The Next Best Action for Banks - Chapter 3: Formulating Recommendations
In this article
A growing number of organizations are leveraging data science to increase user interaction and enrich user experience. Nowadays a wealth of data about individual customers is collected, including their preferences, online activities, historical purchases and more. By studying each customer's behavior, companies can increase the opportunities to serve them.
In the third article of our series aimed at helping banking and financial services organizations better understand how they can leverage data science to become more customer-centric, we will look at Next Best Action (NBA) Recommendation Systems and approaches to financial recommendation modeling.
The term "Next Best Action" (NBA) refers to using AI and machine learning algorithms to evaluate different communication channel messages, based on the susceptibility and response patterns of customers. Companies can then use these evaluations to determine which actions are most likely to materialize in a conversion.
NBAs are dynamic by nature; the environmental conditions and each customer's journey are constantly evolving. Therefore, NBA decisions must adapt to new environments to improve behavioral targeting. The underlying system must recommend the optimal channels (email, text, online offers, direct mail, voice mail, personal phone calls, or social media) and deliver engaging messages for customers at just the right time.
The success of NBAs is typically measured with metrics such as conversion rate or click-through rate, along with an applied methodology like A/B testing, which can be employed to establish confidence about the true impact of the measures. NBA recommendation systems have the potential to change the way organizations communicate with users and allow companies to maximize ROI.
There are several areas of opportunity for banking and financial services organizations to utilize and benefit from by using the NBA approach:
- Attract new customers – Once customer segments are determined and those individuals most likely to become customers are identified, these segments can be used to generate specific product recommendations. Considerations regarding a customer's stage of financial maturity can help define their financial goals and offer products best suited for them.
- Cross-sell/Up-sell opportunities – Additional product offerings or services, such as credit cards, checking or savings accounts, investment accounts, loans/lines of credit, financial advisory services, estate planning and tax planning can be recommended throughout a customer's lifecycle.
- Customer satisfaction / Churn reduction – Surveys can be conducted to measure customer satisfaction and then analyzed to identify areas of improvement for the institution. Preventive measures, such as direct deposit or automatic bill pay, can be implemented to mitigate the risk of an over-drafted account, or missed bill payment. Customers can also participate in a reward or loyalty program for further engagement. The ability to detect early signals of potential churn can initiate proactive engagements with customers and the appropriate retention offers.
NBA systems may make usage of various types of data, including:
- User profiles and identities
- User transactions
- User activity and interactions
- Product/service features
- Contextual information
User profiles and identities are helpful sources of information when it comes to understanding user behavior. Profile-based features, such as location, age group, gender, income bracket, credit score, wealth and housing information, can be crucial for recommending relevant services to users. Data sharing with other companies can provide broader insights into a customer's preferences. One caveat is product offerings that must follow equal-opportunity constraints or the usage of protected category information (e.g., age, gender, ethnicity) need to be carefully vetted to ensure compliance with established regulations.
User transactions are also rich sources of information for use in NBA systems. This would encompass various data sources such as credit card transactions, bank account line items, credit scores, credit utilization, etc. Transactional data typically comes with many detailed attributes, such as timestamps, amount, MCC (Merchant Category Code) and whether a card is present or not present. Additional feature variables can be created to reflect concepts such as transaction recency, frequency, intensity, and concentration.
User interaction with the financial organization is usually available, yet this data set also tends to be underutilized. This is due to the large variety of touchpoints of a user with an organization through physical branch offices, call centers, sales representatives, searches and item clicks, online banking, e-mail, SMS text and phone campaigns. Such activity-related information is owned by different business verticals requiring access permissions and domain knowledge. When combined with users' financial transactions, this activity-related information can be extremely valuable.
Attribute information for financial services such as duration, risk profile, historic monthly/yearly/quarterly return, mode of investment, yield and transferability can help identify effective cross sell and upsell opportunities for a certain stage in the financial planning lifecycle. Together with user profile features, an advanced rule-based recommendation engine can also effectively kickstart targeted selling and marketing of financial services.
These can be either location- or time-dependent features created from the previous sources of data. These features act as an explanation to the recommendation, "you might be interested in Financial service ABC," which is crucial for taking the recommendation from a black-box model output to a user-relevant model insight. Contextual information may also include GPS or WiFi location data.
Besides the data input for the recommendation engine model, it is also important to set up and prepare a validation database for effectively testing the user responses to recommendations. This will help to fine-tune and refine the models according to the changing importance of different features.
The modeling of financial recommendations can be approached from different angles, depending on the nature of available data. These approaches include:
- Customer segmentation
- Regression methods
- Recommendation algorithms
- Performance tracking: A/B testing
Through the profiling of customer attributes, segments can be established to understand the users better, enabling financial companies to make strategic decisions. These decisions can lead to the crafting of specific messages delivered through optimum channels to reach the customer. This approach is often adopted when there is no historical data on the performance of the recommendations.
