Experimental Results & Ablation Study
Systematic empirical evaluation comparing 4 architectural variants across 16 technical, sentiment, and event features.
Hybrid RAG-LSTM Model
Event-Augmented Variant
Production Streaming Ready
8 to 16 Features
Model Variant Benchmark Matrix
Empirical testing with 80/20 deterministic train-test split on daily OHLCV and event news series
| Model Variant | Features | Accuracy | Precision | Recall | F1-Score | ROC-AUC | Latency |
|---|
Visual Metric Comparison
Measures sensitivity to price directional changes. Crucial in trading to avoid missing catastrophic downturns.
Confusion Matrix
SelectedClick any model row in the matrix table above to inspect its classification error distribution.
In quantitative trading, a False Negative represents holding through an unexpected crash catalyst. Our 95.65% Recall significantly outperforms pure technical models.
With FAISS C++ vector indexing and compact 384-dim embeddings, inference runs in just 2.16 milliseconds per sample on standard CPUs without requiring high-end GPUs.
Standard Cross-Entropy fails on choppy sideways markets. Dynamic Focal Loss (γ=2.0, α=0.25) penalizes hard-to-classify edge days, preventing model collapse.