Analyse de la série de la production algérienne de blé et son application à la comparaison des performances d’un réseau de neurones récurrent LSTM vs ARIMA

Analysis of the Algerian wheat production series and its application to the performance comparison of recurrent neural network LSTM vs. ARIMA

الملخص

This article aims to analyze the characteristics of the national wheat production time series from 1960 to 2022, and compare the predictive performance of a recurrent neural network LSTM with the ARIMA model on this series. The KPSS test shows that this series is non-stationary, and a break detection test identified a trend break in 2002, highlighting a growth momentum in production. The unfavorable climatic factors cannot explain this point break; it is much more the agricultural policy of support of the State, initiated at the end of the 90s, which was the cause. The 1960-2002 sub-period was used to compare the predictive performance of a sub-optimal LSTM with the optimal ARIMA model. The respective (lower) values 451.9956 and 0.4254 of the RMSE and the MAPE of the LSTM on the test set, compared to those of the ARIMA with 769.5197 and 0.4631 show the superiority of the LSTM. The usefulness of such a result is that it makes it possible to take the LSTM as a reference model for wheat production forecasts