Realistic Tax Planning with Evolutionary Algorithms

Tax avoidance and non-compliance is a major challenge for governments around the world, and in the United States is estimated to cost the federal government up to 15 % of its annual revenue. Various computational models have been proposed to study tax avoidance and associated loopholes, but most existing models are extremely simple compared to the real-world complexity of tax law. We aim to develop automated decision-support tools that can help government agencies detect and mitigate tax loopholes earlier than is possible with the current manual approaches. Here we introduce a benchmark of realistic tax avoidance scenarios based on simulations of corporations and partnerships and the financial actions that they can take. To establish baselines on this benchmark, we implement two evolutionary representations for taxpayer decisions with two different approaches to representing a business's tax-planning policies. We find that a dynamic representation-which accounts for simulation constraints in an online manner-often performs much better than a fixed action-list representation of tax plans.