This is a summary of the author’s submissions as team “IdleSpeculation” to the National Institute of Justice Recidivism Forecasting Challenge, whose basic structure was to forecast arrest likelihoods of former inmates in each of the 3 years subsequent to their release.
The challenge structure introduces some differences in forecasting methodology, primarily due to the availability of attributes in each year. Where possible, the submissions will be presented as a unified approach with yearly discrepancies highlighted when necessary. Minimal variable transformation or selection was applied to the competition data and the bulk of the effort was aimed at building a diverse set of predictions that were blended into yearly estimates via a model stacking technique. Although these yearly estimates did score well in the racial fairness component of the challenge, there were no special steps taken to produce this outcome. A suggestion is made regarding a potential improvement to this part of the evaluation. Although the author feels the competition went smoothly and was well executed by the host, a few ideas for future competition are provided, as well as a comment regarding suitability of the competition data for practical use in predicting recidivism.
Downloads
Similar Publications
- The Experience of Social Distancing During the COVID-19 Pandemic: A Qualitative Lens on Variability in Compliance
- Assessing the Impact of Plea Bargaining on Subsequent Violence for Firearm Offenders
- A Self-assessment Tool for Helping Identify Police Burnout Among Investigators of Child Sexual Abuse Material