These notes show how we work, on public data, with the sources named. Four are worked studies: intermittent demand forecasting on 2,404 spare-part series, a credit scoring model validation report, small area estimation checked against the French census, and time series cross-validation measured on a simulation. Four describe our methods: how we treat data, how we validate a result, how we detect overfitting, and our record on public benchmarks, including the results that came out empty.
Intermittent demand forecasting, measured
We forecast 2,404 public spare-part series with seven methods, scored them once on a sealed six-month window, and found that the choice of metric moved the ranking more than the choice of method.
Credit model validation, a worked report
A complete validation report for a credit scoring model on public data: benchmark, sealed holdout, discrimination with intervals, calibration, stability, segments and a placebo.
Small area estimation, checked against a census
We drew 300 simulated surveys from the 2023 census of 3,781 communes in Hauts-de-France, estimated each commune's unemployment rate, and checked every margin of error against the census value.
Time series cross-validation, measured
Random k-fold, walk-forward and purged cross-validation with an embargo, compared on a simulation where the true skill is known, so the optimism of each split can be measured.
Our gates
A claim reaches a client only after it survives a fixed sequence of tests, and the sequence is written in code, not in a document.
How we treat data
A result is only as good as the data under it, so we check where every number comes from, when it was known, and whether a second source agrees.
Detecting overfitting
Since June 2026 we have spent much of our time catching results that looked good only because we had searched for them; these are the cases, with their numbers.
Our benchmark record
What our models scored on public benchmarks since September 2026, how far each number can be trusted, and the ones that turned out empty.