Tax Collection Optimization for the State of New York
The Problem
The State of New York collects over $1 billion annually in assessed delinquent taxes. Recent economic conditions, including the need to reduce growing state deficits, require that every dollar that can be collected is collected—and in the most efficient way possible. The debt collection process is highly complex, with a large number of legal and business constraints involved. Because of this complexity, it is common practice to follow rigid, manually constructed rules to guide the collection activities. Even in state-of-the-art rule-based systems for collections, the role of data analytics is typically limited to augmenting the rule engine with scores given by analytics. Manually constructed rules are not adaptive to changes in the environment, and hence the need for a more adaptive approach has been recognized.
The Analytics Solution
In collaborative work between the New York State Department of Taxation and Finance and IBM’s Research and Global Business Services divisions, a novel tax collections optimization solution was developed to address this challenge. The solution is a unique combination of data analytics and optimization based on the unifying framework of constrained Markov decision processes (C-MDP). It optimizes the collection actions of agents with respect to maximization of long-term returns while taking into account the complex dependencies between business needs, resources, and legal constraints.
The Value
The C-MDP system became operational in December 2009, and the results to date are compelling. New York State has seen an $83 million increase in revenue from 2009 to 2010 (8%), using the same set of resources. In addition, the department has increased the productivity from primary enforcement actions with, for example, a 22% increase in the dollars collected per warrant and an 11% increase in the dollars collected per levy. (A warrant is effectively a “tax lien,” and a levy compels financial institutions to turn over debtors’ assets.) Further, the department has seen a 10% decrease in age of case when it is assigned to a field office. (Older cases tend to have lower probability of collection.) Given a typical annual increase of 2–4%, the expected benefit of the developed system is approximately $150 million over three years, far exceeding the initial target of $99 million, and is expected to improve still more in the future as the system further adapts. The system brings together three principal elements of operations research—analytics, optimization, and rules—in a novel, coherent, and effective manner, and simultaneously attains their corresponding benefits of adaptability, optimality, and practicality.
