Automatic Model Selection and Sign-Aware Multiplicative Sentiment Modulation for Algorithmic Trading
Which financial language model should a trading system trust? And how should sentiment change a price signal without reversing its direction? This work connects automatic model benchmarking with sign-aware signal fusion.
Benchmark before selecting.
Fine-tuned DeBERTa-finance reached 94.2% accuracy and 0.9420 macro-F1, a 5.0 percentage-point accuracy lead over FinBERT. Candidates are evaluated using accuracy, macro-F1, and ordinal MAE.
Preserve direction. Modulate conviction.
Sign-aware multiplicative fusion strengthens a price signal when sentiment agrees with its long or short direction and attenuates it when sentiment disagrees. Sentiment changes conviction while preserving direction.
IEEE IMSA 2026 / July 2026 / IEEE document 11700381
| Model | Accuracy | Macro-F1 | MAE ↓ |
|---|---|---|---|
| DeBERTa | 94.2% | 0.9420 | 0.060 |
| FinBERT | 89.2% | 0.8932 | 0.114 |
| pmatorras | 12.5% | 0.028 | 1.157 |
An 81.7 percentage-point gap between the best and worst candidates highlights why a finance label is not a substitute for evaluation. The pmatorras checkpoint scores below the 33% random baseline.
These are held-out sentiment-classification results, not evidence of trading profitability or live financial performance.
