PredPol is based on a decade of detailed academic research into the causes of crime pattern formation. That research successfully linked several key aspects of offender behavior to a mathematical structure that is used predict how crime patterns will evolve from day-to-day, from moment-to-moment.
The mathematics looks complicated, but the behaviors upon which the math is based are very understandable. There are three aspects of offender behavior that make their way into our model.
- Repeat victimization, which describes – taking burglary as an example – that if a house is broken into today, the risk that it is broken into tomorrow actually goes up. This is because it is “rational” for offenders to return to the places where they have been successful before. It makes less sense to go to some other unknown house where they don’t know if the house is empty of people, they don’t know how hard it is to break in, and they don’t know what there is to be stolen. The house they broke into two or three days ago is much less risky.
- Near-repeat victimization, which recognizes that not only is your own house at greater risk of being broken into again, but your neighbor’s house is also at greater risk. Your neighbor is a lot like you: they they have similar socio-economic status, work similar hours, have a house a lot like yours and are going to have much the same stuff to steal. The offending ‘script’ the offender used to break into your house maps to your neighbor’s house almost perfectly.
- Local search ties it all together. We know that offenders rarely travel very far from their key activity points such as their home, work and play locations, meaning that crimes tend to cluster together.
The actual patented algorithm used by PredPol is displayed below:
Implementation: Quick and Easy Deployment
PredPol uses data from your agency’s records management system (RMS) to pull current and historical crime data. We then feed this into our machine-learning algorithm to create our predictions. We work with you and your RMS vendor to make sure that the data we use is accurate and complete. We only use 5 data points for each incident to generate our predictions:
- Incident Identifier – For each crime, we need a unique identifier, such as docket number, incident ID or anything else used by the department to uniquely identify the crime.
- Crime or Event Type – The violation code and/or crime description assigned to a particular incident type as used in your RMS.
- Location of Incident – For best accuracy, latitude and longitude are desired. Your latitude and longitude must use the WGS 84 coordinate system. If latitude and longitude are not available, then the complete address of the incident is required. A complete address is Street Number, Street Name, City, State/Region.
- Timestamps with Start and End Date/Time for Incident – We use these two fields because in some cases the exact date and time that the crime occurred is not known. For example, an auto theft may occur between midnight and 8 AM, or a burglary may occur over a weekend.PredPol calculates the incident occurrence time by taking the midpoint between beginning date/time and ending date/time. Incidents with a span of more than 72 hours between the beginning and ending date/time are excluded because it reduces the accuracy of our predictions.
- Record Modified Date/Time for Incident – This is an optional field, but where possible we also request that you include a “record modified” date/time field to allow us to catch RMS records that may have changed (i.e. crime code has been reclassified).