Date Published: October 24, 2018
Publisher: Public Library of Science
Author(s): Sherry Towers, Siqiao Chen, Abish Malik, David Ebert, Hedwig Eisenbarth.
Improving the accuracy and precision of predictive analytics for temporal trends in crime necessitates a good understanding of the how exogenous variables, such as weather and holidays, impact crime.
We examine 5.7 million reported incidents of crime that occurred in the City of Chicago between 2001 to 2014. Using linear regression methods, we examine the temporal relationship of the crime incidents to weather, holidays, school vacations, day-of-week, and paydays. We correct the data for dominant sources of auto-correlation, and we then employ bootstrap methods for model selection. Importantly for the aspect of predictive analytics, we validate the predictive capabilities of our model on an independent data set; model validation has been almost universally overlooked in the literature on this subject.
We find significant dependence of crime on time of year, holidays, and weekdays. We find that dependence of aggressive crime on temperature depends on the hour of the day, and whether it takes place outside or inside. In addition, unusually hot/cold days are associated with unusual fluctuations upwards/downwards in crimes of aggression, respectively, regardless of the time of year.
Including holidays, festivals, and school holiday periods in crime predictive analytics software can improve the accuracy and precision of temporal predictions. We also find that including forecasts for temperature may significantly improve short term crime forecasts for the temporal trends in many types of crime, particularly aggressive crime.
In recent years, predictive analytics and informatics software has become an essential tool for police forces nationwide to visualize and predict patterns of crime [1–7]. Necessary to the predictive performance of these tools are statistical models that take into account not only secular trends, but also various exogenous factors that might influence the incidence of different types crime; for example, climate, daylight hours, day-of-week, and holidays and festivals.
As seen in Table 2 and Figs 8 and 9, we find that a regression model that includes trend, smooth annual periodicity, holidays and weekdays has significant predictive power for all types of crime examined in this analysis (R2 ≥ 0.43 in all cases). This is in agreement with previous studies [14, 15, 21].
In this analysis, we examined a sample of 5.7 million crimes in the City of Chicago over a 14 year period. We examined a wide array of exogenous factors that might be related to short term temporal trends in crime, including annual periodicity, holidays, day-of-week, paydays, and climate variables, including temperature, humidity, wind, air pressure, and precipitation. In our analysis, we examined various types of crime, corrected for auto-correlation in the data, and used bootstrapping methods for robust model selection. To our knowledge, ours is the first analysis of this type to correct for auto-correlation using harmonic regression methods, and to employ bootstrapping model selection methods that account for the multicollinearities in the explanatory variables. The analysis largely confirms the results of past studies examining the relationship of crime to these exogenous variables, particularly for a linear relationship between aggressive crime and temperature.