📈 Stock Market Analysis using Regression, Classification, and Fusion Techniques

This project explores stock price prediction using both numerical stock data and textual news headlines. We implemented a multi-model pipeline involving polynomial regression, custom random forest classification, and a fusion neural network that combines both data types for enhanced performance. The focus is on Baidu stock prices and their related news headlines.

---

Project Overview

• Regression Model: Polynomial Regression (degree=10) for predicting stock prices using OHLC data.
• Classification Model: Random Forest Classifier using TF-IDF on news headlines to extract sentiment signals.
• Fusion Model: A neural architecture combining both numerical and textual features for superior stock price prediction.

---

1. Polynomial Regression (Numerical Data)

Highlights:

• MSE: 24.38
• RMSE: 4.94
• Significantly outperformed KNN and Random Forest Regressors.

Techniques Used:

• Feature Engineering: 10-day rolling average, polynomial terms.
• Custom implementation of gradient descent.
• Visualizations: Training fit, prediction error, full dataset view, learning curve.

---

2. Text Classification (News Headlines)

Highlights:

• Best Accuracy (Random Forest): ~72%
• Outperformed SVM and Weighted K-NN.

Techniques Used:

• Custom preprocessing: Tokenization, Lemmatization, Stopword filtering.
• Feature Extraction: TF-IDF (Max features: 5000).
• Custom-built Random Forest Classifier (200 trees, depth 10, Gini impurity).

---

3. Fusion Neural Network (Multimodal)

Highlights:

• Fusion Model MSE: 38.24
• Generalization error improved by combining structured + unstructured data.

Architecture:

• Input: Concatenated numerical + textual features.
• Hidden Layers: ReLU + Tanh
• Output: Linear price prediction
• Optimizer: Gradient Descent (Manual backprop)

---

Final Results Summary

Model Type	Metric	Value
Polynomial Regression	Test MSE	24.38
Polynomial Regression	RMSE	4.94
Text Classifier	Accuracy	~70%
Fusion Model	Test MSE	38.24