Audit lead selection and yield prediction from historical tax data using artificial neural networks

Area of Science:

• Data Science
• Machine Learning
• Taxation

Background:

• Traditional tax audit lead selection relies on empirical rules and randomization, which can be inefficient and biased.
• Existing methods may miss significant non-compliance cases, impacting fairness and revenue.

Purpose of the Study:

• To develop and evaluate an artificial neural network-based tool for improved tax audit lead selection.
• To enhance the fairness and effectiveness of tax audit processes through data-driven insights.

Main Methods:

• An artificial neural network model was trained on a dataset of 93,413 restaurant business tax records from Northern California (2007-2017).
• The model was evaluated for its precision, recall, and F1-score in identifying positive audit leads.
• A regressor was used to estimate potential audit gains, and empirical rules were statistically validated.

Main Results:

• The artificial neural network tool achieved 40.1% precision and 58.7% recall (F1-score 0.42) for classifying audit leads.
• The model's regressor estimated audit gains with a Mean Absolute Error (MAE) of $155,490.
• Two out of five evaluated empirical rules were statistically supported by the data.

Conclusions:

• Leveraging data through evidence-based models can significantly improve audit selection processes.
• This approach leads to higher audit yields and promotes fairer selection of businesses for tax audits.
• The study demonstrates the successful application of machine learning in optimizing tax compliance and administration.