Work

Forecasting agricultural value added under drought risk in Morocco: a hybrid Fourier ARDL–neural network approach

Year: 2026
Type: preprint
Abstract: Droughts have become longer and more frequent in Morocco, and their effect on agricultural output is now a major source of uncertainty in macroeconomic forecasts. This paper asks how well agricultural value added can be forecast once crop-year rainfall is known, and what droughts of different intensity and duration would cost. We build a consistent quarterly series of agricultural value added for 1998–2025, benchmarked to the latest national accounts, and link it to rainfall and temperature from the ERA5 reanalysis. We propose a hybrid model in which a Fourier ARDL captures the linear relation and gradual structural change, and a small neural network, selected on a separate validation period, models the nonlinear structure left in its residuals. A 10 percent fall in crop-year rainfall lowers agricultural value added by about 4.4 percent in the long run. Over 2020–2025, the Fourier terms improve four-quarter-ahead forecasts by about 30 percent relative to a standard ARDL, and the hybrid model is the most accurate of six models, with a mean absolute percentage error of 5.4 percent against 11.1 percent for ARIMA, although the network’s own contribution is modest. Scenario simulations show that a two-year drought costs more than a single severe one, about 55 against 41 billion dirhams. JEL Classification: C22 , C45 , C53 , Q10 , Q54
Language: English
Open Access status: green
OpenAlex ID: W7235794592
Cites: 1
Cited by:
Related to: 10