haelsisy/stock-signal

WatchSignal v4.0 — Multi-Horizon 5-Seed Ensemble

BiLSTM + Transformer hybrid with 4 parallel classification heads.

Architecture

• Input: [batch, 30, 55] float32 (30-candle window, 55 v3.1 features)
• Output: [batch, 4, 3] float32 (4 horizons × 3 classes: SELL/HOLD/BUY)
• Backbone: BiLSTM(55→256 bidirectional) + 2-layer Transformer (4 heads, pre-LN)
• Pooling: cat(last_step, mean_pool) → Linear(1024→512) → GELU
• Heads: 4 independent classifiers (512→3), one per horizon

Horizons

Horizon	Days	Buy Threshold	Sell Threshold
Short	5	+0.8%	-0.8%
Medium	10	+1.5%	-1.5%
Long	21	+3.0%	-3.0%
Macro	63	+5.0%	-5.0%

Ensemble

• Seeds: 42, 1, 7, 99, 2025
• Mean weighted accuracy: 0.5690144980676382% ± 0.008819085091941169%
• Per-horizon temperature calibration per seed (20 values total)

Out-of-Sample Backtest (2026-01-09 → 2026-04-10)

(see v4/mh_v4_backtest_report.json for full results)

Inference

• ONNX Runtime Mobile compatible (opset 13)
• Apply per-horizon temperature scaling from mh_v4_seed{X}_temperatures.json
• Ensemble: average softmax probabilities across 5 seeds, then argmax

Files (v4/ directory)

• mh_v4_seed{42,1,7,99,2025}.onnx — ONNX models (~27.7 MB each)
• mh_v4_seed{42,1,7,99,2025}_temperatures.json — per-horizon temperatures
• mh_v4_ensemble.json — ensemble manifest
• mh_v4_backtest_report.json — out-of-sample backtest results
• sha256_manifest.json — SHA-256 hashes for integrity verification

v3.1 Models (root directory)

Previous single-horizon models remain available in the repo root.