Abstract
<title>Abstract</title> <p>The prediction of financial market behavior is difficult because current transformer models have largely overlooked neutral sentiment and domain-specific lexicons. This study presents a two-phase hybrid framework with DeBERTa-v3 financial sentiment classification along with NIFTY 50 technical indicators. Phase A uses a DeBERTa-v3 model with dual-pooling encoder, gated lexicon fusion, Stochastic Weight Averaging (SWA), and capitalization-based test-time augmentation (TTA), fine-tuned on the SEntFiN dataset, with accuracy and macro F1-score of 90.90% and 90.88%, respectively. Inference of batch data containing 198,730 headlines published by Indian and international sources between January 2017 and April 2021 shows that 40.86% of the headlines convey neutral sentiment, a factor usually ignored in binary classification models. Phase B creates a daily sentiment index based on the output of the three-class DeBERTa sentiment model, followed by the fusion of the sentiment score and 15 technical indicators for the NIFTY 50 market. A leakage-free time-series five-fold cross-validation technique was applied to six different predictive architectures, which revealed Logistic Regression with Exponential Moving Average (EMA) smoothing of the features achieves the highest cross-validated accuracy of 0.594 ± 0.032 with directional accuracy of 62.98% on an 80/20 hold-out validation split. The paired t-test confirms that sentiment features extracted from the DeBERTa-v3 model provide statistically significant gains relative to the price-only baseline model.</p>