Catching Lightning in a Bottle: Forescasting Next Events

1
Catching Lightning in a Bottle Forescasting
Next Events

• Presented by
• Dr. Derek J. Paulsen
• Director, Institute for the Spatial Analysis of
  Crime
• Assistant Professor
• Eastern Kentucky University
• 2005 iPSY Conference

2
Spatial Forecasting and Crime Analysis

• Evolution of Crime Analysis in the U.S.
• Increasing focus on Tactical Analysis and
  assistance in major crime investigations.
• Increasing use of advanced technology
• Geographic profiling
• Crime Series Identification software
• Forecasting/Prediction
• Great potential to assist in investigations, but
  research has been limited.
• Developing Crime Series Analysis tools and
  training as part of a NIJ grant.

3
Main Research Questions

• How accurate are traditional strategies in
  comparison to TWKDI at predicting the location of
  a future crime event in an active crime series?
• Under what circumstances do forecasting
  techniques work?
• Are there crime types that are better for
  forecasting than others?
• What case specifics best predict success?

4
Forecasting Strategies Studied

• Traditional Methods
• Standard Deviation Rectangles Gottleib
  Rectangles
• Jennrich/Turner Ellipse
• Minimum-Convex-Hull Polygon
• New Methods
• Modified Correlated Walk Analysis
• Time-Weighted Kernel Density Interpolation
• Control Method
• Modified Center of Minimum Distance

5
Standard Deviation Rectangle
2 Standard Deviation rectangle around the mean
center of the incident locations in the series
6
Jennrich-Turner Ellipse
2 Standard Deviation ellipse based around the
mean center of the incident locations in the
series and drawn around a least squares trend
line
7
Minimum Convex-Hull Polygon
Creates a minimum bounding polygon around all of
the incident locations in the series
8
Modified Correlated Walk Analysis
Uses the CWA as a seed point and creates a search
area by drawing a circle with a radius of the
average distance between crime events in the
series.
9
Time-Weighted Kernel Density Interpolation
Kernel Density Interpolation of crime incident
locations using time as a weighting variable
10
Modified Center of Minimum Distance
Uses the CMD as a seed point and creates a search
area by drawing a circle with a radius of the
average distance between crime events in the
series.
11
Data Used in Study

• 247 serial crime events that occurred in
  Baltimore County, MD between 1994-1997.
• Random sample of 45 cases in which there were 6
  or more incidents.
• Series ranged from 6-14 events
• Burglary, Robbery, Arson, Auto theft, Rape, Theft
• Last Crime was removed from series and remaining
  crimes were used to predict the final event.
• Analysis was conducted using
• Arcview 3.3 and 9.0
• Crimestat 2.0
• Animal Movement Extension/CASE Program

12
Measuring Accuracy of Predictions

• How do you measure accuracy in predicting next
  events in a crime series?
• Accuracy in prediction needs to encompass both
  correctness and the precision of the prediction
  in order to maintain practical utility.
• A prediction may be accurate, but the predicted
  area may so large as to provide little practical
  benefit.
• Methods
• 1. Correct Was the final event location within
  predicted area.
• 2. Search Area Average size of the predicted
  area.
• 3. Search Cost Percent of base search area
  covered by the final predicted area.
• 4. Accuracy Precision of correct forecasts
  divided by the average predicted area.

13
Search Area, Search Cost, and Accuracy Precision
Average base search area was 92 sq. miles
14
Success by crimes in series
Average 57
15
Average distance between crimes
16
Dispersion by Crime in series
17
Search Area size by number of crimes in series
18
Accuracy/Precision by crime number
19
Commercial Burglary Series

• 5 crimes within 6 days.
• Stealing cigarettes from gas stations
• Crime area of approximately 10 square miles
• Over 409 businesses within the area.

20
Commercial Burglary Series

• 8 gas stations within initial crime area
• 22 gas stations within area and 1/2 miles
  surrounding it.

21
Commercial Burglary Series

• Prioritized search into two main areas of .9
  square miles
• Top area contained 3 gas stations
• Second tier area contained 3 gas stations

22
Commercial Burglary Series

• Last station burglarized was within top priority
  search area.

23
Overall Findings

• Time-Weighted is the best at reducing the search
  area while remaining accurate.
• Success most influenced by number of incidents in
  series and the distribution of the crimes.
• Convex-Hull Polygon and modified CMD also
  produced good results, whereas other traditional
  strategies were poor performers.
• While average predicted areas are rather large,
  practical use could reduce them to smaller area.

24
Future Issues

• More research, more data.
• Determine impact of other factors such as crime
  type, city type, and road network.
• Determine case variables that may indicate
  predictive success.
• Develop and analyze other new strategies.
• Temporal as well as spatial forecasting/prediction
  
• More research on serial offender spatial and
  temporal behavior.

25
Data or Suggestions?

• Contact InformationDr. Derek J.
  PaulsenAssistant ProfessorDirector, Institute
  for the Spatial Analysis of CrimeEastern
  Kentucky UniversityRichmond, KY USA
  40507-3102Derek.Paulsen_at_eku.edu859-622-2906