Where does prediction collide?

A model's estimate of an outcome for a case it has not seen, judged by how close that estimate lands to the truth out of sample. In To Explain or to Predict? Shmueli set this against explanation: an explanatory model is built to test a causal account and is judged on fit and on the standing of its coefficients, while a predictive model is built to guess new values and is judged on error on held-out data. The referent is a fitted model applied to fresh input, and the unit is a single estimated value, often a number or a class label. It differs from the neuroscience meaning because nothing here is a signal inside a brain, and from the philosophy meaning because the object is one case, not a theory. In statistics the word is even set against explain, so the internal contrast is built into the field.[1]

Unit

A model's out-of-sample estimate for one case

Recovery means

Not applicable

Measured by

Error on held-out or future data, such as mean squared error or misclassification rate

A signal inside the nervous system: the mismatch between what a level of cortex expected and what it actually received, passed upward so that higher levels can revise their guess. Rao and Ballard proposed in 1999 that visual cortex works this way, with feedback carrying predictions down and only the residual error travelling up, which reproduced several extra-classical receptive-field effects. Friston later folded this into the free-energy account, in which the brain acts to minimise prediction error over time. The referent is neural activity, and the unit is a prediction-error signal in a population of cells, not an estimate a statistician reports. It shares the word with the statistics meaning but names something the brain does automatically, and it differs from the philosophy meaning because no theory is being tested, only sensory input reconciled.[2][3]

Unit

A prediction-error signal in neural populations

Recovery means

Not applicable

Measured by

Neural or computational estimates of the expected-minus-actual signal, from recordings or model fits

A statement a theory forces about a future observation, put forward as something the theory could fail. In Popper's account a theory earns its standing by ruling things out: it must forbid some observable outcome, so that a prediction which could have gone the other way and did not is what supports it. The referent is a theory under test, and the unit is a single forecast that risks the theory, valued most when it is novel, meaning it was not used to build the theory in the first place. This is the root of the prediction-versus-accommodation debate, where a fact predicted ahead of time is treated as stronger evidence than the same fact fitted afterward. It differs from the statistics meaning because what is at stake is a general theory, not one case, and from the neuroscience meaning because the claim lives in an argument, not in tissue.[4]

Unit

A risky forecast that a theory could fail

Recovery means

Not applicable

Measured by

Whether a forecast the theory forbids is observed or not, assessed against the record