Comparative Analysis of Multiple Linear Regression with L1 and L2 Regularization for Stock Price Prediction

Abstract

This research explores the application of L1 (Lasso) and L2 (Ridge) regularization techniques within multiple linear regression frameworks for predicting stock prices in the ACLBSL segment of the Nepal Stock Exchange (NEPSE). A key challenge in stock price prediction over fitting is addressed by incorporating regularization methods that penalize model complexity. Through hyper parameter tuning, optimal alpha values of 0.9541 for Lasso and 0.4715 for Ridge were identified. These values led to improved model performance, reducing Mean Squared Error (MSE) to 514.12 and 521.02, respectively.

The study's findings reveal that Lasso regression not only enhances prediction accuracy but also performs effective feature selection by shrinking less significant coefficients to zero. This enables a more interpretable and simplified model without sacrificing performance. In contrast, Ridge regression retains all features with reduced coefficient magnitudes. The results indicate that Lasso regression is more effective in identifying and leveraging key predictors, thereby providing better generalization to unseen stock price data.

This research contributes to the ongoing efforts in financial modeling by demonstrating that regularization techniques can substantially improve the robustness and reliability of predictive models in the context of NEPSE, providing valuable insights for investors and analysts.