Research Leadership and High Standards in Economic Forecasting: Neural Network Models Compared with Etalon ARIMA Models

Maintaining high standards in socio-economic research and achieving leadership positions in scientific circles requires a scientist to have a perfect command of mathematical tools developing accurate forecasts. Traditional forecasting methods typically involve fitting data to a pre-established relationship between dependent and independent variables, often making specific assumptions about a stochastic process. In contrast, machine learning presents an alternative approach to statistical analysis and forecasting, emphasising a data-driven methodology that does not assume any predefined statistical relationships in the data. The ARIMA method is recognised as the most effective forecasting method in the social sciences and is widely used for analyzing time series data. At the same time, deep learning methods (neural network models) are now forming a serious alternative to traditional econometric models, including ARIMA, as they can exploit nonlinear patterns in the data, which are often hidden from standard linear models. The aim of this study is to evaluate the performance of different neural network (NN) models compared to ARIMA models for economic forecasting. Five different network architectures are studied: Multilayer Perceptron (MLP), Simple Recurrent Neural Network (Simple RNN), Long Short-Term Memory Network (LSTM), Bidirectional LSTM (BILSTM), Convolutional Neural Networks (CNN). The performance of these models is compared with two benchmark ARIMA models: ARIMA based on AIC criteria (ARIMA-AIC) and ARIMA based on BIC criteria (ARIMA-BIC). The performance of these models is compared with two benchmark ARIMA models: ARIMA based on AIC criteria (ARIMA-AIC) and ARIMA based on BIC criteria (ARIMA-BIC). Prior training the NN models, hyperparameters were fine-tuned to obtain optimal performance for each model. The models’ performance on out-of-sample forecasting (test dataset) was evaluated using two metrics: RMSE (root mean square error) and MAE (mean absolute error). The data used to test the effectiveness of these models were monthly year-on-year consumer price inflation data in Algeria. The results show that the MLP model outperformed other models, including the benchmark, in the short and medium term (6-12 months). At the outperformed same time, the LSTM model outperformed all other models, including the benchmarks, in the long term (18-24 month forecasting horizon). Although Simple RNN models performed well in short-term forecasting, their performance deteriorated with the increase in the forecasting horizon. For the benchmark models, the forecasting results were the worst among all models, even in the short term. As a result, MLP and LSTM models were found to be the most appropriate for forecasting Algerian inflation, and deep learning is a promising alternative to traditional time series forecasting methods.