Approach
Models earn trust out of sample.
Tool selection
The data decides the tool.
The first question is never which model; it is what the data permits. The ratio of history to forecast horizon gates everything. With rich history and a short horizon, machine learning earns its keep. With thin history and a long horizon, no model can outrun arithmetic, and the honest deliverable is scenario envelopes, not point forecasts.
We say so when a forecast is not achievable. Declining the impossible job is part of the service.
Validation
Baselines before models.
Every engagement starts with the embarrassing baselines: persistence (tomorrow equals today), seasonal repeats, linear models. A model ships only when it beats them on data it has never seen, walk-forward, the way it will actually be used.
A vendor who never shows you the naive baseline is hiding something.
Uncertainty
Uncertainty is the deliverable.
A forecast is a distribution, not a number. We ship calibrated bands — and verify the calibration: when the model says 90%, reality falls inside 90% of the time, measured on held-out data.
Simulation
Simulation for questions data cannot answer.
Forecasting models learn what has been. For counterfactuals, stress scenarios, and structural what-ifs — situations with no precedent in the data — we build simulations: system-dynamics and agent-based models whose value is exploring the space of what could happen.
Language models
Where language models fit.
Language models are excellent at language — reading documents, turning text into structured data, making a model's outputs legible. In our systems they handle words at the boundaries. Every number is produced by a model that can be checked against reality: a forecast that backtests, a simulation that is audited, arithmetic that is exact.
Craft
Built one-of-one.
Every engagement ships a model designed from scratch for one client and one system — never a re-skinned product. The model is tested out-of-sample, against data it has never seen, before it is allowed anywhere near a decision.
The firm is small by design: the person who scopes the model is the person who builds it.