TODO

​

Current state — Phase 15 (active)

​

OOS performance

​

Phase 15 — Production transition (active)

​

15.0 Broker-agnostic refactor

• config/brokers.json — central broker registry (VT Markets, IC Markets, Exness, FTMO, FundingPips, generic)
• config.json → "broker" key selects active broker; all scripts load spread/leverage/swap from registry
• backtest_config.py — config-faithful backtester; use --max-lot/--cent-account for preset-style comparisons
• scripts/sweep.py — unified sweep replacing sweep_signal.py + sweep_pos_model.py; --target signal|pos|both
• scripts/check_broker_limits.py — fixed main guard; added --symbol flag

​

15.1 Position model validation

• 38d OOS: PF=4.71 vs baseline 4.27 (+10.3%), MaxDD=4.7%, 1 ML_Exit
• Sweep of 13 configs via scripts/sweep.py --target pos
• Thresholds updated: exit=0.80, min_prob_diff=0.25, min_bars_held=4

​

15.2 Automated retrain pipeline

• scripts/weekly_optimize.py — 13-phase autonomous pipeline (SHAP → tune → retrain → sweep → evaluate → push → commit → notify)
• Score improvement gate: 2% minimum to keep new models

0 2 * * 0  cd /path/NOVOSKY && .venv/bin/python scripts/weekly_optimize.py >> logs/weekly.log 2>&1

​

15.3 Cloud monitoring

• scripts/performance_monitor.py — --alert, --lookback, --min-wr, --min-pf, --max-cl flags
• Thresholds: WR < 55%, PF < 2.0, consecutive losses ≥ 7

​

15.4 Telegram bot commands

​

15.5 Broker safety audit

• Run python scripts/check_broker_limits.py against VT Markets live server
• Confirm BTCUSD tick size, stops_level, max/min lot
• Verify swap-free status

​

15.6 Weekly validation cadence

​

In progress

• VM cron (15.2/15.3/15.6): Wire weekly_optimize.py + performance_monitor.py on trading VM
• Phase 15.5: Broker safety audit against live VT Markets — run python scripts/check_broker_limits.py

​

Phase 16 — Signal quality & risk management

​

16.1 Walk-forward validation

• Walk-forward backtest — train 6 months, test next month, roll N times
• Gate model release on median PF > 2.0, all folds WR > 55%

​

16.2 Lot sizing improvements

• Kelly lot sizing backtest (via backtest_config.py lot-cap sweep)
• Confidence-based position sizing — 60/70/80% tiers scale lot size

​

16.3 Drawdown guards

• Enable max_daily_loss — suggested: 15 USD (~3% of $500)
• Equity curve filter — 10-trade rolling drawdown pause

​

16.4 Multi-account support

• Add --account <id> CLI flag to trading.py to select terminal from config/terminals.json
• Each broker/account runs its own mt5api instance on a separate port (e.g. 6542, 6543)
• PM2 multi-instance config — one trading.py process per account
• Shared model files, per-account state_{account}.json for isolated position tracking
• Per-account risk config override (different risk% per account/broker)

​

Phase 17 — Feature engineering

​

17.1 On-chain & derivatives features

• funding_rate — Binance perpetual swap funding rate
• oi_change — Open interest delta from Binance
• fear_greed_index — Daily CNN Fear & Greed (0–100)
• liq_distance_buy/sell — Distance to liquidation cascade zones (Coinglass)

​

17.2 Market regime features

• volatility_regime — ATR percentile bucket (Low/Medium/High)
• corr_spy_20 / corr_dxy_20 — Rolling BTC vs S&P500/DXY correlation
• w1_ema_bias / w1_rsi_norm — Weekly timeframe features

​

17.3 Price structure features

• h4_macd_hist / d1_macd_hist — MACD histogram (not just direction)
• dom_sin / dom_cos — Day-of-month cyclical encoding
• ofi — Synthetic order flow imbalance proxy