Data scientists typically apply unsupervised machine learning algorithms for the clustering of customer data. Traditional methods include K-means clustering, K-nearest neighbors, Hierarchical Clustering and Density-Based Spatial Clustering (DBSCAN). Modern approaches may include the usage of autoencoders from Deep Learning for dimensional reduction and feature generation.
If the historical performance (e.g., success vs. failure) is tracked, then regression/classification algorithms such as logistic regression, random forest, gradient boosting machines and artificial neural networks can be used to automate the recommendations. This class of algorithms is done through a type of machine learning known as "supervised learning" because the model is trained and optimized by using the historical success/failure labels as guidance.
Machine learning recommendation engine algorithms usually are categorized into three classes: collaborative filtering models, content-based models and hybrid filtering models.
- Collaborative filtering models assume customers with similar profiles (e.g., credit score) or behaviors (e.g., consuming behavior) will have similar preferences on financial products. The simplest ones are based on matrix factorization techniques; the goal is to learn the low-dimensional vectors for all customers and products so companies can view how much a product has a certain feature and how much a user likes this feature in products. The factorization can be trained by applying SGD (Stochastic Gradient Descent) to the SVD (Singular Value Decomposition) algorithm.
- Content-based models assume that if a customer likes a financial product, the customer will also like similar financial products. Content-based models solve a "cold start" problem on products. While these models don't require customer feedback data, new or relevant products or services can be identified and recommended to the customer. Some pattern recognition techniques in machine learning have unlocked great improvements in the content-based model.
- Hybrid filtering models ensemble both collaborative filtering and content-filtering strategies together for better recommendation performance. These strategies are usually trained through deep learning or reinforcement learning, allowing companies to realize much finer interactions between customers and products.
Third-party recommendation engine providers include:
- AWS Personalize - AWS Personalize is a leader and proven player among recommendation engine vendors. It is fairly advanced and allows its users to personalize recommendations by choosing between different recipes
- GCP Recommender - GCP Recommender offers strong recommendation engines that are flexible and not tailored to any specific industry
- Microsoft Azure ML - Microsoft is one of the major cloud providers and its Azure ML is suitable for large-sized companies. It supplies customized recommendation algorithms (i.e., Content-based filtering/collaborative filtering/hybrid)
- Salesforce Evergage - Evergage has a superior user interface, which makes it easy for users to implement features such as personalization recipes. The recipes allow marketers to mix and manage multiple algorithms.
- Adobe Target - Adobe Target has a user interface that makes it easy to compare recommendation designs. It supports both AI and logic-based recommendations.
- Survival models for churn – To avoid losing customers, a financial institution needs to understand why its ex-customers have left and then predict the likelihood of churn of its current customers. A survival-analysis model works well to deal with censorship in data (i.e., data sets cut off from the left or the right in a timeline), which is why companies usually implement survival models instead of other regression models for churn analysis.
The Cox Proportional Hazards model is often used to identify the factors associated with customers who churned and predict the likelihood of churn events. One can consider the "start event" as a customer establishing a relationship with a company (i.e., opening a credit card) and the "end event" as a customer ending a relationship with a company (i.e., closing a credit card). An evaluation of Recency, Frequency and Monetary characteristics, also known as an RFM analysis, can help both in the initial exploration as well as in the feature engineering stage.
A/B testing, sometimes called split testing, refers to a randomized experiment measuring the efficacy of business strategies by comparing two versions of a testing target. A/B testing allows a company to judge the result between the original and a variation. This strategy plays an important role in decision-making by evaluating if a company saw better results from using a recommendation engine or through other marketing initiatives.
Here, we focus on discussing how A/B testing helps to evaluate the recommendation engine performance. A/B will measure the performance of a recommendation engine based on whether the sales increased, usages increased, customer experience improved or a reduction in churn rate occurred.
When utilized as a systematic, long-term tactic, A/B testing can help to improve the recommendation algorithms as well. However, the exploration-exploitation tradeoff dilemma must be considered. A formalized approach to leveraging successive A/B tests is known as the "Multi-Armed Bandit" algorithm. The convergence of Multi-Armed Bandit algorithms is typically slow. Sometimes, instead of learning from the real world, computer-simulated virtual "agents" are used to accelerate the model training, in an approach known as the modern "Reinforcement Learning." Reinforcement Learning has proven to be very successful in some areas, such as autonomous driving, computer games and industrial applications. Its ability to emulate human consumers is still in the early stages, as human behaviors are certainly much more complex and tend not to follow rigid rules like in the cases of physical or game simulations.
Next Best Actions have become imperative for all institutions with large customer bases. WWT stands in a unique position to support initiatives that enable results, from Enterprise Architecture to Data Strategy and Data Science. Our mission is to serve as our clients' Trusted Advisor in any digital initiative that leads to optimal operations and satisfaction of end customers. Technology never ceases to make advances, but our clients can rest assured that they will always be at the forefront of innovation.