📈 economics

A Predictive Model to Detect Property Taxpayer Non-Compliance and Estimate Property Tax Liabilities in Rwanda Using Supervised Machine Learning

This study demonstrates that an XGBoost-based supervised machine learning framework, trained on over one million Rwanda Revenue Authority records, effectively detects property tax non-compliance and estimates liabilities, enabling a risk-based audit strategy that could recover 80% of unpaid taxes by targeting the top 20% of high-risk taxpayers.

Original authors: Alexis Nshimiyimana, Charles Ruranga

Published 2026-09-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alexis Nshimiyimana, Charles Ruranga

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In many parts of the world, collecting taxes on buildings and land is a difficult task. Governments need this money to build roads, schools, and hospitals, but the system often relies on people telling the truth about how much their property is worth. When owners declare a value that is too low, the government loses revenue, and honest taxpayers end up paying more than their fair share. To fix this, tax officials have traditionally relied on manual checks, where auditors physically visit properties or review paper files one by one. However, as cities grow and the number of buildings increases, this old way of working becomes too slow and expensive to keep up. In recent years, experts have begun to explore whether computers can learn from past records to spot patterns that humans might miss, helping officials decide who to check first without needing to look at every single file.

This is the challenge researchers from the University of Rwanda set out to solve. They wanted to know if a computer could look at the vast records held by the Rwanda Revenue Authority and figure out two things: first, which property owners are likely not paying the correct amount of tax, and second, what the correct amount should be. The team gathered a massive dataset containing information on more than one million property owners. This data included details about the buildings, how long the owners had been registered, how much they owed, what they had actually paid, and any penalties or interest they had accumulated. Instead of guessing, the researchers taught three different types of computer programs to learn from this history. They tested how well these programs could predict who was non-compliant and how accurately they could estimate the true tax bill for a property.

The researchers found that one specific type of computer program, known as XGBoost, was far better at the job than the others. This program is designed to learn by building many small decision steps and combining them to make a final prediction. When tested on data from 2024 that the computer had never seen before, this model correctly identified non-compliant taxpayers about 88 percent of the time. It was also very good at estimating the correct tax liability, with its predictions being much closer to the actual numbers than those made by the other models. The study revealed that the most important clues for spotting trouble were not always about the building itself, such as its size or location. Instead, the computer learned that a taxpayer's own behavior was the strongest signal. Factors like the ratio of what they paid versus what they owed, how much they owed in penalties, and how long they had been registered were the best indicators of whether someone was likely to be non-compliant.

The real power of this discovery lies in how it could change the way tax audits are conducted. The researchers ran a simulation to see what would happen if officials used these computer predictions to decide who to investigate. In a traditional approach, if officials picked taxpayers at random to audit, they would only recover about 20 percent of the unpaid money they were looking for. However, when they used the computer to rank taxpayers by risk and focused their efforts on the top 20 percent of the most likely offenders, the simulation showed they could recover roughly 80 percent of the unpaid liabilities. This suggests that by using data to guide their work, tax authorities could find the vast majority of missing revenue while spending far less time and money on audits.

The study concludes that this approach offers a practical path forward for Rwanda and other countries facing similar challenges. By integrating these predictive tools into their daily work, tax administrators could move away from random checks and toward a system where resources are directed exactly where they are needed most. The researchers emphasize that while the computer model is powerful, it should be used as a tool to support human decision-making, not to replace it entirely. They recommend that the Rwanda Revenue Authority adopt this risk-based system, while also continuing to improve the quality of their data and ensuring that the process remains fair. This work demonstrates that even in a resource-constrained environment, smart use of existing records can lead to a more efficient and equitable tax system.

Technical Summary: Predictive Modeling for Property Tax Compliance and Liability in Rwanda

Problem Statement
Property tax remains a significantly under-exploited revenue source in Rwanda and other developing economies, hindered by administrative complexity, incomplete property registers, and inaccurate self-declared valuations. In Rwanda, self-declared building values are often substantially lower than modeled values (e.g., a 42% gap identified in Kigali), and collection rates for penalties and interest lag far behind principal tax collections. Traditional compliance detection methods, relying on manual audits and rigid rule-based flags, are resource-intensive and fail to scale with the country's rapid urbanization. There is a specific gap in the literature regarding the integration of non-compliance detection and liability estimation for property tax within a single framework using nationally representative administrative data.