​

17.4 Multi-datasource OHLCV pipeline

• Design consensus/averaging logic (median OHLCV across sources, flag divergence > 0.5%)
• ml/data_sources/ module — pluggable source adapters
• Conflict resolution: use MT5 as primary; fill gaps with secondary sources
• Impact: more data, cross-broker validation, reduced broker-specific bias

​

Phase 18 — Model architecture

​

18.1 Ensemble upgrades

• Stacked meta-learner — replace majority-vote; target +2–5% OOS accuracy
• Calibrated probabilities via CalibratedClassifierCV(method='isotonic')
• Ensemble weighting by recency (last 1,000 bars validation accuracy)

​

18.2 Training improvements

• Increase Optuna to 150 trials + MedianPruner
• Online/incremental weekly updates via XGBoost/LGB native continuation
• Regime-switching model (ADX > 0.25 trending vs ranging)

​

18.3 Multi-instrument expansion

• ml_config_{SYMBOL}.json — instrument-specific feature list and labeling params
• models/{SYMBOL}/ — separate ensemble, position, risk model files
• datasets/training_data_{SYMBOL}.csv — instrument-specific OHLCV dataset
• Separate backtest_config.py --symbol {SYMBOL} run for OOS validation

• XAUUSD (gold CFD) — different ATR ($5–30 vs$50–200 BTC), different session sensitivity
• EURUSD (forex) — pip-based P&L, tighter spreads, different news calendar
• Cross-asset features (BTC–Gold correlation, BTC–ETH spread) — Phase 17 first

​

Phase 19 — Infrastructure & operations

​

19.1 Observability

• Live Streamlit dashboard — equity curve, open positions, signal feed
• Grafana + InfluxDB time-series metrics pipeline
• /health HTTP endpoint from trading process

​

19.2 Reliability

• API failover — alert + SAFE_PAUSE after 3 consecutive failures
• Graceful shutdown — finish current bar, cancel pending orders, write shutdown state
• Config hot-reload — poll config.json mtime every 60s

​

Phase 20 — MQL5 EA (standalone native client)

​

20.1–20.2 Core EA framework + scaler

• mql5/src/NovoskyEA.mq5 skeleton with OnInit/OnDeinit/OnTick
• Load all 6 ONNX models via OnnxCreate() with #resource paths
• ScalerConstants.mqh auto-generated from scripts/extract_scaler_constants.py

​

20.3 Feature computation in MQL5

• Replicate all 56 features in native MQL5
• Multi-timeframe resampling (H4, D1 from M15) without API calls
• Parity test: MAE < 0.001 vs Python on same 100 bars

​

20.4–20.5 Signal + position inference

• Majority-vote inference over 3 signal ONNX models
• ATR-based SL/TP + trade_stops_level enforcement
• 59-feature position inference — EXIT on threshold

​

20.6 Distribution

• Package EA + resources ZIP into single EX5 under 20 MB
• Submit Free tier (signal only) + Premium tier (signal + position) to MQL5 Market

​

Completed phases (1–12)

​

Phase 1 — Fix the model

• SELL recall improved 3% → 30% by adding 5 directional features: ema_stack, candle_direction, volume_delta, rsi_slope, consecutive_direction
• TP=0.3% / SL=0.25% tuning: WR 60.6%→77.3%, PF 1.48→2.85, MaxDD 12%→6.3%
• Training data extended 365 → 730 days; lookahead_candles 12 → 24
• LightGBM feature-name warning fixed (pass numpy array directly)

​

Phase 2 — Multi-timeframe features

• H4/D1 features resampled from M15: h4_ema_bias, h4_rsi_norm (#5 SHAP), h4_macd_dir, d1_trend, price_vs_d1_open (#7 SHAP)
• Session flags: is_london_session, is_ny_session, is_asian_session, session_hour_sin/cos
• Volatility/momentum: atr_percentile (top feature), volume_surge, bb_squeeze, price_acceleration
• S/R proximity: dist_to_round_number (#1 SHAP overall), near_daily_high_low, adx_14, market_quality, momentum_decay, adverse_candle_ratio
• ATR-aware labeling groundwork: create_labels_atr_aware() added (activated in Phase 10)

