sharpe", "sharpe_ratio", ".3f"),
]
for label, key, fmt in metrics_to_show:
row = f"{label:<20}"
for tf in tf_list:
r = results[tf]
val = r.get(key, r.get(key.replace("oos_", "").replace("is_", ""), 0))
s = f"{val:{fmt}}" if fmt != "d" else str(int(val))
row += f" {s:>{col_w}}"
print(row)
for tf, r in results.items():
logger.info(
" %s Sharpe=%.3f (OOS=%.3f) | DD=%.1f%% | WR=%.1f%% | Trades=%d",
tf, r["sharpe_ratio"], r.get("oos_sharpe_ratio", r["sharpe_ratio"]),
r["max_drawdown"] * 100, r["win_rate"] * 100, r["n_trades"],
)
print(f"\n-> Recommended timeframe: {winner} (higher OOS Score)")
print(f" Run: python main.py --mode train --tf {winner} --broker {broker} --balance {balance:.0f}")
def cmd_export(args):
tf = args.tf.upper()
broker = args.broker
xgb_path = get_ensemble_path(tf, broker)
if not xgb_path.exists():
xgb_path = ENSEMBLE_PKL_PATH
try:
models_ensemble, _, xgb_metrics = load_xgb_ensemble(xgb_path)
except FileNotFoundError:
logger.error("No trained ensemble model found. Run --mode train first.")
sys.exit(1)
feature_cols = xgb_metrics.get("feature_cols", FEATURE_COLS)
n_features = len(feature_cols)
paths = export_onnx_ensemble(models_ensemble, n_features=n_features)
print(f"\nONNX ensemble exported ({n_features} features: {feature_cols}):")
for bucket, path in paths.items():
print(f" [{bucket:>4}] {path}")
print("\nCopy these files to your MT5 MQL5/Files/ directory.")
print("The EA selects the model based on the current ATR volatility bucket.")
def cmd_sync_validate(args):
"""Download recent MT5 bars then run the model validation gatekeeper."""
from src.mt5_sync import sync_mt5_data
from src.validator import run_validation
balance = _resolve_balance(args)
tf = args.tf.upper()
reconfigure_for_tf(tf)
logger.info("Syncing MT5 data [%s] period=%s ...", tf, args.period)
try:
df = sync_mt5_data(tf=tf, period=args.period)
logger.info(
"Sync complete: %d bars %s -> %s",
len(df), df.index.min(), df.index.max(),
)
except Exception as exc:
logger.error("MT5 sync failed: %s", exc)
sys.exit(1)
try:
result = run_validation(tf=tf, broker=args.broker, account_size=balance)
except Exception as exc:
logger.error("Validation error: %s", exc)
sys.exit(1)
print(f"\n=== Validation Result [{tf}] ===")
_fdd = result.get("max_dd", 0.0)
_eff = result.get("avg_efficiency", 0.0)
_cost_e = result.get("cost_efficiency", 0.0)
_payout = format_payout(result.get("total_return", 0.0), balance, args.broker)
print(
f" [{tf} LIVE] Score: {result.get('score', 0.0):.2f}"
f" | RF: {result.get('recovery_factor', 0.0):.2f}"
f" | PF: {result.get('profit_factor', 1.0):.2f}"
f" | Payoff: ${result.get('expected_payoff', 0.0)*balance:.4f}"
f" | MaxDD: {_fdd*100:.1f}% (Floating)"
)
print(
f" Efficiency: {_eff:.2f}x ATR/Spread"
f" | CostEff: {_cost_e*100:.1f}%"
f" | Total Payout: {_payout}"
)
if _eff < 1.2:
print(" [!] WARNING: Low Market Efficiency — Spread is eating your edge.")
if _cost_e < 0.50:
print(f" [!] WARNING: Broker is consuming >{(1-_cost_e)*100:.0f}% of gross profit.")
print(f" Sharpe: {result['sharpe']:.3f} | Trades: {result['n_trades']} | WR: {result['win_rate']*100:.1f}%")
print(f" Status: {result['status'].upper()}")
print(f" {result['message']}")
if result["status"] == "fail":
print(
"\nABORTING: Validation failed. "
"Retune with --mode optimize then --mode train before going live."
