One of the ten largest US banks

Predicting time series for multiple days is an exceedingly hard problem. Small errors accumulate rapidly. If several days need to be forecasted, doing this for all of them at once might be preferable to propagating a model of one time step forward. Different types of transaction patterns might need different modeling techniques - dormant accounts, for example, likely won't have changes. Reconstructing a balance as it would be displayed at an ATM could be surprisingly non-trivial. You might need to understand the intricacies of the whole system to get this value: the exact definition of different types of balances, when batch jobs run and what the cut-off is for which kinds of transactions, what time the time stamps actually record. How do you quantify uncertainty in your forecast if you use neural nets, and how do you design a user interface to properly convey this concept in the first place? When and how should users be notified?

We did extensive data exploration, looking at both individual balance time series, and building interactive visualizations for millions of data points. We created novel account segmentation workflows, using transaction-level data and state-of-the-art dimensionality reduction and clustering algorithms.  We then built both tree-based and seq2seq-type neural network based models predicting the end-of-day ledger balance using an in-house cluster.  We created extensive documentation of relevant model metrics for internal model validation teams.

The new feature is deployed in the production app for millions of retail consumer accounts. Furthermore, the account segmentation derived from transaction-level data could be used in other contexts: detection of suspicious behavior for anti money laundering efforts, and mitigation of the use of consumer accounts for commercial purposes.