Predicting Household Expenditure Using Machine Learning Techniques: A Case of Cambodia

This study aimed to predict household expenditure using a combination of survey and geospatial data. A web-based application operating on the Google Earth Engine platform has been specifically developed for this research, providing a set of satellite-based indicators. These data were spatially averaged at the district level and integrated with household nonfood expenditures, a proxy of socioeconomic conditions, derived from the World Bank’s 2019 Living Standards Measurement Study (LSMS). Four machine learning algorithms were applied. By using root mean square error as the goodness-of-fit criterion, a random forest algorithm yielded the highest forecasting precision, followed by support vector machine, neural network, and generalized least squares. In addition, variable importance and minimal depth analyses were conducted, indicating that the geospatial indicators have moderate contributive powers in predicting socioeconomic conditions. Conversely, the predictive powers of variables derived from the LSMS were mixed. Some asset ownership yielded a high explanatory power, whereas some were minimal. The attained results suggest future development aimed at enhancing accuracy. Additionally, the findings revealed an association between economic activity density and household expenditure, recommending regional development promotion through urbanization and transition from agriculture to other economic sectors.