)
sys.exit(1)
if result["status"] == "warn":
print("\nWARNING: Borderline performance — consider re-optimising before live trading.")
def cmd_demo(args):
"""Connect to MT5 demo account and start the live signal execution loop."""
from src.mt5_trader import run_live_loop
balance = _resolve_balance(args)
tf = args.tf.upper()
if not _check_m5_readiness(tf, args.broker):
sys.exit(1)
_check_model_staleness(tf, args.broker, args)
logger.info("Starting demo loop — TF=%s broker=%s balance=$%.0f",
tf, args.broker, balance,
)
run_live_loop(tf=tf, broker=args.broker, account_size=balance,
profit_target=getattr(args, "profit_target", None))
def cmd_live(args):
"""Connect to MT5 live account and start the live signal execution loop."""
from src.mt5_trader import run_live_loop
balance = _resolve_balance(args)
tf = args.tf.upper()
reconfigure_for_tf(tf)
if not _check_m5_readiness(tf, args.broker):
sys.exit(1)
_check_model_staleness(tf, args.broker, args)
if not args.yes:
print("\n" + "=" * 60)
print(" WARNING: LIVE ACCOUNT — real money is at risk.")
print(" Ensure --mode sync_validate passed before continuing.")
print(" Ensure the GoldRegimeX EA is removed from the XAUUSD chart.")
confirm = input(" Type YES to confirm live trading: ").strip()
if confirm != "YES":
print("Aborted.")
sys.exit(0)
print("=" * 60 + "\n")
logger.info(
"Starting live loop — TF=%s broker=%s balance=$%.0f",
tf, args.broker, balance,
)
run_live_loop(tf=tf, broker=args.broker, account_size=balance,
profit_target=getattr(args, "profit_target", None))
def cmd_report(args):
balance = _resolve_balance(args)
broker = args.broker
tf = args.tf.upper()
reconfigure_for_tf(tf)
try:
params = get_best_params(balance=balance, broker=broker, tf=tf)
except Exception:
params = {}
df = process_pipeline(obs_cov=params.get("obs_cov"), trans_cov=params.get("trans_cov"),
save=False, tf=tf)
_hmm_path = hmm_model_path(tf, broker)
if not _hmm_path.exists():
_hmm_path = None # let the except branch fit a fresh one
try:
if _hmm_path is None:
raise FileNotFoundError
model_hmm = load_hmm(_hmm_path)
from src.engine_hmm import predict_states, STATE_NAMES_3, STATE_NAMES_2, STATE_NAMES_4
states = predict_states(model_hmm, df)
n = model_hmm.n_components
state_names = {2: STATE_NAMES_2, 3: STATE_NAMES_3, 4: STATE_NAMES_4}.get(n, STATE_NAMES_3)
except Exception:
model_hmm, states, state_names = fit_hmm(df, n_states=params.get("n_states", 3))
# Load the saved feature scaler so the report uses identical scaling to training
try:
_feat_scaler = load_feature_scaler(tf=tf, broker=broker)
except FileNotFoundError:
_feat_scaler = None # old model without scaler — prepare_features fits fresh
X, y, df_aligned, _ = prepare_features(df, states, feature_scaler=_feat_scaler, tf=tf)
_xgb_path = get_ensemble_path(tf, broker)
if not _xgb_path.exists():
_xgb_path = ENSEMBLE_PKL_PATH
try:
models_xgb, thresholds_xgb, metrics = load_xgb_ensemble(_xgb_path)
# If the saved ensemble was trained with different features (e.g. before DXY
# was added), retrain in-memory so the report stays consistent with the data.
if metrics.get("feature_cols") != list(X.columns):
raise ValueError("Feature mismatch — retraining for report.")
