Big Data and Predictive Analytics in Government Finance: Transforming Fraud Detection and Fiscal Oversight

Abstract

Governments around the world are experimenting with Big Data and predictive analytics. They deploy various software applications like predictive policing, fraud detection, and capacity demand prediction while at the same time developing and investing in broader data analytical infrastructure and analytical skill sets. Implementing Big Data and predictive analytics can be a challenging endeavor, however, as these analytics often rely on open-source algorithms that are unsupervised and black-boxed. Equally challenging is how government institutions endeavor to use a waterfall approach from exploratory to model predictive analytics, yet how often predictions are only perfunctory and unempirical [1]. To shed light on how government finance organizations conceptualize and prepare for such analytical disruption pertaining to predictive analytics specifically, data processes, stakes and concerns were articulated based on in-depth interviews with 23 analysts who work with predictive analytics in a regional government agency. Descriptive coding of the interviews revealed that changes to data processes are being prepared or enacted, but many foreseen stakes and concerns about changes to both data processes and knowledge processes remain unresolved. New agendas to address such issues and better understand the approaches adopted were proposed.

Governments across the world are aiming to exploit Big Data and associated predictive analytics to govern more effectively and efficiently. These analytics come in many sorts and varieties. In government finance, the topical applications of predictive analytics have to date mainly been found in fraud detection, capacity demand prediction, budget revenue prediction, and the prediction of homelessness and recidivism. A plethora of software applications built on open-source predictive analytics algorithms exist, encompassing packages for forecasting demand, and estimating regression models including linear and logistic types. However, there is some hesitancy in adopting most of these analytics, as open-source predictive analytics algorithms are rarely supervised and almost always deployed as black-boxed. Black-boxified analysis is countercultural to the emancipation and democratization of knowledge advocated in government, as well as other more mundane concerns about accountability and validity. With black-boxification work piling-up on agency knowledge processes, concerns arise about how this analytical work is handled, displayed, devised and/or aggregated to produce knowledge that meets government quality expectations of reproducibility, replicability, auditability, and trainability.