sato2ru/omega-v4-trading

OMEGA v4 — Crypto Trading Signal Model (Baseline)

Summary

Binary directional classifier (Up / Down) for 5 crypto pairs at 1h resolution. Ships as a documented baseline — evaluation revealed the model's confidence scores are not predictive of accuracy at this resolution, making Kelly sizing unreliable. v5 will address this with 4h bars and news masking.

Architecture

• CNN feature extractor → RegimeGate → Transformer encoder
• 184K parameters, D_MODEL=64, 3 layers, 4 heads
• Input: SEQ_LEN=24 bars × 40 ratio-based features + 5 regime features

Training

• Symbols: BTC, ETH, SOL, XRP, ADA
• Bars: ~9,800 × 5 symbols (1h, ~14 months)
• Split: 70/15/15 chronological
• Optimizer: Adam, LR=3e-4, label_smoothing=0.05, PATIENCE=20
• Device: NVIDIA Tesla T4

Evaluation (test set)

Metric	Value
Accuracy	51.6%
ROC-AUC	51.6%
Brier score	0.2997
Temperature (calibration)	5.01 (hit clamp ceiling)

Known Limitations

• Confidence scores are not predictive of accuracy (flat acc-vs-conf curve)
• Temperature scaling hit ceiling (T=5.0) — logits are severely compressed
• Per-symbol hi-conf accuracy is at or below 50% for 4/5 symbols
• Kelly sizing is not recommended with this checkpoint

Artefacts

File	Description
omega_v4_weights.pt	Model weights + config
omega_v4_scalers.pkl	Per-symbol StandardScalers
omega_v4_calibration.json	Temperature + thresholds
omega_v4_features.json	Feature list + regime cols

Roadmap

v5 changes planned:

• 4h bar resolution (cleaner signal, less noise)
• ForexFactory news masking (remove high-impact event bars)
• Larger dataset (N_PAGES=8, ~20k bars per symbol)
• Revisit SEQ_LEN=48 at 4h resolution