Searching Stable Solutions For Stock Predictions: A Stacking Approach

The goal of the competition is to predict stock positions for holding, selling or buying stocks of companies from the S&P 500. Firstly the data is read in and the missing values are imputed with the median. Categorical data is one-hot encoded. A classification approach with mainly tree based methods is used. The models used are HistGradientBoosting, XGBoost, MLP and SVC whose parameters are chosen and modified through a grid search. For the stacking the models' prediction results are summed up and the result is mapped to the three positions. It is found that the result is a bit overfitted to the competition's test data which makes sense in regard to it being a competition. The stacking improves the score drastically. Concluding it can be said that machine learning models can hint in the right direction when it comes to handling stocks but fail at giving good financial advice.

---

Track: Data Mining Competition