except (FileNotFoundError, ValueError):
models_xgb, thresholds_xgb, metrics = train_xgb_ensemble(
X, y,
max_depth=params.get("max_depth", 4),
learning_rate=params.get("learning_rate", 0.1),
n_estimators=params.get("n_estimators", 200),
subsample=params.get("subsample", 0.8),
colsample_bytree=params.get("colsample_bytree", 0.8),
min_child_weight=params.get("min_child_weight", 5),
gamma=params.get("gamma", 1.0),
reg_alpha=params.get("reg_alpha", 0.1),
)
_, probabilities = get_predictions_ensemble(models_xgb, thresholds_xgb, X)
states_aligned = states[df.index.isin(df_aligned.index)]
# Ensure regime_stats are present (may be absent in pre-Z-Score saved models)
if not metrics.get("regime_stats"):
_rs_split = metrics.get("split_idx") or int(len(X) * 0.8)
_X_is = X.iloc[:_rs_split]
_hs_is = states_aligned[:len(_X_is)]
metrics["regime_stats"] = compute_regime_stats(models_xgb, thresholds_xgb, _X_is, _hs_is)
# split_idx in saved metrics may come from a different TF's training run
# (e.g. M15 split_idx ~186k used in an H1 report with only ~58k bars).
# Recompute from current data when the stored value doesn't fit.
split_idx = metrics.get("split_idx")
if split_idx is None or not (0 < split_idx < len(X)):
split_idx = int(len(X) * 0.8)
result = vectorized_backtest(
df_aligned, probabilities, states_aligned,
split_idx=split_idx, account_size=balance, broker=broker, tf=tf,
regime_stats=metrics.get("regime_stats"),
evaluator_config=({"Z_CUTOFF_BULL": params["z_cutoff_bull"], "Z_CUTOFF_BEAR": -params["z_cutoff_bull"]}
if params.get("z_cutoff_bull") else None),
)
arm = AdaptiveRiskManager(balance, broker=broker)
limits = arm.get_trade_limits()
display_params = dict(params)
display_params.update({
"broker": broker,
"balance_usd": balance,
"tier": "small" if arm.is_small_account else "growth",
"pos_per_trade": limits["pos_per_trade"],
"tf": tf,
})
paths = generate_full_report(
df_aligned, states_aligned, state_names, model_hmm,
X, probabilities, metrics, result, display_params,
split_idx=split_idx, tf=tf, broker=broker,
account_size=balance,
)
print(f"\n=== Report Generated [{tf} / {broker}] ===")
for p in paths:
print(f" {p}")
print(f"\n{len(paths)} charts saved to reports/{tf}_{broker}/")
def cmd_audit(args):
"""Print (and optionally send) the daily MT5 performance report."""
from src.auditor import get_daily_report
from src.notifier import send_telegram_msg
balance = _resolve_balance(args)
report = get_daily_report(broker=args.broker)
print(report)
sent = send_telegram_msg(report)
if sent:
print("\nReport also sent to Telegram.")
else:
print("\n(Telegram not configured — see .env.example)")
def cmd_guardian(args):
"""Start the multi-TF health monitor loop."""
from src.guardian import run_guardian
balance = _resolve_balance(args)
tfs = [t.strip().upper() for t in args.tf.split(",")]
interval = getattr(args, "interval", 3600)
run_guardian(
tfs=tfs,
broker=args.broker,
account_size=balance,
period=args.period,
interval_sec=interval,
)
def cmd_consolidate(args):
"""Consolidate multi-asset CSV exports in data/raw/ into per-TF master files.
Processes 5 assets × 3 timeframes = 15 master files in data/processed/.
Assets: USDCHF, XAGUSD, XTIUSD, US500, USDJPY.
Timeframes: H1, M15, M5.
