Forecasting financial budget time series: ARIMA random walk vs LSTM neural network

Maryem Rhanoui (IMS Team, ADMIR Laboratory, Rabat IT Center, ENSIAS, Mohammed V University in Rabat)
Siham Yousfi (School of Information Sciences)
Mounia Mikram (Mohammed V University)
Hajar Merizak (School of Information Sciences)

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol 8, No 4: December 2019

Financial time series are volatile, non-stationary and non-linear data that are affected by external economic factors. There is several performant predictive approaches such as univariate ARIMA model and more recently Recurrent Neural Network. The accurate forecasting of budget data is a strategic and challenging task for an optimal management of resources, it requires the use of the most accurate model. We propose a predictive approach that uses and compares the Machine Learning ARIMA model and Deep Learning Recurrent LSTM model. The application and the comparative analysis show that the LSTM model outperforms the ARIMA model, mainly thanks to the LSTMs ability to learn non-linear relationship from data.