The concept, its research origins and what the term means in practice today.
Elizabeth Sramek
Published 7 July 2026
The phrase describes a system that does what a statistician does on first contact with a dataset: propose a structure, fit it, check it, revise it, and then explain what was found. The distinguishing feature is the last step. A system that returns a well-tuned model is an automated modeller. A system that returns a description you can argue with is something else.
The original research came out of the University of Cambridge, with collaborators at MIT, and framed the problem as search over a grammar of models. Because Gaussian process kernels can be added and multiplied, a small set of building blocks generates an open-ended space of structures — and every structure in that space decomposes into parts with names. The research page covers the papers and credits their authors.
A language of models expressive enough to capture real phenomena. A search procedure to explore it. A principled evaluation criterion. A translation step that turns structure into prose, figures and limitations.
Not “the model achieved 0.94”, but “the series has a smooth upward trend of about 2.7% per year, an annual cycle whose amplitude is growing with the level, and residual variation of about 3%; the model fits poorly in 2008, where behaviour appears to change.” The airline example shows the full form.
Loosely. Most tools marketed as automatic statisticians or AI analysts automate profiling and model fitting and then generate commentary with a language model. That commentary is not derived from a fitted structure, so it cannot be checked against one. The distinction is worth insisting on.
No. AutoML optimises predictive performance over a model space. An automatic statistician optimises for an interpretable description, which sometimes means preferring a slightly worse-fitting but structurally clearer model.
It removes the mechanical part of the work and produces a draft that a person still has to criticise. The framing of the question and the judgement about whether the answer is plausible remain human.