NCJ Number
85339
Date Published
1981
Length
24 pages
Annotation
This study assesses the naive model and three other forecasting approaches -- an extrapolative-seasonal model, the Holt-Winters model, and Box-Jenkins procedures.
Abstract
The findings suggest that the naive model (crime in the month to be forecast is assumed identical to and compared with that for the same month in the previous year) often used by law enforcement agencies for public information and evaluation purposes will usually be associated with forecast errors consistently larger than when the Holt-Winters and Box-Jenkins models are used. When applied to crime data, forecasting models with capabilities for coping with changing trends and seasonality (the Holt-Winters and Box-Jenkins models) will provide forecasts superior to those of models with fixed seasonality (the Holt-Winters and Box-Jenkins models) will provide forecasts superior to those of models with fixed seasonality and fixed or no trend terms (the extrapolative-seasonality and naive models); however, for some crimes heavily dominated by random noise, none of the models tested may have forecast errors small enough to allow reliable crime deterrence program evaluation. Although results based on the comparison of a limited number of models for a limited number of crime types for one city are by no means conclusive, this study suggests how methods found useful in business and economics may be applied to crime series and other social data that tend to be underanalyzed. Tabular and graphic data and 20 references are provided. (Author summary modified)