Private Banks are looking to digitalize their business processes and use the latest techniques pioneered by on-line business models. A bank’s core business involves a client relationship manager guiding a client’s investment decisions. We use techniques developed by retailers like Amazon to pioneer recommender technologies in banking. We leverage these tools to design a recommender system for private banking with personalized and automated selection of next best products for private banking clients.
Private Banking – Not a Typical Use Case for Recommenders?
Practically every on-line recommendation uses an AI-based recommender engine. Generally, the application of such systems has been in entertainment (movies, books and songs) and e-commerce. These two fields share several beneficial traits:
- Partial availability of explicit feedback as a client/user rating such as 5 stars or like / dislike
- Time-constant product meta-data. For example, if a movie is classified as a western it remains classified as such.
- A bad recommendation has usually no long-term effect on the user
These traits do not apply when we design a recommender system for private banking. Private banking clients typically don’t provide ratings to the financial products they buy or get proposed by their bank. Furthermore, product features can change over time; for example, a product which was classified as a value stock a year ago could now be a momentum stock. Finally, as the recommendations have financial impact they must be robust, transparent and fit into the client’s current portfolio.
Despite these challenges a recommender engine is highly desirable. There is a large range of investment products about which a relationship manager cannot possibly know all the details. Furthermore, a relationship manager typically oversees 50 to 100 client portfolios and has to monitor financial markets and economies on a global scale. In spite of this, clients expect a personal service without detailing all their preferences. A recommender system for private banking seeks to address these challenges.
Private Banks employ portfolio optimization techniques to recommend a suite of products satisfying a client’s risk profile and investment goals. However, this process does not consider the client’s affinity to the selected financial products. We augment the current portfolio tools available at private banks with Collaborative Filtering (CF) techniques to make personalized suggestions for the next best product that fits best into the client’s current portfolio in terms of utility (portfolio optimization) and affinity (AI-based recommender system).
Collaborative Filtering Techniques
Collaborative Filtering techniques attempt to find correlations in the consumption patterns among all clients. The starting point is the ratings matrix. The rows and columns of the matrix refer to users (e.g. private banking clients) and items (e.g. financial products) respectively. The matrix can be filled with positive and/or negative ratings, e.g. a tick in row i and column j would indicate, user i likes item j. A cross could indicate user i does not like item j. A missing value would mean, we don’t know, whether user i likes item j.
These matrices are typically very sparse, having as many as 99% or more missing values. The sparsity is a function of the number of available products; the larger the product universe, the sparser the ratings matrix.
The Ratings Matrix and Implicit Ratings
There are no explicit ratings available that can be exploited by a recommender system for private banking. How then can we construct the ratings matrix? Provided that we have access to transaction data or historical portfolio holdings, we can work out a proxy, called implicit rating. The simplest approach is to assign a value of 1 to products which were once a part of the client’s portfolio. Here we are equating purchase with a positive rating. A slightly more sophisticated model takes the investment holding period as a proxy. Here the implicit rating is the length of time that a product has been hold in a client’s portfolio. The assumption here is that the longer a client holds a product, the happier he is. This is a coarse proxy to a complicated decision-making process. However, it should capture more nuance than the first approach.
These are not the only approaches to model implicit ratings. One could consider further indicators such as price evolution. Moreover, patterns in transaction data combined with market data also yield valuable insights into modeling of implicit ratings.
In our experience, the approach to model implicit ratings is a key consideration when applying a recommender system for private banking and has high impact on the prediction performance.
The Collaborative Filtering Algorithm
Missing values in the ratings matrix do not indicate a negative rating. Rather, they expose our lack of knowledge of the client preferences. Initially, the ratings matrix contains a zero in these cells. However, this does not indicate a negative preference. The Collaborative Filtering algorithm predicts these values by uncovering the correlations in the ratings matrix. The output is a new ratings matrix with imputed values. The imputed values derive from similar transaction patterns between clients. As a result, they should produce interesting recommendations on average. This is in essence how a recommender system for private banking adds value through statistical relationships between clients and transactions.
