Madecasse Group · federation of 8 companies · applied research (MSc)

FedCash — federated learning to forecast customer collections

Measuring when federated learning helps — or hurts — probabilistic forecasting of customer collections across several companies of a group, and how to weight and then adapt the shared model.

Research in progressOngoing

Context

Customer collections are the most uncertain component of short-term cash management. Companies within one group differ in data volume, customers and payment habits. They are also separate legal entities, responsible for their customers’ data: pooling detailed data cannot be taken for granted, and the information system keeps changing.

Problem

Learning together can help companies with little data, but it can also blur each company’s own habits (negative transfer). How much weight should each company carry in aggregation, and how far should the shared model then be readapted for each one, compared with a local or centralised model?

Approach

  1. A discrete-time hazard network estimates, for each outstanding invoice, the probability of being paid within 7, 14, 21 and 28 days; these are aggregated into a company-level forecast with an 80% interval.
  2. Local training, then federated averaging of parameters (FedAvg): only parameters, sample counts and aggregated losses leave a company, never its invoices.
  3. Factorial design: local, centralised or federated training; volume or uniform weighting; local adaptation, from simple recalibration to the full network; simple baselines and gradient boosting, under an equal tuning budget.
  4. Hypotheses, metrics (log-loss, calibration, interval coverage) and decision criteria pre-registered before any performance is measured: no gain would be a valid result.

Stack

  • Python
  • PyTorch
  • FedAvg
  • Discrete-time survival
  • LightGBM (baseline)
  • Calibration
  • Conformal prediction

Status

Research in progress

Protocol pre-registered before any performance is measured; no results are published at this stage.

What I learned

The most useful work happens before the first model: fix the question, the metrics and the decision criteria, then seal them. And within a group, heterogeneity between companies is not noise to smooth out — it is the very object of study.