​

Phase 3 — Hyperparameter tuning + SHAP

• Optuna 50-trial tuning: WF 44.9%→45.7%, PF 2.82→2.96, MaxDD 7.6%→3.8%, Sharpe 7.23→8.63
• SHAP analysis: ml/shap_analysis.py with TreeExplainer; beeswarm plots; models/shap_summary.json
• Top SHAP features: h4_rsi_norm > atr_14 > hourly_return > price_vs_ema200 > session_hour_sin
• Feature pruning tested and rejected — removing low-SHAP features hurt performance (WR 55.9%→50%)
• Sequence model deferred (CPU-only hardware too slow)

​

Phase 4 — Risk management (superseded)

• EMA trend filter, partial profit taking, trailing stop — all superseded by ML active management
• max_weekly_drawdown_pct added to config.json (2026-04-11)

​

Phase 5 — Backtesting

• backtest_config.py built: config-faithful OOS backtester with ONNX inference; WR, PF, Sharpe, MaxDD, Return
• v8 results (1yr OOS, 48 features): Setup A WR=76.4% PF=3.20 MaxDD=2.6% | Setup D Return=+747%
• Walk-forward backtest deferred to Phase 16

​

Phase 6 — Automation & monitoring

• scripts/performance_monitor.py — WR/PF/consecutive losses; Telegram alert
• Signal logging to models/ml_performance.csv
• scripts/dashboard.py CLI trade summary

​

Phase 7 — Live trading integration

• Migrated trading.py from MetaTrader5 Python package to NOVOSKY HTTP API
• --dry flag added for safe testing
• Paper trade validated via OOS backtest: WR=82.7%, PF=4.27, MaxDD=5.2%

​

Phase 8 — ML active trade management (2026-04-10)

• Dedicated position model: RF+XGB+LGB on 60 features (56 market + 4 position-state)
• Labels: HOLD / EXIT / ADD — ml/position_labeling.py + ml/position_trainer.py
• PositionPredictor.get_position_action() with 2/3 majority vote
• Kelly-adjusted lot sizing, ML-based SL/TP scaling, partial close, trailing stop
• Results: Signal WF=43.66%, Position ensemble=73.50% | Setup E: WR=78.8% PF=4.52 MaxDD=1.6%

​

Phase 9 — Growth config sweep (2026-04-10)

• Goal: maximize monthly return for $10k account, IC Markets RAW
• Best: conf=0.55, risk=20%, max_lot=10, all Phase 8 active management disabled
• Result: WR=57.4%, PF=1.75, MaxDD=15.0%, Sharpe=4.68, Return=+8449%, ~340 trades/yr

​

Phase 10 — Deep optimization (2026-04-11)

1. Label-execution mismatch — activated atr_aware labeling in ml_config.json
2. Class imbalance — replaced downsample with compute_sample_weight('balanced')
3. Spread underestimation — backtest_config.py spread fallback $0.30 →$16.95
4. ADX regime filter — new adx_regime_filter block in config.json

​

Phase 11 — M15 scalping + local timezone support (2026-04-11)

• H1 → M15 timeframe; 112 trades/yr → 477 trades/yr (1.31/day)
• ATR-aware labels: SL=0.8×ATR, TP=1.5×ATR, lookahead=48 bars
• sl_atr_multiplier 1.0→0.8; min_atr 50→15; ADX filter disabled
• Local timezone support via config.json; Telegram redesign
• Backtest: WR=63.7%, PF=3.05, MaxDD=19.6%, Sharpe=7.29, Return=+1,284,866%

​

Phase 12 — Production hardening (2026-04-12)

• API retry with exponential backoff (3× on network errors)
• Atomic state.json writes via .tmp → os.replace()
• tracked_positions persisted to state.json — survives PM2 restarts
• risk_percent 6→2; max_consecutive_losses 10→5; max_weekly_drawdown_pct 0→20
• Full retrain: fresh 2yr M15 data, Optuna 50-trial local tuning, OOS backtest