Predicting the Indian Stock Market with Machine Learning: A Complete Analysis of a Technical Indicator-Driven Price Prediction Project

The open-source project analyzed in this article was developed by Kartikey Kumar Tripathi. It uses linear regression and feature engineering with 14 technical indicators to predict the stock prices of three major Indian listed companies: Reliance Industries, TCS, and Infosys, covering the complete ML pipeline from data acquisition and feature construction to interactive dashboards.

Extracting value from the massive transaction data in the stock market is a core topic in fintech. This project was developed by a third-year computer science student in India, aiming to apply ML theory to real financial data. The research targets three core sector companies listed on India's NSE (in the energy, IT, and consulting fields).

The project converts classic technical indicators into ML features, totaling 14: trend indicators (MA_10/20/50), volatility indicators (Bollinger Bands, BB_Width), momentum indicator (RSI), lag features (Lag_1/3/5), and auxiliary features (daily return, price range, volatility, trading volume). The author points out that the high R² is partially due to the autocorrelation effect of Lag_1.

Three years of OHLCV data (about 750 trading days) were obtained using the yfinance library. Preprocessing steps to avoid data leakage: MinMaxScaler is only fitted on training data; time series are split chronologically into 80/20, without random shuffling.

Linear regression was chosen (due to its strong interpretability and low overfitting tendency). Evaluation metrics include RMSE, MAE, MAPE, and R². A 30-day rolling prediction was implemented to simulate continuous decision-making scenarios, revealing the limitations of linear models in long-term predictions.

The accompanying dark-themed HTML dashboard includes three modules: individual stock analysis (KPI cards + four-panel charts), comparative analysis (indicator comparison of three stocks), and methodology explanation (feature design/process/limitations). Data is dynamically populated via metrics.json, separating the data layer from the presentation layer.

Limitations: High R² due to Lag_1 autocorrelation; linear assumption fails to capture non-linearity; lack of external factors (news/financial reports/macroeconomic indicators). Improvements: Ensemble models (Random Forest/XGBoost), rolling window validation, Streamlit application, expanding the range of stocks.

Value for beginners: Converting domain knowledge into features, preventing data leakage, time series processing, and visualization productization. For quantitative finance learners: Complete pipeline and highly readable code. Reminder: For educational purposes only, not for investment decisions.