Source: Department for Transport, Bank interpretation and calculations.
The data for 2019 is estimated. The Department for Transport publishes a list of all the vehicles that have been recalled because of a safety issue. The list includes over 5,500 recall notices that have affected over 33 million cars, trucks, coaches, vans and motorcycles since 1992. For each vehicle that is recalled the concern, defect and the remedy are provided along with the year of the product recall, the number of affected vehicles and the make and model of the vehicle. We have allocated each product recall into categories ('software', 'hardware/sensors' or 'other') based on the cause of the recall. In order to do this we manually categorised 20% of the entries and used these to calibrate various machine-learning techniques. Each machine-learning technique was calibrated using c.800 manually categorised recalls. Each technique was then tested against the remaining subset of c.200 manually categorised entries to assess its accuracy. The most accurate machine learning technique was a unigram model in conjunction with a random forest algorithm. The technique involved estimating the categorisation based on thousands of decision trees, each of which classified the data according to a set of rules based on single word counts in the text, and then denoting the ultimate categorisation as the most common classification amongst the trees. The machine-learning model that provided the greatest degree of accuracy in the test sample was then used to assign the categories 'software','hardware/sensors' or 'other' to each of the remaining c.4500 product recalls. We then reviewed each of the c.600 product recalls that were allocated to the categories 'software' or 'hardware/sensors' and checked these had been allocated correctly. Eleven of the c.600 were identified to have been inappropriately allocated and were reallocated manually. We were then able to see trends in the number of recalls and the number of vehicles recalled each year due to issues with their software or with their hardware/sensors. We have allocated all product recalls and not only those specifically in relation to the ADAS features included within Chart A. The data included recall notices up until 12 September 2019 and so we scaled up the 2019 results on a pro-rata basis to reflect a full year.
With thanks to the Bank of England’s Advanced Analytics Division for their support in employing the machine-learning techniques.
The data used from the Department for Transport to create this chart is public sector information licensed under the Open Government Licence v3.0. The Open Government Licence can be found here.
