Incorporating High-Frequency Weather Data into Consumption Expenditure Predictions
This work addresses the challenge of accurately mapping short-term well-being fluctuations in data-sparse regions, though it is incremental by building on existing methods for welfare prediction.
The paper tackled the problem of predicting volatile welfare measures like consumption expenditure by incorporating high-frequency weather data, resulting in a significant improvement in prediction accuracy compared to models using only satellite imagery.
Recent efforts have been very successful in accurately mapping welfare in datasparse regions of the world using satellite imagery and other non-traditional data sources. However, the literature to date has focused on predicting a particular class of welfare measures, asset indices, which are relatively insensitive to short term fluctuations in well-being. We suggest that predicting more volatile welfare measures, such as consumption expenditure, substantially benefits from the incorporation of data sources with high temporal resolution. By incorporating daily weather data into training and prediction, we improve consumption prediction accuracy significantly compared to models that only utilize satellite imagery.