Association for Public Policy Analysis & Management

Traditional poverty measures, which are estimated using income, do not adequately reflect the true incidence of poverty. Meanwhile, the funding limits faced many local governments require more precise poverty targeting. In this paper, we propose a new measure of poverty, Composite Financial Status (CFS) to gain a more complete picture of households’ financial situations, and we capitalize on machine learning methods to improve poverty prediction. Using data from the China Family Panel Studies (CFPS), a nationally representative longitudinal survey of Chinese households that contains asset information, we combine households’ disposable income and net worth to create the CFS measure. CFS takes into account the quintile distributions of families’ income and asset, and disaggregates household financial status into four categories: Poor (both income and asset poor), Transient Poor (income poor and asset non-poor), Transient Stable (income non-poor and asset poor), and Stable (both income and asset non-poor), as defined by us.

With these CFS groups, we studied the mobility patterns among different groups during our observation period from 2010 to 2014. To investigate the socioeconomic factors underlying the intra-group mobility, we used a multinomial logistic regression model with year fixed effects. Moreover, applying discriminant analysis and Neural Network—a machine learning method—we explored different modeling specifications that produce the most accurate household-level predictions of future poverty.

Overall, we find evidence that CFS illustrates distinctive poverty patterns not shown by traditional poverty measures. Compared with the stable group, families in the transient poor group, though with their asset above the asset-poverty line, are indiscriminant in their characteristics as the poor group. Whereas families who are transient stable—having above-threshold income but insufficient asset—are relatively well off in terms of health and employment. There is high mobility among different CFS groups of around 41 percent, and nearly 38 percent of the poor move into stable group during the four-year period, an indication of effective poverty reduction. Testing the accuracy of different forecasting models based on cross-validation, our neural network models achieve a correct classification rate of 73 percent with minimal variables and up to 97 percent when enough information is provided. We conclude that CFS appears to be an appealing alternative measure of household poverty and presents opportunities for governments to better target financial resources by using advanced forecasting techniques.