Methodology
This study employs a quantitative, predictive research design using secondary administrative data from the Rwanda Revenue Authority (RRA).

Dataset: The analysis utilizes 1,048,590 taxpayer records across 27 variables, covering property characteristics, assessed values, reported liabilities, payments, and compliance status.
Pre-processing: Missing values were imputed (mean/median for numerical, mode for categorical), outliers were handled via Z-scores and interquartile ranges, and categorical features were encoded. Synthetic Minority Oversampling Technique (SMOTE) was applied to address class imbalance in the compliance target variable.
Modeling Tasks: Two supervised learning tasks were defined:
1. Classification: Detecting taxpayer non-compliance (compliant vs. non-compliant).
2. Regression: Estimating property tax liabilities.
Algorithms: Three supervised algorithms were trained and compared: Decision Tree, Random Forest, and XGBoost.
Validation Strategy: A temporal validation approach was used to simulate real-world deployment. Data from 2020–2023 served as the training set (with stratified five-fold cross-validation for hyperparameter tuning), while 2024 records were held out as a completely unseen test set.
Evaluation Metrics: Classification performance was measured using Accuracy, Precision, Recall, F1-score, and AUC. Regression performance was assessed via Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination ( $R^2$ ).

Key Results
The study found that XGBoost consistently outperformed both Random Forest and Decision Tree models across all metrics for both tasks.

Non-Compliance Detection (Classification):
- XGBoost achieved an Accuracy of 0.88, Precision of 0.85, Recall of 0.83, F1-score of 0.84, and an AUC of 0.91.
- These results surpassed Random Forest (AUC 0.86) and Decision Tree (AUC 0.71).
Liability Estimation (Regression):
- XGBoost demonstrated the strongest fit with an $R^2$ of 0.87, the lowest RMSE (78.9), and the lowest MAE (58.1).
Feature Importance:
- The most influential predictors of non-compliance were payment ratio, penalty dues, interest payable, and registration age. Notably, dynamic payment history and arrears data were more predictive than static property attributes like land value.
Operational Simulation:
- An audit-targeting simulation revealed that focusing audits on the top 20% of taxpayers ranked by XGBoost-predicted risk could recover approximately 80% of unpaid liabilities. In contrast, random audit selection would recover only about 20%.

Key Contributions

Integrated Framework: The study provides a unified supervised machine learning framework that simultaneously detects property taxpayer non-compliance and estimates tax liabilities, addressing a gap where previous Rwandan studies treated valuation and compliance separately.
Nationally Representative Data: Unlike prior work focused on specific cities (e.g., Kigali) or specific tax types (e.g., VAT), this study utilizes nationally representative administrative data for property tax.
Operational Translation: The research translates predictive risk scores into a practical audit-targeting simulation, demonstrating the efficiency gains of risk-based selection over random selection for the Rwanda Revenue Authority.

Significance and Claims
The paper claims that supervised machine learning, specifically XGBoost, can substantially improve risk-based compliance monitoring and liability estimation in a resource-constrained tax administration. The authors conclude that these models can be integrated into the RRA's compliance risk management framework to prioritize audit resources effectively.

However, the authors maintain a modest stance regarding operational deployment. They acknowledge that:

The XGBoost model is relatively opaque compared to single decision trees, necessitating complementary interpretability tools (e.g., SHAP) and explicit fairness auditing before deployment.
Compliance labels are derived from payment records rather than audited ground truth, meaning they may contain noise.
The current data lacks geospatial and socio-economic covariates, suggesting future work should incorporate richer features and validate model-selected audits against conventional field practices.

The study recommends that the RRA invest in data quality and staff capacity to sustain such systems but cautions that the current results serve as a proof-of-concept for integrating predictive analytics into existing administrative workflows.

Technical Summary: Predictive Modeling for Property Tax Compliance and Liability in Rwanda

More like this