Forecasting Potato Production in Major South Asian Countries: a Comparative Study of Machine Learning and Time Series Models

Abstract

This study analyzed and forecasted potato production in eight major South Asian countries from 1961 to 2028 using advanced time series and machine learning approaches. Annual potato production data was modelled with autoregressive integrated moving average (ARIMA), state space, and extreme gradient boosting (XGBoost) models. The models were trained on 1961–2009 data and evaluated on a 2010–2021 validation set. On the training set, XGBoost showed the best performance. However, on the validation set, ARIMA and state space models significantly outperformed XGBoost, indicating issues with overfitting. The ARIMA models produced the lowest forecast errors for Afghanistan, Bangladesh, China, and Myanmar. Meanwhile, state space models were optimal for India, Nepal, Pakistan, and Sri Lanka, demonstrating that no one approach was uniformly best. The top performing models forecast potato production up to 2028, Afghanistan’s production is expected to remain stable at around 860–862 thousand metric tons. Bangladesh’s output is forecasted to stay constant at 9887 thousand metric tons. In contrast, China is predicted to see a steady increase from 94,625 to 96,193 thousand metric tons. India’s production is anticipated to grow significantly from 54,704 to 62,396 thousand metric tons. Conversely, Myanmar’s production is projected to decline from 460 to 426 thousand metric tons. Nepal’s output is expected to steadily increase from 3395 to 4011 thousand metric tons. Pakistan’s production is forecasted to rise substantially from 5573 to 7045 thousand metric tons. Lastly, Sri Lanka’s potato production is projected to experience a modest increase from 77 to 84 thousand metric tons. These forecasts reflect the different levels of potato demand, consumption, and trade in each country, as well as the effects of climate change, pests, and diseases on potato yields. The rigorous comparative application of advanced time series and machine learning techniques provides valuable data-driven insights into future South Asian potato supply. The forecasts can assist food security planning and agricultural policymaking in the region.