Forecasting murder within a population of probationers and parolees: a high stakes application of statistical learning