Time-series forecasting for strategic analysis

Matt Ashby
matthew.ashby@ucl.ac.uk

UCL Department of Security and Crime Science

What forecasting is (and isn’t)

Predictive policing

streets

hours/days into future

Horizon scanning

national/international

(many) years into future

Crime forecasting

national/force/district

weeks/months/(a few) years into future

Forecasting uses data on the frequency of events …

… to estimate how many will occur in the future

Past crime patterns are usually the best-available guide to future crime patterns …

… but they’re not a perfect guide

How forecasting could be useful to policing

A commander wants to know how many crimes are likely to occur each month for the next three years

A rosters officer wants to know how many crimes are likely each day for the next three months.

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

How time-series forecasting works

Forecasting breaks 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.

Forecasting relies on structural stability

We must assume the causes of crime patterns in the past will continue in the future …

Forecasting relies on structural stability, but …

… new technologies change crime patterns

… the context of policing changes

… unexpected events happen

(we’ll come back to this)

Forecasting methods

There are many forecasting methods, but …

  • some require expensive specialist software
  • or detailed expertise
  • or a lot of time to implement

Intelligence/performance analysts …

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

⇒ need forecasting to be semi-automatic

We tested 10 different forecasting methods

We tested 10 different forecasting methods

We tested 10 different forecasting methods

We tested 10 different forecasting methods

We tested 10 different forecasting methods

We test how good a forecast is by looking at the difference between the forecast and reality

We test how good a forecast is by looking at the difference between the forecast and reality

We test how good a forecast is by looking at the difference between the forecast and reality

We can compare the accuracy of different types of forecast

… but comparing forecasts repeatedly is better

So we tested each forecasting method:

  • using data from 12 different cities
  • for each of 3 different scenarios
  • for 48 different time periods

⇒ we can say which method is generally most reliable

… but comparing forecasts repeatedly is better

… but comparing forecasts repeatedly is better

… but comparing forecasts repeatedly is better

… but comparing forecasts repeatedly is better

… but comparing forecasts repeatedly is better

Some tips for forecasting

  • If you have time, test multiple methods on previous data and use the most-accurate method
  • If not, use the online forecasting tool
  • Use at least 3 years of previous data for monthly forecasts and 90 days of data for daily forecasts
  • Update forecasts frequently to make use of the latest data
  • Combine quantitative forecasts with other information, e.g. intel reports

But what about structural stability?

… new technologies change crime patterns

… the context of policing changes

… unexpected events happen

But what about structural stability?

Forecasts should be expressed as:

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

Any analysis using forecasting should highlight specific threats to structural stability

These slides:
lesscrime.info/talk/forecasting-sebp/

Questions:
matthew.ashby@ucl.ac.uk