ARIMA vs. MACHINE LEARNING IN TERMS OF EQUITY MARKET FORECASTING
Iulian-Cornel LOLEA1, Ioan-Radu PETRARIU2, Adriana GIURGIU3
1Bucharest University of Economic Studies, Bucharest, Romania
2Bucharest University of Economic Studies, Bucharest, Romania
2Bucharest University of Economic Studies, Bucharest, Romania; Department of International Business, Faculty of Economic Studies, University of Oradea, Oradea, Romania
Abstract: Through this paper we aimed to develop a comparison between ARIMA, Prophet, KNN and Neural Networks in terms of stock prices forecasting. After reviewing the literature, we noticed that there is a plethora of studies that address this problem of forecasting, but very few have made comparisons that include ARIMA, machine learning, but also the Prophet forecasting model developed by Facebook, which brought interesting results for certain data series. Based on methodologies validated by other authors, we compared these models in our paper and we sought to obtain promising results regarding performance evaluation. The comparison was made in-sample, the training period being 01/01/2010 – 31/07/2021, but also out-of-sample (01/08/2021 – 31/10/2021). The study was performed for Societe Generale’s stock, using daily observations. Statistical loss functions such as RMSE, MPE, MAPE, MAE, and ME were used for comparison. The results indicated an outperformance of Neural Networks, both in-sample and out-of-sample, this model being on the 1st place according to the aggregated score. It is also noteworthy that the ARIMA model was in second place in-sample, ahead of KNN, but for out of sample these two algorithms changed their positions. On the other hand, the Prophet algorithm performed the weakest, both in-sample and out of sample. Also, we must underlie that all four algorithms had a clear tendency to overestimate the price of Societe Generale, according to the results of the statistical loss functions ME and MPE. Finally, it should be noted that the results were consistent with what other authors found out, especially for the out-of-sample period, where the machine learning models performed best.
Keywords: loss functions; machine learning; autoregressive; equity markets.
