Designing a model for financial forecasting using the integration of neural networks, Box Jenkins and Holt Winters methodology

The present study designs a model for financial forecasting by integrating neural networks. This retrospective comparative study uses the average price data of OPEC oil from ۲۰۰۳ to ۲۰۲۲ to forecast the period from June ۲۰۲۲ to May ۲۰۲۴. To this end, two time-series models (Box-Jenkins and Holt-Winters) were examined, which in the second stage were incorporated into a hybrid model based on artificial neural networks. The neural network model was developed using Matlab, and the Box-Jenkins time-series model was constructed using SPSS and Eviews software. Based on the results of the error analysis of the Box-Jenkins methodology, among the time series processes ARIMA(۵,۱,۵), ARIMA(۴,۱,۵), ARIMA(۳,۱,۵), and ARIMA(۵,۱,۳), the models demonstrated the best accuracy with MSE values of ۶۱.۸۶, ۶۳.۲۱, ۶۳.۲۹, and ۶۳.۶۲, respectively. The accuracy of the Holt-Winters method was not suitable for time-series forecasting due to the nature of the data. Therefore, the best artificial neural network was designed for combining forecasting methods. This neural network included an input layer with ۵ neurons, a hidden layer with ۵ neurons, and a single-neuron output layer. The network was trained using the Levenberg-Marquardt algorithm and employed a linear sigmoid activation function. The results indicated that the designed hybrid neural network significantly improved the accuracy of the forecasting methods and enhanced the MSE, MAPE, AIC, and BIC indices.

کلیدواژه ها:

Prediction ، time series ، ARIMA ، Neural Network