About
Tips in Sberbank Online are a tool based on artificial intelligence that can help users change financial habits for the better, save money, time and effort or just remember something important at a certain moment of time.
Innovation presentation
Sberbank presents a service called Tips. It is a part of Sberbank Online mobile app for private customers. It is based on artificial intelligence that help users save money, time and effort. Tips analyse the user’s financial behavior. It uses a set of machine learning models, including experiments with neural networks.
This service is a part of the deep evolution of personal finance tools. It makes it possible to break free from traditional analysis of expenses, look to the future and change one’s financial habits.
It has been created at the intersection of digital assistants such as Siri and financial services like Mint and offers the advantages of each: proactivity, analysis of financial behavior, machine learning, provision of recommendations for everyday tasks, and not just money-related ones.
In comparison with similar services of financial recommendations Tips in Sberbank Online have extra features like collection of feedback, links that allow going by advice, etc.
Eight Sberbank departments worked to create Tips (to deliver the concept, develop the product, generate content, create algorithms, support clients, etc.).
In February, Tips become available to 20 million active users of Sberbank Online. Currently the service is used by around 200,000 users daily. We are gathering feedback, assessing existing recommendations and preparing new ones. There are 50 tips available now, in a month there will be 100, and this number will continue to grow.
Besides adding new tips we intend to expand functionality, experiment with new possibilities (USG, clarifying questions etc), and add third-party content suppliers.
Uniqueness of the project
Our service combines the characteristics of two types of financial products. On the one hand, there are PFM services, where our competitors are Mint, LearnVest and Moven. They also provide their clients with advice, but it is primarily linked to the purchase of relevant banking products. Advice in Sberbank Online goes beyond banking services by informing users about offers of partners, as well as events and life hacks that aren’t related to any brands.
On the other hand, Tips competes with digital assistances such as Siri from Apple and the Assistant from Google, but they are less related to financial behaviour of users and do not provide advice if users do not ask for it.
This means that Smart Tips creates unique value by combining financial data, machine learning and personalised recommendations about money and everyday life.
We've seen similar services of financial recommendations in other banking applications and pilot projects. Here are a few benefits of our Tips for Sberbank Online users:
• The user can easily give feedback that allows us to target recommendations more precisely.
• It's not just about advice – many tips end with the script to go to a website or moving within the application with filled-in fields. Thus, the recommendations are more likely to lead to concrete action and help the client.
• We are steadily expanding to different branches of life scenarios, rather than randomly moving in different directions, so you can be sure that in certain areas the service will provide its recommendations.
• Now we are working on functions of socialization for the Tips users the opportunity to share them with each other; the ability to share information about a sales point in a particular institution - they will come to a friend also in the form of advice (e.g. "Look, they have great soups here! Definitely come and check them out, you'll love it") after the transaction.
The service is built on the basis of technology that includes several methods of machine learning. For example, the gradient boosting algorithm helps predict events in the life of clients – for instance, it is possible to predict that the client is going on holiday, and the assistant may provide them with personal advice. Meanwhile, reinforcement learning algorithms help take into account feedback from clients in the optimal way for elaboration of models, and show only the most valuable advice – each new piece of advice is first shown to a small part of the target audience for feedback and then the target audience is optimised automatically.
The digital assistant accumulates a very large amount of feedback from our clients and uses it for self-training. If a client says several times that he/she is not interested in the advice on a specific topic, the system decides if this is bad advice or this very client is not interested in such advice. On the basis of this feedback, the digital assistant will continuously improve its advice targeting skills with respect to our huge audience. Our objective is to make the digital assistant at a certain abstract level generate and provide clients with advice, without interference from our content managers.
Logically all the advice can be presented like a big tree or a mind map. Every node of the tree is an insight about the user which is defined with a mathematical model. Models differ in terms of complexity and the algorithms which they are based on. Simple models are based on trigger detection (the user bought something or did something) whilst complex models are based on machine learning algorithms which can predict the user’s behaviour (for example we can detect that the user is going on a trip somewhere or he/she is going to buy a car or other big item).