"""
from src.data_consolidator import consolidate_asset, ASSET_CONFIGS
assets = list(ASSET_CONFIGS.keys())
tfs = ["H1", "M15", "M5"]
n_ok = 0
for asset in assets:
for tf in tfs:
result = consolidate_asset(asset, tf)
if not result.empty:
out = ASSET_CONFIGS[asset][tf]["output"]
print(f" {asset} {tf}: {len(result)} rows → data/processed/{out}")
n_ok += 1
else:
logger.warning("%s %s: no source files found — skipping.", asset, tf)
print(f"\nConsolidation complete: {n_ok}/{len(assets) * len(tfs)} master files produced.")
def cmd_listen(args):
"""Start the Telegram remote control listener + nightly report scheduler."""
import threading
from src.remote_control import run_listener
from src.notifier import send_telegram_msg
from src.auditor import get_daily_report
balance = _resolve_balance(args)
# ── Nightly audit scheduler (runs in background thread) ──────────────────
def _run_scheduler():
try:
import schedule
except ImportError:
logger.warning(
"schedule package not installed — nightly report disabled. "
"Install with: pip install schedule"
)
return
def _send_nightly():
report = get_daily_report(broker=args.broker)
send_telegram_msg(f"Nightly Report\n{report}")
logger.info("Nightly report sent via Telegram.")
schedule.every().day.at("23:55").do(_send_nightly)
logger.info("Nightly report scheduled at 23:55 UTC.")
while True:
schedule.run_pending()
time.sleep(60)
sched_thread = threading.Thread(target=_run_scheduler, daemon=True)
sched_thread.start()
# ── Blocking Telegram listener ────────────────────────────────────────────
run_listener()
def cmd_train_tcn(args):
"""Train the TCN signal confidence scorer (replaces LSTM ensemble).
The TCN learns from 100-bar sequences to predict whether the next bar's
direction will match the current HMM regime. At inference time the output
(a confidence multiplier in [0.7, 1.3]) scales the Z-Score cutoff:
effective_z = base_z × multiplier
multiplier < 1.0 → relaxed entry (strong, clear regime)
multiplier > 1.0 → tightened entry (noisy / uncertain regime)
Full training:
python main.py --mode train_tcn --tf H1 --broker headway_cent --epochs 100
Fine-tune on last 2 years:
python main.py --mode train_tcn --tf H1 --broker headway_cent \\
--epochs 20 --fine_tune --recent_years 2
"""
from src.engine_tcn import SignalConfidenceTCN, get_tcn_dir
from src.processor import load_data_with_hmm_labels
tf = args.tf.upper()
broker = args.broker
epochs = getattr(args, "epochs", 100)
seq_len = getattr(args, "seq_len", 100)
temperature = getattr(args, "temperature", 1.5)
fine_tune = getattr(args, "fine_tune", False)
recent_years = getattr(args, "recent_years", 2)
reconfigure_for_tf(tf)
try:
df = load_data_with_hmm_labels(tf, broker)
except FileNotFoundError as exc:
print(f"\nERROR: {exc}\n")
sys.exit(1)
n_states = int(df["hmm_state"].nunique())
save_dir = get_tcn_dir(tf, broker)
