Weekly Time-Series Forecast
NOTE (7.30.2020): The events throughout the US as of late bring up a point which I did not mention in the original creation of this forecaster that I should have addressed; One has to be very careful in selecting what factors are used as inputs to determine criminality and crime in general. Machines can very well be racist, but being a forecaster based on only time and location, this algorithm, I think, has it's benefits in the research in pattern recognition and neighborhood-by-neighborhood 'criminal activity' (as determined by the City of Chicago) identification. My overall thoughts on crime and the criminal justice system at large are.. opinionated.. to say the least. I'll simple leave this by stating:
This project was a research project and in no way condones or suggests that a benefit to society can be found in the idea of 'predictive policing' based on any human or machine decided algorithm/methodology.
This is a crime rate forecaster for the City of Chicago. It predicts overall weekly occurrence rates and maps them out to their expected distribution across city neighborhood and time grouping (8-hour intervals).
It uses a SARIMA time-series forecast, detailed here, trained on data obtained from Chicago's Data Portal. Evaluation in training resulted in an accuracy of 94.2%, with a mean error per location-time matrix element of 0.2.
Data, model, and visualizations are updated through an automated system, lightly detailed in a post here.
This will be updated every Sunday.
Or view a larger version of the Forecast.
Or view a larger version of the Prediction Matrix.