Leveraging Machine Learning for Countering Customs Tax Evasion in Zambia

Policy Context

Zambia relies heavily on customs taxes to fund critical public services, processing around half a million commercial shipments annually. However, the nation faces a severe tax gap currently estimated between 47% and 56%, largely driven by systemic customs tax evasion and revenue leakage. Customs administrations face severe manpower constraints, and the existing digital clearance system generates an unmanageably high volume of shipments requiring physical inspection. Because officers cannot possibly inspect every import, optimizing resource allocation through risk-based targeting is crucial to catch sophisticated tax evaders while speeding up legitimate trade. We do this by using machine learning to analyze historical customs data and assigning real-time risk scores to incoming shipments.

Study Design 

To address this challenge, researchers partnered with the Zambia Revenue Authority (ZRA) Artificial Intelligence unit to adapt a Random Forest machine learning algorithm, previously piloted in Paraguay, to detect customs fraud in Zambia. The team first conducted a qualitative scoping study, performing workflow observations and semi-structured interviews at four distinct, functionally major customs ports to understand daily operations and evasion typologies. They also systematically queried the ZRA’s existing customs SQL database to evaluate the structural availability, quality, and completeness of the historical data needed to train the predictive model. Ultimately, the project plans a randomized pilot evaluation to rigorously measure the algorithm’s true impact on state revenue collection and its ability to reduce the overall volume of physical inspections.

Results and Policy Lessons

Results forthcoming.