Unlocking Stock Return Predictions: Using Financial Statements with Random Forest and PCA

Financial statements are pivotal for forecasting the future performance of stocks. Harnessing the random forest machine learning model, this study aims to enhance the prediction of quarterly stock returns by focusing on twelve critical financial indicators. This paper utilized Principal Component Analysis (PCA) for dimensionality reduction and feature selection, aiming to optimize the model's predictive accuracy. The dataset encompassed quarterly financial statements and stock data for the 100 constituent stocks of the NASDAQ 100 index from 2010 to 2020. The PCA analysis revealed that reducing the input features to six dimensions significantly improved the model's predictive performance, as indicated by Mean Squared Error (MSE) and Mean Absolute Error (MAE). This finding suggests that an overabundance of components can introduce unnecessary complexity, potentially detracting from the model's predictive capabilities. The feature importance assessment, conducted using the random forest algorithm, identified Volatility, Revenue Growth Rate, and Return as the most influential predictors. Notably, the optimal predictive performance was achieved with the inclusion of seven and five top features, respectively, highlighting the non-linear relationship between the number of features and model performance. This comprehensive study underscores the utility of the random forest model in predicting stock returns and emphasizes the critical role of dimensionality reduction and feature selection refinement in enhancing predictive accuracy.