Integrating Transformer-Derived Sentiment Signals with Technical Indicators for Emerging Market Direction Prediction: A NIFTY 50 Study

Technical Summary: Integrating Transformer-Derived Sentiment Signals with Technical Indicators for Emerging Market Direction Prediction

Problem Statement
Financial market prediction remains challenging due to the limitations of existing transformer models, which often overlook neutral sentiment and fail to incorporate domain-specific lexicons. While momentum-based technical indicators (e.g., RSI, MACD) detect price patterns, they are reactionary and ignore sentiment generated by news. Furthermore, prior research in financial Natural Language Processing (NLP) frequently relies on binary classification (excluding neutral classes), lacks integration of domain knowledge, and often fails to provide formal statistical proof linking sentiment to stock prices, particularly within emerging markets like the NIFTY 50.

Methodology
The study proposes a two-phase hybrid framework designed to bridge the gap between transformer-based sentiment analysis and technical market indicators.

• Phase A: Hybrid Sentiment Classification

  • Architecture: A modified DeBERTa-v3 model is employed, enhanced with a dual-pooling encoder (concatenating CLS token and average token embeddings) and a gated lexicon fusion mechanism. The lexicon fusion integrates 14 scalar features derived from a financial lexicon (including counts of positive, negative, upward, downward, negation, and uncertainty terms).
  • Training Techniques: The model utilizes Stochastic Weight Averaging (SWA) to converge to flatter loss minima for better generalization. Test-Time Augmentation (TTA) based on capitalization variations is applied to reduce prediction variance.
  • Training Data: The model is fine-tuned on the SEntFiN dataset (14,404 samples), which includes entity-level annotations for positive, negative, and neutral sentiments.
  • Inference: The trained model processes 198,730 financial headlines from Indian and international sources (Jan 2017–Apr 2021) to generate a Daily Sentiment Index (DSI). Four indices are constructed: a simple net sentiment ($s_1$), a confidence-weighted net sentiment ($s_2$), a three-category probabilistic index ($s_3$), and a 30-day moving average ($s_4$).

• Phase B: Market Direction Prediction

  • Feature Engineering: The sentiment indices are fused with 15 NIFTY 50 technical indicators (including RSI, MACD, Bollinger Bands, EMA, ATR, OBV) and derived features like log returns and volatility. Lag features (1st, 2nd, and 3rd order) are applied to ensure temporal causality.
  • Experimental Design: Six predictive architectures are evaluated using a leakage-free Time Series Split (five-fold cross-validation) and an 80/20 chronological hold-out. The models include Vanilla LSTM, CNN-LSTM, GRU with Attention, Ensemble LR-LSTM, Standard Logistic Regression, and an Optimized Logistic Regression with Exponential Moving Average (EMA) smoothing.
  • Statistical Validation: Performance is assessed via accuracy and directional accuracy. Statistical significance is tested using paired t-tests (comparing price-only vs. sentiment-augmented models) and OLS regression with HC3-robust standard errors.

Key Contributions

1. Novel Hybrid Architecture: Introduction of a DeBERTa-v3 framework incorporating dual-pooling, gated lexicon fusion, SWA, and TTA specifically for three-class financial sentiment analysis.
2. First Transformer-Derived DSI for NIFTY 50: Creation of a Daily Sentiment Index based on 198,730 multi-source headlines, revealing that 40.86% of financial news conveys neutral sentiment—a factor often ignored in binary models.
3. Rigorous Comparative Analysis: A leakage-free comparison of six predictive models using sentiment-augmented features, demonstrating that simpler linear models with EMA smoothing can outperform complex deep learning architectures on this dataset.
4. Formal Statistical Testing: Application of paired t-tests and HC3-robust OLS regression to validate the statistical significance of sentiment features and analyze their relationship with market returns.

Results

• Sentiment Model Performance: The Lexicon Fusion Model achieved an accuracy of 90.90% and a macro F1-score of 90.88% on the SEntFiN test set. The neutral class, often the most ambiguous, was classified with 86.8% accuracy.
• Market Prediction Performance: The Optimized Logistic Regression with EMA features achieved the highest cross-validated accuracy of 0.594 ± 0.032 and a directional accuracy of 62.98% on the hold-out set. This outperformed all deep learning models (LSTM, CNN-LSTM, GRU) and the majority class baseline (53.10%).
• Statistical Findings:
  • A paired t-test confirmed that the sentiment-augmented model provided a statistically significant improvement ($p < 0.05$) over the price-only baseline. The Optimized LR model achieved 59.54% accuracy compared to 50.87% for the Price Only (LR) baseline, representing an increase of approximately 8.67 percentage points.
  • Conversely, OLS regression of the sentiment index against daily returns yielded a non-significant correlation ($p = 0.70$, $R^2 = 0.003$), supporting the Efficient Market Hypothesis at daily frequencies. The authors note this discrepancy arises because OLS tests a univariate linear relationship, whereas the t-test evaluates the utility of a multi-dimensional feature vector in a nonlinear model.

Significance and Claims
The paper claims that integrating transformer-derived sentiment signals with technical indicators offers a statistically significant advantage in predicting NIFTY 50 market direction, particularly when utilizing a three-class sentiment framework that accounts for neutral news. The study highlights that while daily market efficiency may obscure simple linear correlations between sentiment and returns, sentiment features provide valuable discriminative utility in a multi-dimensional predictive context. The research fills a gap in the literature by addressing the lack of neutral sentiment analysis, the omission of domain-specific lexicons, and the absence of formal statistical validation in Indian market studies.

Limitations
The authors acknowledge limitations including the lack of hourly timestamps (hindering precise news-to-price matching), the exclusion of social media sentiment and regulatory filings, and the absence of regime-specific analysis (e.g., during the COVID-19 outbreak). The daily frequency of the analysis also ignores intra-day relationships between sentiment and price.