The purpose of this applied research is three-fold: (1) to link theory and appropriate data/measures to the practice of predictive policing; (2) to determine the accuracy of various predictive policing algorithms to include traditional hotspot analyses, regression-based analyses, and data-mining algorithms; and (3) to determine how algorithms perform in a predictive policing process.
The project's primary phases included: (1) employing report card strategies to analyze, review and evaluate available data sources, software, and analytic methods; (2) reviewing the literature on predictive tools and predictive strategies; and (3) evaluating how police agencies and researchers tested predictive algorithms and predictive policing processes.
Primarily, this study utilized a secondary analysis of time-series crime data obtained from DPD (2013-2015). These data were used to create benchmarks for evaluating the effectiveness of various crime forecasting algorithms.
As part of the data assessment process, researchers created data report cards to evaluate publicly available crime datasets on data quality and usability. For practitioners seeking implementation, vendor assessment tools to evaluate predictive software were created. Researchers also compiled a matrix of available software and models for predictive analysis, highlighting type of data used, mode of analysis, and crimes predicted.
All crimes reported to DPD between January 1, 2013, and December 31, 2015 were included in the datasets for analysis.
All records of crimes reported to Denver Police Department (DPD) within the study time frame.
Crime dataset from Denver Open Data Catalog
Crime dataset variables include date and time of crime event and reporting, offense description, geographic grid, and NIBRS and UCR codes.
In the grid cells datasets, crimes are categorized into 5 groups: assault (excluding homicide and sexual assault), auto crimes, burglary, robbery, and gun-involved crimes. Variables include aggregated monthly counts for each type, forecasts for various baseline algorithms, and evaluations (hit, miss, false negative, false positive).