One of the ten largest US banks
For the re-launch of their mobile app, this bank wanted to be able to predict when retail account balances are likely to go below a certain threshold. This would allow to alert customers who are likely to overdraft their account before they incur late fees.
Challenge
Multi-day time series forecasting is hard — small errors accumulate rapidly. Different transaction patterns require different modeling techniques, and reconstructing a balance as displayed at an ATM is surprisingly non-trivial: the exact definition of balance types, batch job timing, and transaction cut-offs all matter. On top of the modeling challenge, uncertainty in the forecast needs to be quantified and communicated clearly in the app’s user interface.
Solution
We did extensive data exploration, building interactive visualizations for millions of data points. We created novel account segmentation workflows using transaction-level data with dimensionality reduction and clustering algorithms. We then built both tree-based and sequence-to-sequence neural network models predicting end-of-day ledger balances, with extensive documentation for internal model validation teams.
Results
The feature is deployed in the production app for millions of retail consumer accounts. The account segmentation derived from transaction-level data also proved valuable in other contexts: detection of suspicious behavior for anti-money-laundering efforts and identification of consumer accounts used for commercial purposes.