Interpretable forecasting for major e-commerce checkout solution

By breaking down a complex time series-based credit-risk challenge into distinct ML subproblems, Modulai produced robust, interpretable forecasts for precise risk assessment.

Background
Forecasting future balances is highly valuable for financial institutions managing customer payments. Insight into how a customer’s balance is likely to evolve allows proactive action rather than reactive intervention, which is critical for effective risk management.

At the same time, the task is far from trivial. Financial time series are often noisy, volatile, and influenced by irregular events, making reliable forecasting difficult. At the same time, it is also important to understand what drives the predictions instead of relying on a black box system. By encoding prior knowledge into the modeling framework, the complexity of the problem can be reduced while maintaining strong predictive performance.

Solution
At a high level, the solution combines four machine learning models. First, two separate models predict the total incoming and outgoing amounts over the forecasting period without assigning transactions to specific days. Next, two additional models forecast how these volumes should be distributed across days within the forecasting period, creating transaction time series. Finally, these transaction forecasts are combined with the company’s balance policies to compute predicted customer balances over time.

This modular approach constrains the final balance forecast by total transaction volumes, improving accuracy while maintaining interpretability. Each step can be traced back to individual models, making the results easier to explain, validate, and act upon in a risk management context.

Outcome
Using this approach, Modulai’s client was able to identify high risk customers with high accuracy. Although only about 0.5 percent of customers were classified as high risk, 50 percent of them were identified at a 75 percent precision threshold. This will enable the client to take more accurate precautionary actions and ultimately provide better service for the vast majority of customers who are not considered “risky”.