- Silo (cross-silo federated learning)
- An organisation that keeps its data on its own premises. “Cross-silo” federated learning brings together a few organisations (hospitals, banks, companies), each holding a lot of data, as opposed to the “cross-device” setting (millions of phones). Here, each group company is a silo.
- Heterogeneity (non-IID)
- The participants’ data do not follow the same distribution: volumes, customers and payment habits differ from one company to another. This is the central difficulty of federated learning.
- Federated averaging (FedAvg)
- In each round, every silo fits the shared model on its own data; the coordinator takes the weighted average of the parameters received (weights ws ∝ nsα) and sends it back to all. The data never leave their silo.
- Local adaptation (personalisation)
- Fitting the shared model to a single company’s data, from lightest to strongest: recalibration of probabilities (A0), relearning the last layer (A1), careful fine-tuning of the whole network (A2).
- Negative transfer
- The case where learning with the other companies yields worse forecasts than learning alone, with one’s own local model.
- Discrete hazard
- The probability that an invoice is paid during a given week, given that it has not been paid before. Chained together, weekly hazards give a cumulative probability of payment that can only increase with the horizon.
- Log-loss (cross-entropy)
- An error score for probabilistic forecasts: the lower it is, the better the forecasts. It mainly penalises overconfidence and cannot be gamed: it is a proper scoring rule.
- Calibration
- Agreement between stated probabilities and observed frequencies: invoices forecast as “70% likely to be paid” should be paid about 7 times out of 10. The calibration slope equals 1 when probabilities have the right spread; below 1, the model is overconfident.
- ECE (expected calibration error)
- The average gap between stated probability and observed frequency, computed over equal-count bins of forecasts; 0.03 corresponds to 3 percentage points.
- Interval coverage
- The share of cases in which the amount actually collected falls within the stated interval. An “80%” interval should contain the outcome about 8 weeks out of 10, while staying as narrow as possible.
- Equivalence margin ε
- The smallest difference deemed useful in practice, set in advance. Below it, two methods are considered equivalent, even if a tiny difference exists.
- Pre-registration
- A dated freeze of the protocol, metrics, decision criteria and evaluation code before any performance is measured, evidenced by a Git tag and a SHA-256 hash: the rules can no longer be adjusted after seeing the results.