📈 Stock Market Prediction Using Machine Learning

Predicting TCS stock price movements using classical ML algorithms — regression, classification, and sentiment analysis combined.

---

📖 Table of Contents

• Overview
• Key Results
• Project Structure
• Dataset
• Models Implemented
  • Regression Models
  • Classification Models
• Sentiment Analysis Pipeline
• Feature Engineering
• Getting Started
• Model Comparison
• Key Findings

---

🌐 Overview

The stock market is highly volatile and nonlinear — classical ML models struggle, but with the right feature engineering they can still extract meaningful signals.

This project applies 6 different ML algorithms (3 regression + 3 classification) to TCS (Tata Consultancy Services) historical stock data, enriched with news sentiment scores from a fine-tuned FinBERT model. Each script is self-contained: load data → preprocess → train → evaluate.

Goals

• Predict the next day's opening price (regression)
• Predict whether price will go UP or DOWN the next day (classification)
• Measure the impact of lag features and sentiment on prediction quality
• Compare all models side-by-side with a visualizer script

---

🏆 Key Results

Regression (Predicting Next Day's Price)

Model	R² Score	RMSE
Linear Regression	0.9848	49.81 ✅ Best
Random Forest Regressor	0.6836	227.71
KNN Regressor	0.3824	318.15

Classification (Predicting UP/DOWN Direction)

Model	Accuracy	Notes
Random Forest	75.68% ✅ Best	Highest overall
Logistic Regression	75.09%	Best precision/recall balance
KNN Classifier	70.99%	Improved significantly with lag features

Baseline: ~50% (coin flip). All models beat the baseline, with Random Forest and Logistic Regression reaching ~75%.

---

📁 Project Structure

Stock-Market-ML/
│
├── 📊 Data
│   └── Stocks_TCS2.csv              # Main dataset: OHLCV + Sentiment scores
│
├── 🤖 Models
│   ├── StockMarket_LinearRegression.py      # LR for price prediction
│   ├── StockMarket_LogisticRegression.py    # LogReg for direction (+ ROC curve, confusion matrix)
│   ├── StockMarket_KNN.py                   # KNN classifier
│   ├── StockMarket_KNNReggresor.py          # KNN regressor
│   ├── StockMarket_RandomForestClassifier.py # RFC with decision tree visualization
│   └── StockMarket_RandomForestRegressor.py  # RFR
│
├── 🔧 Data Pipeline
│   ├── textb.py         # Downloads TCS stock data via yfinance
│   ├── merge.py         # Merges stock data with FinBERT sentiment scores
│   └── remover.py       # Cleans rows with zero sentiment values
│
├── 📈 Analysis
│   └── comparizer.py    # Side-by-side bar charts for all model results
│
├── requirements.txt
└── README.md

---

📦 Dataset

File: Stocks_TCS2.csv

Column	Description
Close	Closing price
High	Daily high price
Low	Daily low price
Open	Opening price
Volume	Trading volume
Sentiment	FinBERT sentiment score from TCS news headlines (-1 to +1)

• Source: TCS.NS historical data via yfinance (2015–2025)
• News Source: CF-AN equities TCS headlines dataset
• Sentiment Model: ProsusAI/FinBERT — a BERT model fine-tuned on financial text

How Sentiment Was Calculated

Positive headline → +score
Negative headline → -score
Neutral headline  →  0
Daily sentiment   = mean of all headline scores for that trading day

---

🤖 Models Implemented

Regression Models (Predict Price)

All regression models predict the next day's opening price using Open.shift(-1) as the target.

Linear Regression

Target: Next day's Open price
Features: OHLCV columns
Result: R² = 0.9848, RMSE = 49.81
Note: Lag features slightly hurt LR performance — dropped for this model

KNN Regressor

Target: Next day's Open price
Features: OHLCV + lag features
n_neighbors: 10
Scaler: StandardScaler (required for KNN)
Result: R² = 0.3824, RMSE = 318.15

Random Forest Regressor

Target: Next day's Open price
Features: OHLCV + lag features
n_estimators: 100
Result: R² = 0.6836, RMSE = 227.71

---

Classification Models (Predict Direction)

All classification models predict 1 (price UP) or 0 (price DOWN) for the next day.

