Random forest and stock market data

Abstract

Since random forest relies on data being independent and identically distributed (IID), it has largely been ignored in time series literature. To the best of my knowledge, limited research has been conducted using random forest for multi-step financial time series forecasting. This thesis aims to assess the performance and practicality of employing random forest for multi-step forecasting on financial time series using the direct strategy and to compare the results to conventional statistical models. The direct strategy involves creating a unique model for each prediction horizon. The models are trained independently on the same training data, and the predictions from each model are combined to produce a multi-step forecast. The performance of random forest and traditional statistical models were evaluated on actual stock price. The results showed that random forest provided a more accurate prediction on three of the five stocks evaluated, based on the RMSE and MAPE scores. Due to the limited number of time series examined, no definitive conclusion can be drawn in favour of either method.In addition, it cannot be ruled out that the performance of random forest may be attributed to the direct strategy itself, but not necessarily the algorithm. It should be noted, however, that a person with no prior knowledge of market efficiency or the underlying data generation process may expect to get similar results to traditional time series models when using random forest combined with the direct strategy.