Integrating ARIMA-LSTM Architecture for Forecasting Avocado Production in Major Producing Countries
Аннотация
Avocado ( Persea americana Mill.) has emerged as one of the most rapidly expanding horticultural commodities in global agricultural trade, with production trajectories exhibiting complex non-linear dynamics driven by shifting consumer demand, climatic variability, and agronomic intensification. Accurate forecasting of avocado production is therefore critical for informed policy decisions, supply chain management, and agricultural planning. The present study develops and evaluates a hybrid Autoregressive Integrated Moving Average (ARIMA)–Long Short-Term Memory (LSTM) modelling framework for forecasting annual avocado production across five major producing countries: Colombia, the Dominican Republic, Indonesia, Mexico, and Peru. The modelling strategy follows a three-stage approach: (i) a classical ARIMA model is first fitted to capture the linear structure of each series; (ii) an LSTM deep learning model is subsequently trained individually and on the ARIMA residuals to extract residual non-linear dynamics; and (iii) the ARIMA forecast and the LSTM-predicted residual are combined to yield the final hybrid forecast. Prior to model development, series stationarity was assessed via the Augmented Dickey–Fuller (ADF) test, and optimal ARIMA orders were selected using the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC). LSTM hyperparameters including lag order and the number of hidden units were optimised through cross-validated Mean Absolute Percentage Error (MAPE) on the training set. The selected ARIMA specifications (ranging from ARIMA(1,1,1) to ARIMA(2,2,1)) demonstrated adequate residual whiteness as confirmed by the Ljung–Box test ( p > 0.05) for all countries. The standalone LSTM models achieved test MAPE values ranging from 3.97% (Mexico) to 17.81% (Colombia), while the hybrid model produced refined forecasts by correcting systematic ARIMA residuals, with the most notable improvement observed for Indonesia and Mexico. The findings demonstrate that the hybrid ARIMA-LSTM framework offers a robust and flexible forecasting tool for time series characterised by mixed linear-non-linear dynamics, and provides actionable insights for stakeholders engaged in the global avocado supply chain.