AI4Invest-Indian-FinBERT

AI4Invest-Indian-FinBERT

🧠 Model Overview

AI4Invest-Indian-FinBERT is an institutional-grade financial sentiment analysis model specifically optimized for the Indian stock market (NSE/BSE).

Unlike other regional models trained on LLM-hallucinated or synthetic labels, this architecture was fine-tuned exclusively on a purely human-annotated dataset, ensuring mathematical reliability and zero pseudo-labeling bias. It classifies financial headlines and market chatter into three distinct categories: Bullish, Bearish, and Neutral.

🏗️ Architecture & Training Mechanics

• Base Model: ProsusAI/finbert
• Training Data: 11,931 human-annotated financial headlines/tweets (zeroshot/twitter-financial-news-sentiment).
• Learning Rate: 2e-5 (Safe descent to preserve foundational English grammar).
• Epochs: 3 (Optimized via Early Stopping to completely prevent Overfitting).
• Weight Decay: 0.01 (L2 Regularization applied to prevent buzzword-fixation).
• Batch Size: 16

📊 Institutional Benchmarks (The Gladiatorial Arena)

To prove deployment readiness, this model was subjected to a blind evaluation on a strictly held-out 20% validation set (2,387 unseen headlines) alongside two industry baselines.

Note: The Matthews Correlation Coefficient (MCC) of 0.73 mathematically validates that the model deeply understands contextual nuance and is not succumbing to the accuracy paradox by over-predicting the "Neutral" majority class.

📈 Confusion Matrix Telemetry

          Predicted
          Bullish   Bearish   Neutral
        +-------+-------+-------+
Bullish |  399  |   22  |   59  |
Bearish |   25  |  283  |   50  |
Neutral |   86  |   93  | 1370  |
        +-------+-------+-------+