Which financial language model should an algorithmic trading system trust? And how should sentiment modulate a price signal without flipping its fundamental direction? This paper introduces empirical model selection paired with directional signal fusion.
A common pitfall in sentiment-augmented trading models is signal inversion: high-confidence sentiment can erroneously flip a short signal to a long position or vice versa, bypassing technical risk controls.
We propose Sign-Aware Multiplicative Modulation. In this formulation, sentiment modulates the conviction (magnitude) of the underlying price signal, while preserving its directional integrity:
Where Sprice ∈ [-1, +1] is the base market signal, Ssentiment ∈ [-1, +1] is the calibrated sentiment polarity, and α ∈ [0, 1] is the modulation intensity factor.
When sentiment agrees with the market direction (e.g. positive sentiment on a long signal), conviction is amplified. When sentiment contradicts market technicals, the signal is dampened toward zero without reversing direction.
Candidate architectures were evaluated on held-out stratified splits of the Financial PhraseBank benchmark across classification accuracy, macro-F1, and ordinal Mean Absolute Error (MAE):
| Model Checkpoint | Accuracy | Macro F1 | Ordinal MAE |
|---|---|---|---|
| DeBERTaTop Candidate | 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 proves that a "financial" label on Hugging Face is no guarantee of reliability; domain-specific benchmarking is an mandatory pre-condition before production deployment.
If you build upon this work in academic or quantitative research, please cite:
@inproceedings{elmolla2026sentiment,
author = {Mohamed Ehab El Molla},
title = {Automatic Model Selection and Sign-Aware Multiplicative Sentiment Modulation for Algorithmic Trading},
booktitle = {Proceedings of the IEEE International Conference on Intelligent Multi-System Applications (IMSA)},
year = {2026},
publisher = {IEEE},
doi = {10.1109/IMSA.2026.11700381}
}