4th National Crime Mapping Conference Crime Forecasting for Police Deployment: Results from a Large-Scale Research Programme by Wil Gorr 1 Objective • Assess time series forecasting methods for use by police – Using over 10 years of offense reports and computer aided dispatch calls from Pittsburgh, Pennsylvania, and Rochester, New York (crimeforecasting.org) • Forecast Methods – Univariate extrapolative – Leading indicator – Tracking signal 2 Crime Early Warning System (CEWS) 3 Scenario • It’s the first day of July in Rochester, New York • You have made time series forecasts for July: – For part 1 property crimes (P1P) – By patrol district • You are about to view the forecasts in CEWS 4 Crime Level: July Forecast P1P July Forecast 5 Crime Change: July Forecast 6 Crime Analysis: June’s Crimes 7 Crime Time Series Forecasting 8 Crimes Analyzed P1V (Part 1 Violent Crimes) = Aggravated Assault + Murder + Rape + Robbery P1P (Part 1 Property Crimes) = Burglary + Larceny + Motor Vehicle Theft + Robbery 9 Pittsburgh, PA Rochester, NY City Area (sq. miles) 2000 Population Density Population (persons/sq. mile) Pittsburgh 56 335,000 6,000 Rochester 36 220,000 6,100 10 Geographies Pittsburgh Rochester 42 Patrol Districts 38 Patrol Districts 178 Census Tracts 90 Census Tracts 11 Monthly Time Series Data: Rochester, NY 3500 3000 2500 2000 1500 1000 P1P P1V x 10 500 200012 199912 199812 199712 199612 199512 199412 199312 199212 199112 0 12 Crime Count Extrapolative Forecast Month Now 13 Crime Count Leading Indicator Model Leading Indicator P1V Now Months 14 Leading Indicator Crimes for P1P and P1V • • • • • • • • Arson CAD Drugs CAD Shots Fired Criminal Mischief Disorderly Conduct Drug Offenses Family Violence Gambling • • • • • • • Liquor Law Violations Prostitution Public Drunkeness Simple Assault Tresspass Vandalism Weapons 15 Crime Forecast Accuracy 16 Crime Changes Forecasted Change(t) = Forecasted Crime(t+1) - Actual Crime(t) Actual Change(t) = Actual Crime(t+1) - Actual Crime(t) 17 Sample Decision Rule and Performance Measures • If Forecasted Change >= 4 then issue an Exception Report • Actual Change for Exception Reports >= 4 is a Positive (10%) = 2, 3 is a “Medium Positive” (15%) <= 1 is a False Alarm (75%) • Positive Rate = 100*Positives/Actual Positives 18 Forecast Methods • Chance – Randomly choose 10% of areas each month – Strawman • Extrapolative method: simple smoothing – Deseasonalized with multiplicative classical decomposition • Leading indicator: regression model – Time and space lags of leading indicators – Monthly seasonal indicators – Neighborhood factors • Leading indicator: neural network model – Same variables as in linear regression – www.neuralware.com 19 Forecast Experiments • Design – One-month-ahead forecasts – Rolling, 5-Year estimation window • Cities – Rochester • Forecasted January 1996−December 2001 – Pittsburgh • Forecasted January 1995−December 2001 20 Crime Change Forecast Accuracy Pittsburgh Tracts Positive Monthly Rate Exception Reports Monthly Positives Monthly Medium Positives Monthly False Positives Crime Method P1V Chance 11% 17.0 2.1 3.1 11.8 P1V Smoothing 21% 17.3 5.4 4.1 7.8 P1V Regression 25% 12.6 4.8 3.3 4.6 P1V Neural Net 47% 31.8 8.9 8.5 14.6 P1P Chance 10% 17.0 1.8 2.7 12.5 P1P Smoothing 25% 12.7 4.4 3.2 5.1 P1P Regression 29% 21.7 5.2 5.1 11.5 P1P Neural Net 44% 40.4 7.7 9.1 23.6 178 tracts in Pittsburgh 21 Crime Change Forecast Accuracy Pittsburgh Patrol Districts Positive Monthly Rate Exception Reports Monthly Positives Monthly Medium Positives Monthly False Positives Crime Method P1V Chance 11% 4.2 0.5 0.8 2.8 P1V Smoothing 25% 3.3 1.2 0.9 1.3 P1V Regression 25% 3.8 1.1 1.0 1.6 P1V Neural Net 43% 7.8 2.0 2.0 3.9 P1P Chance 10% 4.2 0.4 1.0 2.7 P1P Smoothing 19% 2.5 0.8 0.8 0.9 P1P Regression 28% 6.9 1.2 2.3 3.5 P1P Neural Net 39% 11.0 1.6 3.3 6.1 42 patrol districts in Pittsburgh 22 Monitoring Crime Time Series Data 23 Crime Count Counterfactual Forecast Year ago Month Now 24 Crime Count Counterfactual Forecast Forecast Error Year ago Month Now 25 l-0 1 l-0 2 02 2 l-0 3 03 3 l-0 4 04 4 Ju 05 l-0 5 Ap r- 5 26 Oc t- 0 5 Step Jump Ja n-0 Oc t- 0 4 Ju Ap r- Ja n-0 Oc t- 0 3 Ju Ap r- Ja n-0 Oc t- 0 2 Ju Ap r- Ja n-0 Oc t- 0 1 Ju 01 1 40 Ap r- Ja n-0 911 Drug Calls 70 Spike 60 50 Step Jump 30 20 10 Spike 0 Approach • Had Pittsburgh Police crime analysts code a sample of jumps and spikes of potential importance • Tuned monitoring techniques based on coded sample • Use cross validation to make performance comparisons among alternative techniques 27 Tracking Signal Alternatives • Comstat method Change from last year Critical value for crime count differences • Standardized frequency Z score for crime count with critical value • Standardized forecast errors Z score for forecast error with critical value • Trigg tracking signal Smoothed forecast errors divided by smoothed absolute forecast error with critical value 28 Cross Validation of Tracking Signals 29 Recommendations 1. Build a crime early warning system 2. Use tracking signals to evaluate performance and trigger crime analysis 3. Use a neural network leading indicator model to trigger crime analysis for prevention 4. Drill down to leading indicators and recent P1 crimes and apply traditional crime analysis methods 30