Target = (Open.shift(-1) > Open).astype(int)

KNN Classifier

n_neighbors: 10
Scaler: StandardScaler
Accuracy: ~70.99%
Impact of lag features: +20% accuracy improvement (from ~50% to ~70%)

Logistic Regression

max_iter: 1200
class_weight: balanced
Outputs: Confusion Matrix + ROC Curve
Accuracy: ~75.09%
Note: Lag features had minimal impact on LogReg

Random Forest Classifier

n_estimators: 200
class_weight: balanced
Outputs: Decision tree visualization
Accuracy: ~75.68%
Impact of lag features: Huge — from ~55% to ~70%+

---

💬 Sentiment Analysis Pipeline

The sentiment pipeline uses FinBERT to score TCS news headlines and append daily sentiment to the stock dataset.

textb.py     →  Downloads TCS.NS price data from yfinance → saves TCS_STOCKS!2.csv
     ↓
merge.py     →  Loads headlines CSV + stock CSV
             →  Filters stocks to only dates with news coverage
             →  Runs FinBERT on each headline
             →  Averages daily sentiment scores
             →  Saves Stocks_TCS2_date.csv
     ↓
remover.py   →  Removes rows where sentiment = 0 (no news day)
             →  Saves Stocks_TCS4.csv

⚠️ GPU recommended — merge.py runs FinBERT with device=0 (CUDA). Switch to device=-1 for CPU.

---

🔧 Feature Engineering

Lag features are the most impactful preprocessing step in this project:

stocks['lag_1']       = stocks['Open'].shift(1)       # Yesterday's open
stocks['lag_2']       = stocks['Open'].shift(2)       # Two days ago open
stocks['rolling_3']   = stocks['Open'].rolling(3).mean()  # 3-day moving average
stocks['per%_change'] = stocks['Close'].pct_change()  # Daily % change

Impact of Lag Features by Model

Model	Without Lag Features	With Lag Features	Δ
KNN Classifier	~50%	~70%	+20%
Random Forest Classifier	~55%	~75%	+20%
Logistic Regression	~75%	~75%	~0%
Linear Regression	0.985 R²	0.985 R²	Slightly worse
KNN Regressor	baseline	+0.01 R², -800 RMSE	Slight improvement
Random Forest Regressor	baseline	+0.003 R², -1000 RMSE	Marginal improvement

---

🚀 Getting Started

Prerequisites

Python >= 3.11
CUDA GPU (optional, for faster FinBERT inference)

Installation

git clone https://github.com/your-username/Stock-Market-ML.git
cd Stock-Market-ML
pip install -r requirements.txt

requirements.txt

pandas
numpy
scikit-learn
matplotlib
transformers
yfinance
statistics
torch  # for FinBERT GPU inference

Note: The requirements.txt lists Transformer — install as transformers (lowercase) from Hugging Face.

Running the Models

# Run any individual model
python StockMarket_LinearRegression.py
python StockMarket_RandomForestClassifier.py
python StockMarket_LogisticRegression.py

# Compare all results visually
python comparizer.py

# Re-build the dataset from scratch (requires news headlines CSV)
python textb.py          # Download stock data
python merge.py          # Merge with sentiment
python remover.py        # Clean zeros

All model scripts expect Stocks_TCS2.csv to be present. The pre-built dataset is already included in the repo.

---

📊 Model Comparison

Run comparizer.py to generate side-by-side bar charts for both regression (RMSE) and classification (accuracy):

Regression:
┌─────────────────────────────────────┐
│  KNN   │ RF    │ Lin Reg            │
│ 318.15 │ 227.7 │ 49.81 ← Best      │
└─────────────────────────────────────┘

Classification:
┌─────────────────────────────────────┐
│  KNN   │ Log Reg │ Random Forest    │
│ 70.98% │ 75.08%  │ 75.68% ← Best   │
└─────────────────────────────────────┘

---

🔍 Key Findings

1. Linear Regression dominates regression tasks With an R² of 0.985, linear regression is surprisingly the best price predictor. Stock prices on adjacent days are highly correlated — which is exactly what LR captures well. Lag features actually hurt it slightly, suggesting multicollinearity.

2. Lag features are critical for tree-based and KNN classifiers Adding just two lag values and a rolling mean pushed KNN and Random Forest from ~50–55% (barely better than random) to ~70–75%. This is the single biggest lever in the project.

3. FinBERT sentiment adds signal Incorporating news sentiment into the feature set provides additional predictive information beyond price history alone, particularly on high-news days.

4. SVM underperformed The SVM (with RBF kernel, C=100, gamma=0.1) performed worse than random on this dataset — noted in the script itself as "worse than deciding on coin toss." Kernel and hyperparameter tuning would be needed for improvement.

5. The 75% ceiling Both Logistic Regression and Random Forest plateau around 75% accuracy for direction prediction. Breaching this ceiling likely requires deeper features — options data, order book depth, or LSTM-based sequence modeling.

---

🔮 Potential Improvements

• Add more technical indicators (RSI, MACD, Bollinger Bands)
• Try LSTM or Transformer-based sequence models
• Expand sentiment to include social media (Twitter/Reddit)
• Hyperparameter tuning via GridSearchCV
• Walk-forward validation instead of a single train/test split

---

_{Built with 🐍 Python · scikit-learn · FinBERT · yfinance}