Data

Modelling & Machine Learning

Models built to be checked, not just fitted.

Prediction and inference are different problems, and conflating them is how models end up confidently wrong. DeepTeal’s modelling work is grounded in formal statistical training, which mostly shows up as scepticism: about leakage, about validation that flatters the model, and about results that look too good to be true.

Two models fitted to the same observations. The dark one is flexible enough to pass through every single point, so its error on the data it was given is exactly zero. The hollow points arrive afterwards, and neither model has seen them. The perfect score does not survive them; the duller one barely moves. That is the whole reason a model is judged on data it has never met.

Areas of work

Forecasting and predictive models Demand and revenue forecasting, churn and retention models, segmentation and propensity scoring, with validation designed to survive contact with new data.
Experiment and study design A/B and multi-variant tests, and observational designs for the cases where a controlled test is not possible. Sample size and power, what is being controlled for, how multiple comparisons are handled, and what the result can and cannot be used to claim.
Model review Assessing existing models for validity, drift and misuse, including cases where the honest conclusion is that a simpler method would do the job.

If this is close to something you are considering, a short description of the problem and the systems already involved is the most useful place to start. Get in touch.