Forecasting the Nasdaq-100 index using GRU and ARIMA

Abstract: Today, there is an overwhelming amount of data that is being collected when it comes to financial markets. For forecasting stock indexes, many models rely only on historical values of the index itself. One such model is the ARIMA model. Over the last decades, machine learning models have challenged the classical time series models, such as ARIMA. The purpose of this thesis is to study the ability to make predictions based solely on the historical values of an index, by using a certain subset of machine learning models: a neural network in the form of a Gated Recurrent Unit (GRU). The GRU model’s ability to predict a financial market is compared to the ability of a simple ARIMA model. The financial market that was chosen to make the comparison was the American stock index Nasdaq-100, i.e., an index of the 100 largest non-financial companies on NASDAQ. Our results indicate that GRU is unable to outperform ARIMA in predicting the Nasdaq-100 index. For the evaluation, multiple GRU models with various combinations of different hyperparameters were created. The accuracies of these models were then compared to the accuracy of an ARIMA model by applying a conventional forecast accuracy test, which showed that there were significant differences in the accuracy of the models, in favor of ARIMA.