Comparison of machine learning models for market predictions with different time horizons

Abstract: Stock market prediction is, when successful, a means of generating large amount of wealth. It remains an unanswered question if stock prices can be predicted consistently, due to the randomness and seemingly unpredictable nature of stock prices. A previous study has shown that using machine learning they were able to predict with high accuracy whether stocks on the Indian stock exchange BSE Sensex would increase or decrease in price the following trading day. They compared several machine learning models which were all trained using technical indicators from the stocks they researched. They also introduced a method of converting continuous valued technical indicators into discrete values and found it to improve the accuracy of the models. In this thesis we apply a similar method, using technical indicators converted into discrete values, to model and compare different models on stocks listed on American stock exchanges. We also compare models when predicting stock price movements ten and 30 days into the future, instead of just one. Our results do not show the same high accuracies seen in the previous study. Overall, we found that the simple strategy of always predicting that a stock’s price will increase the following trading day provided higher accuracy than using a model. Our results are therefore in support of the random walk theory, which states that stock prices evolve randomly and therefore cannot be predicted.