Using Machine Learning to Create a Property Tax Roll: Evidence from the City of Kananga, D.R. Congo

Governments in the world’s poorest countries face severe revenue constraints. They typically collect less than 10 per cent of GDP in taxes, compared to 25–50 per cent in high-income countries. The literature on state capacity and development argues that inability to collect taxes is at the heart of why low-income countries are as poor as they are. It suggests that the path to economic prosperity may begin with investment in governments’ capacity to collect the tax revenue necessary to provide public goods that enhance productivity.

Property taxation is often the primary source of government revenue at the local level, and is essential for provision of local public goods.1 However, it remains one of the most under-utilised taxes in developing countries. This is partly because taxing properties requires mapping and assessing the value of properties, which is complex and expensive. Only 39 per cent of non-OECD countries and 15 per cent of sub-Saharan African nations have mapped their largest city’s private plots.

Several approaches have been proposed to map and value properties (see Zebong, Fish and Prichard (2017) for a review). Some countries rely on in-person appraisal visits, but, while accurate, these are typically costly and prone to corruption. For this reason, many countries, such as Pakistan, Sierra Leone, and Malawi, have instead adopted simplified valuation methods. The most common approach is points-based valuation, which consists of assigning points based on the surface area of the land and buildings. Additional points are awarded for positive features, and deducted for negative features.