All financial services providers need more personalized offerings when interacting with clients and especially new clients. As a client I want to get some personalized recommendations on what to do next or which product to choose. Banks and other providers have had very limited personalization capabilities so far, this might be due to regulatory requirements or lack of historical data on past client behavior. This article reviews known recommendations techniques and their applicability for financial services

Personalization approaches: Collaborative Filtering

There are a number of very good articles and courses on recommender systems ( for example, so the idea is not to talk about them in depth, but to talk about their application in financial services

User-User Recommendations

This type of recommendation system tries to find an item that similar users have already chosen. An item will be for example an article, a movie or a shopping article. Similar users are those that have a similar interaction or buying profile in our case. Based on similar users the system will propose items that are popular with similar users.

Their application in financial services seems quite difficult, but feasible technically. The challenge is to find enough user interactions to be able to infer relevant products. Usually clients do not have many products in their portfolio and might have a few accounts, credit cards, a saving account, a trading portfolio and maybe mortgages. On top of that clients might not buy products very frequently, which again impacts the volume of data that can be available to make relevant recommendations.

Item-Item Recommendations

In this case, the underlying algorithm will try to find similar items a user has already selected in the past. So if you have purchased for example a book from Harlan Coben, it may propose next time a similar crime book from another author. Obviously if you are a new user, proposed products might be random or picked from popular products.

The application in financial services does not seem to make sense in most cases. Let us assume somebody has opened purchased a mortgage, so other mortgage types might be found as similar by the system given mortgage characteristics for example. It does not make sense for somebody in most cases to purchase a mortgage again unless their current mortgage is expiring. This means that such a system would have to be implemented together with a number of business rules that might override the algorithm in most cases, so better implement a rule-based system probably if we want to go that route.

Matrix Factorization

Matrix factorization in some way combines both approaches and defines a matrix of features that are relevant for both users and items. We won’t go into further details here, as this approach will suffer the same issue as the user-user approach.

Personalization approaches: Content Based

 In content based approaches, the problem becomes a classification problem whereby we try to predict what kind of article or item the user is likely to buy or click. The features used are characteristics of the user that allow to classify user against potential products. The challenge here is to have enough user data to build a model and also have historical data to have a solid predictive model. This is in my opinion the best suited model for financial services. Providers need to collect  in any case enough user information to open a relationship during onboarding and this can be used to propose best matching products. For existing clients, enough data has been collected to have enough features we can use to predict future or best action.

Building an incremental personalization engine

We have 2 options to build a recommendation system:

·      Collect existing data and build the model based on historic data. The issue with this approach is that it might not reflect the current products and therefore would have less value when new products are added. This could be circumvented using some item similarity for example when proposing products. In this case we would use a superposition of 2 models, the classifier and the item-item model

·      Build an incremental model from scratch that will adapt automatically when new products are added and that will improve over tome without manual intervention

The latter solution is the preferred one as this offers flexibility and adaptability, It also offers automatic updates when new data arrives. Most of the models need to be retrained regularly due to concept drifting. The phenomenon is due to change of behaviors of users for example or in our case due to changing products.

If you are interested to discuss more about the topic, please contact me for more details.

(Visited 48 times, 1 visits today)
%d bloggers like this: