Using domestic weather disturbances and price transmission for maize price predictions in Southern Africa

In this paper we combine traditional econometric time series techniques and machine learning algorithms to construct skillful monthly maize price prediction models for four southern African countries – namely, Malawi, Mozambique, Zambia, and Zimbabwe. Theoretical models of price transmission commonly assume that shocks are transmitted from an external market (typically modeled as the world market) to the largest domestic city or port within a country and then, depending on the degree of market integration within the country, these shocks are transmitted to local markets. However recent evidence suggests that internal shocks have a larger impact on prices than external shocks. In an analysis of 554 local commodity markets across 51 countries during the period between 2008-2012, Brown and Kshirsagar (2015) find that 20% of local market prices were affected by domestic weather disturbances in the short-run in comparison to 9% by international price changes. This finding has prompted more recent literature to relax assumptions about international price transmission to investigate how shocks are transmitted through local and regional markets. 

Here we investigate the effects of domestic weather disturbances on regional maize price transmission. We then use these results of to build skillful price prediction models that use limited price data, weather disturbances, and other readily accessible free secondary data to predict monthly grain prices three, six, and nine months ahead in four Southern African countries. The collection of subnational price data in developing countries is costly and often difficult to obtain. We limit the amount of price data used by first determining if monthly price series in each country co-move and how these co-movements are influenced by domestic climate disturbances. We then use bivariate error correction models to both assess whether price movements in each country follow well-defined paths and identify influencing and influenced markets.

From this analysis we classify markets that act as price anchors in each country. Because local climate conditions have been found to affect and accurately predict agricultural prices, price dispersion, and yields in developing countries we use climate conditions at both the market location and anchor market locations as predictors. We show that during periods classified by drought, price prediction models using anchor market prices and high-resolution climate data have high degrees of predictive accuracy. We hope the results presented in this paper will assist policymakers, government stakeholders, and researchers in systematically constructing subnational price forecasts with minimal price data to be used in early warning and food security monitoring models.