# CS318 Course Project — Synopsis

---

## Topic

**AI-Powered Stock Analysis Platform: Fine-Tuned BERT Sentiment and Cross-Sectional Momentum Strategy with Full-Stack Deployment**

---

## Group / Team Members

| S.No. | Name          | Roll No.   |
|-------|---------------|------------|
| 1.    | Harsh Yadav   | 23/CS/165  |

---

## Abstract

This project presents a full-stack deep learning application for stock market analysis that combines NLP-based sentiment analysis with a quantitative cross-sectional momentum trading strategy. The core deep learning component fine-tunes a pre-trained `bert-base-uncased` model on the Financial PhraseBank dataset (~4,840 expert-annotated sentences) for three-class financial sentiment classification (Positive / Neutral / Negative), using a custom classification head (768→256→3 with ReLU and Dropout) trained with AdamW optimizer and linear warmup scheduling. The fine-tuned model achieves **97.1% accuracy and 97.1% F1-score**, outperforming the industry-standard pre-built FinBERT (93.2% accuracy) by 4 percentage points, with a decisive +18.4% advantage in detecting positive financial sentiment. The quantitative module implements a cross-sectional momentum strategy that ranks ~300 S&P 500 stocks monthly using a composite alpha signal, going long the top quintile. Backtested over 2008–2025, the strategy achieves a **14.9% CAGR and 0.90 Sharpe ratio**, outperforming the S&P 500 Buy & Hold benchmark (11.6% CAGR, 0.78 Sharpe), growing $1 to $11.49 vs $6.85. Both models are deployed in a Flask REST API backend with live news via GNews API and market data via Twelve Data API. The React.js (Vite) frontend provides an interactive dashboard with stock charts, sentiment feeds, and cross-sectional rankings. The application is Dockerized and deployed on Hugging Face Spaces with Supabase caching.

*(~200 words)*

---

## Sign of Student

| S.No. | Name          | Signature         |
|-------|---------------|-------------------|
| 1.    | Harsh Yadav   | _________________ |

---