Modelling Wealth from Call Detail Records and Survey Data with Machine Learning: Evidence from Papua New Guinea

Call detail records (CDRs) provide a significant opportunity to understand human development at a high spatiotemporal resolution, specifically in developing countries, which face financial, human, and capacity constraints. This study attempts to model and identify features derived from CDRs that can best predict relative wealth and poverty across Papua New Guinea (PNG), by combining it with tele-survey data. Our findings show promising results on the prediction of dichotomous variables consisting of self-reported household assets with an Area Under the Curve (AUC) score is equal to 0.88 or higher. Meanwhile, the prediction of the numerical wealth index, which was built using a dimensional reduction method did not provide satisfactory results. For the target variable Principle Component Analysis (PCA) derived numerical wealth index, the Root Mean Squared Error (RMSE) 0.69 is lower compared to its standard deviation 0.74. The numerical index was further classified into quintiles, and this was also used as a response variable separately and was subjected to a multi-class classification approach. The best F1 score for multiclass-classification of the quintiles derived from PCA was 0.7. Findings from the Multiple Correspondence Analysis (MCA) derived index and quintiles add the robustness to our study. The overall results suggest that in the case of PNG, CDRs are better suited for developing proxy indicators related to individual household assets and classified quintiles based wealth index, but not for a numerical relative wealth index. hi general, our results add more confidence in harnessing mobile network data to predict wealth and poverty.