Forecasting stock returns via valuation models and macro indicators

Abstract

In a world that is constantly jostled by political and macroeconomic shocks, and simultaneously with AI rise, especially machine learning methods, I felt compelled to strip away the noise and focus on the long-term game. For this purpose, my aim was to see what factors affect the S&P 500 and to test algorithms that could make practical use of those factors for a thorough fundamental analysis. The variables used were of quantitative nature, specifically the valuation models built from com-pany-level financial ratios and a set of macroeconomic indicators, plus a set of machine learning tools – CNN, LSTM and XGBoost that often are applied in stock analysis. However, due to the diminished dataset, I turned to the multiple linear regression that accurately identified the stock price predictors and was also used to back test against the S&P500. The results showed that not every variable mattered equally. For example, the Graham number was a poor predictor for Apple, hinting that this metric may falter with tech stocks. The final recommendations point to the use of large, reliable datasets, which in turn demands solid APIs. Therefore, future research should in-clude that but also qualitative data to better derive insights for long-term and real-time investments. Finally, I challenge all the researchers to build a user-friendly platform that would allow investors and analysts to plug in these metrics and get forecasts across different time horizons.