Andersen's team was approached by a retail bank serving over 10 million clients. All of them regularly interact with the bank and use its services. To improve business efficiency and make the most of this client base, the customer wanted to obtain a tool capable of identifying clients with a high propensity to accept cross-sale offers. Also, they requested the capability to determine the best possible personalized approach for each client. Correspondingly, a 360-degree view solution was needed.
Andersen's customer is now running a unified 360-degree client view database. Each of the profiles it stores contains a history of interactions, a history of offers, and calculated NPVs for each product and channel. The offers are assigned on a black-box basis, with test and control groups. While participating in this initiative, our team wasn't aware of the exact logic applied to assigning test and control groups or offers. This situation is quite typical because the offers can be managed by different departments, teams, and systems – both manually and on the basis of A/B tests.
When working on the project, we focused on:
MongoDB, PostgreSQL, Couchbase
Apache Spark, SAS
To start with, our specialists carefully reviewed and analyzed the customer’s current state of affairs and their tech capabilities against the backdrop of their requirements. Andersen's team also did extensive market research. After that, we outlined the main idea to improve the customer’s system; particularly, we decided to migrate it to a data lakehouse based on open-source technologies.
The logic used for test and control assignments was unknown to us. More precisely, there was no guarantee that those assignments were not proposed at random. Therefore, any performance evaluation of the customer’s model had to be based solely on historical data. This data might have reflected arbitrary biases.
In an ideal case, we would evaluate every model version via randomized experiments. However, such testing is generally expensive and risky from the customer experience standpoint. Therefore, we had to develop an offline evaluation methodology that would reduce the number of online trials and mitigate the risk of adopting catastrophic policies.
Andersen's team generated a procedure for bias control and correction to address potential biases in the historical data. We built a core model focused on the product and channel balance. Its mission was to estimate the probability of sales, resulting from different types of offers. This further helped determine the optimal next-best offer strategy for each client, product, and channel.
Based on probability scores (i.e. the prescriptive part of the next best offer model), we implemented the offer type selection logic. Andersen's specialists developed an evaluation procedure to assess the quality of the model using the bias-corrected data.
As a result of our work, a single source of correct information allowed the customer to calculate and improve a definitive loan-to-value. All that data is accessible within one central repository, which reduces the time and cost needed to integrate it.
The contribution of our analysts consisted of:
What happens next?
An expert contacts you after having analyzed your requirements;
If needed, we sign an NDA to ensure the highest privacy level;
We submit a comprehensive project proposal with estimates, timelines, CVs, etc.
Customers who trust us