Revisiting Feature Scaling in Linear Regression: An Empirical Study on Microsoft Stock Price Prediction

Abstract

Stock price prediction occupies a central position in quantitative finance, bearing directly on risk management, portfolio construction, and investment decision-making. This study evaluated the effect of feature scaling on linear regression performance in predicting Microsoft (MSFT) stock prices. A quantitative experimental design was employed, drawing on historical MSFT stock data spanning 2014 to 2024. Preprocessing involved data cleaning, outlier treatment via the Interquartile Range (IQR) method, and feature standardization through Z-score normalization. Two experimental conditions were tested: linear regression without feature scaling and linear regression with feature scaling. Model performance was assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²). Both conditions produced nearly identical results — R² approaching 0.99, with negligible divergence across all error metrics. The evidence suggests that feature scaling does not meaningfully alter the predictive behavior of linear regression. For simple linear models operating without regularization, scaling appears to be an unnecessary preprocessing step, a finding that warrants more deliberate evaluation of preprocessing decisions in machine learning pipelines.