Part two: How Tink launched in the UK in record time
Here comes the second piece of our three-part Project Win UK series. In part one, we showed how our engineers smashed expectations to launch Tink in the UK in four weeks. Now we dive deeper to walk you through the first two weeks of the project, showing how we built connections with the banks and trained our categorisation model – at lightning speed.
Launching in the UK was always going to have its challenges. There are Open Banking UK regulations on top of PSD2. Consumer spending patterns and consumer behaviour is different in the UK, and we didn’t have any data to base a categorisation model on. So where did we start?
Our first job was to integrate with the banks using the APIs. With standardised APIs written into UK regulation, this was actually less complicated than we are used to. But even with prescribed APIs there is room for banks to interpret the rules differently. So once we had established a connection, the testing phase was really important.
Then step two was to build our categorisation model – which cleans up the transaction descriptions and segments them into the proper categories (restaurant, rent, insurance). To get real value from the data that’s aggregated from a customer’s bank accounts, it needs high-quality categorisation. This way, our partners can offer insights based on a more complete understanding of how a customer spends their money – and help them manage it better.
So we set about feeding our machine-learning algorithm with UK data. Lots and lots of data. It’s called training data because it teaches the algorithm what the data means and how to interpret it to offer more accurate conclusions. Collecting good training data – both in terms of quality and size – is a true challenge of scale, and it can’t be done solely by machines. Categorising what someone spent their money on can be subjective. We need real people to help train the system. So that’s exactly what we did.
We got users in the UK to fill our system with transaction data to help train the model. In the first two weeks, we could bug test the connections with banks, learn which transactions were being re-categorised, and keep feeding it back into our machine-learning algorithm.
As we were doing this, we noticed a big difference in the UK. It is a nation of chain restaurants, bars and shops. Which means the accuracy of categorisation had the potential to be the highest in Europe, because there are many more repetitive transactions.
By the end of the first two weeks, we could aggregate account information from all of the UK’s largest nine banks, and we could categorise the data to a really high level. We could have stopped there, and Tink could have launched in the UK in two weeks. But we wanted more. We wanted to show what we really meant by “Project Win UK”.
We knew weeks three and four were going to be busy.
Join us for part three of our Project Win UK series, when every Tink engineering team gets on board for the final push – to add more services, connect to more banks, and make better integrations – to complete our mission to go live in the UK and “win” it.