Improving the performance of micro-simulation models with machine learning: The case of Australian farms

Micro-simulation models are widely used to measure the effects on businesses or individuals of policy changes or other shocks, including the effects on farms of changes in weather conditions and prices. Typically, economic micro-simulation involves econometric analysis of microdata to estimate parametric models. In contrast with the existing literature, this paper presents a non-parametric machine learning based micro-simulation model. In this study, a multi-target regression tree algorithm is combined with farm and weather panel data, to produce an economic micro-simulation model of Australian farm businesses. This approach captures the complex non-linear and farm specific effects of weather and price shocks on profits, with out-of-sample tests showing performance gains over conventional methods. Model results demonstrate the sensitivity of Australian farm profits to weather risk, particularly drought, and show an increase in weather risk exposure over the last 20 years. [Display omitted] •An economic micro-simulation model of Australian farm businesses is developed.•The model applies a novel non-parametric machine learning approach.•Out-of-sample tests show performance gains over typical parametric methods.•Australian farm profits are highly sensitive to weather risk particularly drought.•Australian farms have been exposed to increased weather risk over the last 20 years.