What forecasting is not

Predictive policing

Geography: streets

Time: hours/days into future

Horizon scanning

Geography: national/international

Time: (many) years into future

Crime forecasting

Geography: national/force/district

Time: weeks/months/(a few) years into future

Forecasting uses data on the frequency of events …

… to estimate how many events will occur in the future

Forecasting breaks down time series into components

Time series = trend +

Time series = trend + season +

Time series = trend + season + other

Many repeating and one-off events can influence time series

Public holidays

Sporting events, festivals, etc.

Emergencies

Forecasting relies on structural stability

But …

… new technologies change crime patterns

… policing itself changes

… unexpected events happen

Forecasts should be expressed as:

If the factors underlying crime remain broadly the same, we expect …

How forecasting can be useful to policing

Scenario 1

A police commander preparing a budget request wants to know how many crimes are likely to occur in a city each month for the next three years

Scenario 2

The officer in charge of shift assignments wants to know how many crimes are likely to occur each day for the next three months.

Scenario 3

A senior detective wants to know how many serious assaults are likely to occur in a district each month for the next year.

But crime analysts …

  • must master lots of skills
  • have limited time
  • have limited software access

⇒ need forecasting to be semi-automatic

Forecasting methods

There are many forecasting methods

There are many forecasting methods

There are many forecasting methods

There are many forecasting methods

There are many forecasting methods

We can assess how good forecasts are by measuring the difference between the forecast and the actual frequency of crime

We can compare the accuracy of different types of forecast

But instead of comparing methods using only one set of forecasts …

… we can use a rolling origin design to compare multiple forecasts

Multiple forecasts help us better understand forecast errors

Multiple forecasts help us better understand forecast errors

Multiple forecasts help us better understand forecast errors

Multiple forecasts help us better understand forecast errors

But …

  • We have only tested forecasts using data from one city
  • We have only tested one scenario

Many US cities make detailed crime data public

Many US cities make detailed crime data public

It is also useful to test multiple realistic scenarios

Scenario 1

how many crimes are likely to occur in a city each month for the next three years

Scenario 2

how many crimes are likely to occur each day for the next three months.

Scenario 3

how many serious assaults are likely to occur in a district each month for the next year.

36 rolling-origin forecasts for 3 scenarios in 12 cities each

⇒ 1,296 tests of the performance of each forecasting method

Conclusion

  • Reasonably accurate forecasting is possible
  • Combining multiple forecasts is typically most accurate
  • Method choice matters: some specific methods (e.g. FASSTER) are wildly inaccurate

Remember

Forecasts are only useful as long as there is structural stability

Slides:
lesscrime.info/talk/forecasting-scs-seminar/

Questions:
matthew.ashby@ucl.ac.uk