Data after collection, but prior to pre-processing in R code

This data was used for examining various machine learning methods for predicting S&P 500 returns using monthly data from November 1987 to February 2022, utilizing 11 predictors after principal component analysis. The research addresses the literature gap of applying state-of-the-art machine learning algorithms, specifically Diverse Representation Canonical Interval Forest (DrCIF), ARSENAL, and HIVE-COTE 2.0 (HC2), which have demonstrated competitive predictability in various domains but have not been explored for financial forecasting. Additionally, this study includes long-term geometric cumulative total returns in the evaluation of machine learning strategies, filling another gap in the literature. The strategies were evaluated based on out-of-sample test simulation in terms of predictability, risk-adjusted excess returns, and terminal portfolio value. The results indicated that DrCIF and Long Short-Term Memory (LSTM) provide greatest added economic value compared to passive index-following strategies, traditional logistic regression, and other common machine learning algorithms. The practical implication is that industry practitioners can increase their cumulative and risk-adjusted excess returns relative to the S&P 500 by implementing defensive market timing adjustments of their equity weighting, guided by these predictive machine learning models.