# TF-specific sequence lengths — CLI --seq_len overrides only if explicitly
# different from the default (100). TF defaults give the best context window.
TF_SEQ_LENGTHS = {"H1": 50, "M15": 80, "M5": 100}
default_seq_len = TF_SEQ_LENGTHS.get(tf, 100)
seq_len = default_seq_len if seq_len == 100 else seq_len
if seq_len != default_seq_len:
logger.info("Using CLI --seq_len=%d (TF default was %d)", seq_len, default_seq_len)
else:
logger.info("Using TF-default seq_len=%d for %s", seq_len, tf)
tcn = SignalConfidenceTCN(
seq_len=seq_len,
n_states=n_states,
temperature=temperature,
)
if fine_tune:
import os
if not os.path.exists(os.path.join(save_dir, "tcn_confidence_model.keras")):
print("\nERROR: No existing TCN model to fine-tune. "
"Run without --fine_tune first.\n")
sys.exit(1)
tcn.load(save_dir)
logger.info("Fine-tuning TCN [%s/%s] on last %d years…", tf, broker, recent_years)
tcn.fine_tune(
df, df["hmm_state"], df["log_return"],
epochs=epochs, recent_years=recent_years,
)
else:
logger.info("Training TCN [%s/%s] — %d bars n_states=%d T=%.1f",
tf, broker, len(df), n_states, temperature)
tcn.train(df, df["hmm_state"], df["log_return"], epochs=epochs)
tcn.save(save_dir)
# Quick inference test on last bar
recent_seq = df.iloc[-(tcn.seq_len + 10):]
multiplier = tcn.predict_confidence(recent_seq)
if multiplier is not None:
logger.info(
"Quick test — last bar confidence multiplier: %.3f "
"(Z-cutoff %.0f%% of base)",
multiplier, multiplier * 100,
)
print(f"\n TCN confidence model saved: {save_dir}")
print(f" Run --mode demo/live to use dynamic Z-Score adjustment.\n")
def cmd_sensitivity(args):
"""Z-Score sensitivity analysis on already-trained models.
Loops over Bull/Bear Z-Score cutoff values (1.5 .. 3.0 in 0.25 steps),
reruns the OOS backtest for each, and prints a comparison table so you can
see whether the current cutoff is optimal. MR (Chop) cutoffs are held
constant throughout.
Usage:
python main.py --mode sensitivity --tf H1 --broker headway_cent --balance 15
"""
from src.sensitivity import run_sensitivity
balance = _resolve_balance(args)
tf = args.tf.upper()
broker = args.broker
reconfigure_for_tf(tf)
# Load processed data + models exactly as cmd_report does
try:
params = get_best_params(balance=balance, broker=broker, tf=tf)
except Exception:
params = {}
df = process_pipeline(
obs_cov=params.get("obs_cov"), trans_cov=params.get("trans_cov"),
save=False, tf=tf,
)
_hmm_path = hmm_model_path(tf, broker)
if not _hmm_path.exists():
_hmm_path = None
try:
if _hmm_path is None:
raise FileNotFoundError
model_hmm = load_hmm(_hmm_path)
from src.engine_hmm import predict_states
states = predict_states(model_hmm, df)
except Exception:
logger.warning("No saved HMM found for %s/%s — fitting fresh.", tf, broker)
model_hmm, states, _ = fit_hmm(df, n_states=params.get("n_states", 3))
try:
_feat_scaler = load_feature_scaler(tf=tf, broker=broker)
except FileNotFoundError:
_feat_scaler = None
X, y, df_aligned, _ = prepare_features(df, states, feature_scaler=_feat_scaler, tf=tf)
_xgb_path = get_ensemble_path(tf, broker)
if not _xgb_path.exists():
_xgb_path = ENSEMBLE_PKL_PATH
try:
models_xgb, thresholds_xgb, metrics = load_xgb_ensemble(_xgb_path)
if metrics.get("feature_cols") != list(X.columns):
raise ValueError("Feature mismatch")
except (FileNotFoundError, ValueError):
logger.warning("No saved ensemble found for %s/%s — can't run sensitivity.", tf, broker)
return
_, probabilities = get_predictions_ensemble(models_xgb, thresholds_xgb, X)
states_aligned = states[df.index.isin(df_aligned.index)]
if not metrics.get("regime_stats"):
_rs_split = metrics.get("split_idx") or int(len(X) * 0.8)
metrics["regime_stats"] = compute_regime_stats(
models_xgb, thresholds_xgb,
X.iloc[:_rs_split], states_aligned[:_rs_split],
)
split_idx = metrics.get("split_idx")
if split_idx is None or not (0 < split_idx < len(X)):
split_idx = int(len(X) * 0.8)
run_sensitivity(
tf=tf,
broker=broker,
balance=balance,
df_aligned=df_aligned,
probabilities=probabilities,
states_aligned=states_aligned,
split_idx=split_idx,
regime_stats=metrics["regime_stats"],
)
def main():
parser = argparse.ArgumentParser(description="Gold Regime X — Hybrid ML Trading System")
parser.add_argument(
"--mode",
choices=["process", "optimize", "train", "compare", "export", "report",
"sync_validate", "demo", "live", "audit", "guardian", "listen",
"consolidate", "wfa", "sensitivity", "train_tcn", "extract_consensus"],
required=True,
)
parser.add_argument("--trials", type=int, default=250)
parser.add_argument("--n_jobs", type=int, default=1,
help="Parallel Optuna trial workers (default 1). "
"See optimizer.py for caveats with n_jobs>1.")