Ensemble Techniques and Boosting
Boosting is a common feature of AI systems. The aim is to maximize predictive power by combining an ensemble of predictors. Each component predicts particularly well a subset of the data. The Collaborative Filtering algorithm is essentially a predictor of unknown ratings which are used to select the next best product for each client. Thus, we apply boosting techniques to improve these predictions. We developed AdaCF, a boosted collaborative filtering algorithm, and presented it on the 5th Swiss Conference on Data Science.
Product Binning or Categorization
We address the sparsity problem by applying a categorization technique to the product space. Accordingly, a defined set of product features defines discrete product bins. We will refer to these product bins as ‘items’. This procedure brings several benefits:
- Reduction of the product universe leads to a less sparse ratings matrix
- Reduction of the dimensionality leads to a faster running time
- The cold-start problem for new products disappears; a new product will belong to a pre-existing product bin
The output of the Collaborative Filtering algorithm is a defined number of recommended product categories/bins per client. Proper product categorization is a crucial process when designing a recommender system for private banking. The next section describes how we produce the final recommendations of tradeable securities.
Next Best Product Selection
The product selection component chooses the next best product for a client from a recommended category/bin. Initially, the candidate products are filtered according to the bank’s business logic. This logic differs from bank to bank. For example, a common requirement is that a recommended product is not already a part of the client’s current portfolio. Furthermore, other filters consider regulatory requirements or explicit client preferences.
After having imposed the business logic we need to choose the best product for the client in question. We take a novel approach to this problem by applying a feature-based algorithm based on the client’s attributes. The approach chooses the product based on the consumption patterns of other clients who are most similar to the client in question. Similarity is defined with a suitable metric based on the client meta-data. Thus, the resulting framework is a hybrid model combining the advantages of Collaborative Filtering with a feature-based model.
The Long-Tail Problem
The long-tail problem is common to most recommendation systems. In general, most users consume only a small sub-set of items. Conversely, most items are purchased by very few users. The diagram below illustrates this:
One of the goals of a recommender system for private banking is to suggest novel products from the long-tail region.The fewer users have consumed an item, the less likely it is that the Collaborative Filtering algorithm will detect a correlation pattern. Thus, product categorization improves this situation as it leads to a denser distribution.
How to Measure Quality of a Recommender System for Private Banking?
How can we assess how good our recommendations for the next best product are? One technique is cross-validation. Here we split the data into training data and test data. The training data is used to train the algorithm. Next, we run the algorithm on the test data and calculate several quality metrics. A common metric is the so-called Area Under Curve. This measures how likely items consumed by the user come out on top in the final ranking matrix.
The Collaborative Filtering algorithm uses several parameters whose optimal settings depend on the data set. We apply a parallelized grid search to find these optimal parameters. Here, optimality is defined in terms of the best quality metrics.
A recommender system for private banking running in production provides further data which we can use to assess the recommendations. Besides just evaluating quality metrics on historical data we have a unique opportunity for online evaluation. This means we get live feedback on recommendations. We currently store feedback from relationship managers on whether they think the client would buy the product. This feedback has two advantages. First, a positive response gives us more confidence in the recommendations. Second, we can use this explicit feedback to improve the training of the recommender models.
Improving Performance with Explicit Feedback
Using explicit feedback, we can begin to close our knowledge gap. As previously mentioned, we don’t know how to translate missing ratings. They could mean a client doesn’t like a product. Conversely, they could mean the client simple has no knowledge of the product. Explicit feedback gives us this information.
Although this feedback is also sparse, it leads to better recommendations. As it is accumulated over time, the quality of the recommender engine increases significantly.
Novelty versus Familiarity
Metrics such as the Area Under Curve are vital to gain confidence in the recommendations. As we train the algorithm using this measure we are not including bias for novel items. We address this by calculating the inverse popularity of an item as a proxy for novelty. Furthermore, we can combine this with the final ratings matrix delivered by the Collaborative Filtering algorithm to penalize popular items and reward novel items.
Outlook: Next Best Product in Portfolio Context
The recommender system also needs to consider the current portfolio of the client. This reflects the day-to-day concerns of a relationship manager when advising a client. For example, a key concern is to limit exposures to certain economic factors. Such factors could be a currency or an industrial sector. The ulimate goal is to arrive at well diversified portfolios that maximize a client utility with financial products and investments the client has an affinity to.
The question how we integrate a recommender system into a private banking advisory process considering the portfolio context of the client will be discussed in one our next blog posts.