ANWB: predicting roadside assistance/
The Dutch automobile association (ANWB) worked with us to improve their ability to predict the number and location of calls for roadside assistance. The result was a significant reduction in error and more efficient planning of their roadside assistance service.
ANWB handles more than a million service calls every year (3000 per day, on average). Their goal is, then, to allocate resources (both people and materials) as efficiently as possible. Consequently, they approached Scyfer for help predicting the number of cars that will break down as accurately as possible in terms of:
- Time, i.e. the number of service calls per half-hour
- Place, i.e. the locations of the service calls
Approach of ANWB
In collaboration with ANWB, we analyzed the data and defined two goals:
- Predict more accurately per location, using 33 specific locations instead of just 4 regions
- Predict the expected number of service calls per half-hour.
Data used to achieve these goals included:
- Historical data on service calls and customer base
- Historical weather data and weather forecasts
- Special holidays, high/low seasonal effects (bank holidays, vacation period, etc.) and other seasonal data.
We then adjusted our deep-learning technology to train the predictive model.
In the end, we successfully reduced the level of error in the prediction of service calls from 7% down to 5%, so ANWB’s dispatching department is able to plan resources more efficiently on a monthly and weekly basis. Resource planning can be adjusted on a half-hourly basis to apply the latest available data and recalculate expected demand. The figures below show predicted (red line) and actual (blue line) service calls for each half-hour period per service location.