Combining rather than selecting multi-horizon forecasting of an agricultural commodity price index

Wholesale food prices in emerging economies move sharply, with direct consequences for farm incomes and food security. Individual commodity series are short, gap-ridden and shock-dominated, so this study forecasts a composite instead: the Agricultural Commodity Price Index, an inverse-volatility weighted aggregate of 135 daily wholesale series from Nepal’s Kalimati market, 2013–2023, with 55 strictly causal predictors including Nepali festival windows. Twelve architectures spanning penalised linear models, tree ensembles, recurrent networks and direct neural forecasters are compared under a protocol that fits on 2013–2020, chooses every configuration on 2021 to mid-2022 by blocked cross-validation, and reads July 2022 to September 2023 only for reporting. A shrunk combination improves on the model that validation would have selected by 1.3, 1.5, 2.0 and 11.2 %, and on a no-change benchmark by 2.1, 3.2, 7.8 and 22.2 %. The models that proved best on the test window ranked third, seventh, fifth and fourth on validation, so none could have been chosen in advance. A supporting result explains why these models work at all: the recursive, level-target strategy common in this literature caps tree ensembles at their training range and compounds recurrent error, and a direct change target reduces error by 33-53 %.