Machine Learning Forecasts of Risk to Inform Sentencing Decisions

Abstract

When threats to public safety are a factor in sentencingdecisions, forecasts of ‘‘future dangerousness’’ are neces-sarily being made. Sometimes the forecasts are effectivelymandatory. Federal judges, for example, are required toassess risk in every case. Under 18 U.S.C. § 3553(a)(2)(C),‘‘[t]he court, in determining the particular sentence to beimposed, shall consider...(2) the need for the sentenceimposed...(C) to protect the public from further crimes ofthe defendant...’’ A judge must look into the future,determine the likelihood and seriousness of criminalbehavior, and within certain bounds, sentence to minimizethe harm that could result.Ideally, the forecasts should be highly accurate. Theyalso should be derived from procedures that are practical,transparent, and sensitive to the consequences of forecast-ing errors. However, there is usually no compelling guid-ance on precisely how these goals can best be achieved.Subjective judgment, sometimes called ‘‘clinical judg-ment,’’ is an approach that relies on intuition guided byexperience. As discussed below, the resulting risk assess-ments are often wildly inaccurate and their rationale opa-que. ‘‘Actuarial’’ methods depend on data that allow one tolink ‘‘risk factors’’ to various outcomes of interest. Theassociations found can then be used to forecast those out-comes when they are not known. Over the past severaldecades, regression statistical procedures have dominatedthe actuarial determination of empirically based risk fac-tors. By and large, this enterprise has been a success. Butthe increasing availability of very large datasets coupledwith new data analysis tools promise dramatically bettersuccess in the future. Machine learning will be a dominantstatistical driver.There is now a substantial and compelling literature instatistics and computer science showing that machinelearning statistical procedures will forecast at least asaccurately, and typically more accurately, than olderapproaches commonly derived from various forms ofregression analysis.