parser.add_argument("--interval", type=int, default=3600,
help="Guardian check interval in seconds (default 3600 = 1h).")
parser.add_argument("--min_cap", type=float, default=15.0,
help="Account capital in USD (legacy, use --balance)")
parser.add_argument("--balance", type=float, default=None,
help="Account balance in USD — overrides --min_cap")
parser.add_argument("--broker", type=str, default="standard",
choices=["standard", "headway_cent"])
parser.add_argument("--tf", type=str, default="H1",
help="Timeframe: H1 | M15 | M15,H1 (process/compare accept comma list)")
parser.add_argument("--period", type=str, default="3m",
help="Lookback window for MT5 sync, e.g. '3m' '6m' '12m'.")
parser.add_argument("--yes", action="store_true",
help="Skip the interactive live-account confirmation (used when launched as a subprocess).")
parser.add_argument("--profit_target", type=float, default=None,
help="Quick-profit close threshold in USD. M5 defaults to 4.0; "
"other TFs disabled unless set. Pass 0 to disable on M5.")
parser.add_argument("--skip_stale_check", action="store_true",
help="Bypass the model-staleness gate on --mode live/demo. "
"Use when intentionally running an older model (e.g. demo testing).")
parser.add_argument("--train_days", type=int, default=None,
help="WFA IS window in calendar days (default: H1=365, M15=180, M5=90).")
parser.add_argument("--test_days", type=int, default=None,
help="WFA OOS step size in calendar days (default: H1=90, M15=60, M5=30).")
parser.add_argument("--epochs", type=int, default=100,
help="TCN training epochs for --mode train_tcn (default 100).")
parser.add_argument("--seq_len", type=int, default=100,
help="TCN sequence length in bars (default 100).")
parser.add_argument("--temperature", type=float, default=1.5,
help="Temperature for TCN probability calibration (default 1.5). "
">1.0 softens; 1.0 = raw output; <1.0 sharpens.")
parser.add_argument("--fine_tune", action="store_true",
help="Fine-tune an existing TCN on recent data instead of full retraining.")
parser.add_argument("--recent_years", type=int, default=2,
help="Years of recent data used for --fine_tune (default 2).")
parser.add_argument("--fast_wfo", action="store_true",
help="Use faster WFO window sizes for --mode optimize and --mode wfa.")
parser.add_argument("--top_n", type=int, default=10,
help="Top-N trials to aggregate for --mode extract_consensus (default 10).")
parser.add_argument("--min_wfe", type=float, default=0.0,
help="Minimum WFE ratio filter for --mode extract_consensus (default 0).")
args = parser.parse_args()
{
"process": cmd_process,
"optimize": cmd_optimize,
"train": cmd_train,
"compare": cmd_compare,
"export": cmd_export,
"report": cmd_report,
"sync_validate": cmd_sync_validate,
"demo": cmd_demo,
"live": cmd_live,
"audit": cmd_audit,
"guardian": cmd_guardian,
"listen": cmd_listen,
"consolidate": cmd_consolidate,
"wfa": cmd_wfa,
"sensitivity": cmd_sensitivity,
"train_tcn": cmd_train_tcn,
"extract_consensus": cmd_extract_consensus,
}[args.mode](args)
if __name__ == "__main__":
main()