Prediction versus Explanation

The distinction between prediction and explanation is one of the foundational problems of Philosophy of Science. A predictive model outputs accurate forecasts about future or unobserved states of a system. An explanatory model says why those states arise — it identifies mechanisms, causes, or structural constraints that make the outcome intelligible rather than merely expected.

The distinction matters because prediction and explanation can come apart. A model that achieves high predictive accuracy on known data distributions — such as AlphaFold predicting protein structures from sequence databases — may do so through statistical correlation with no mechanistic content. Such a model does not explain why the correlation holds, and it will fail precisely where explanations are most needed: on novel inputs, under distributional shift, or where the causal structure changes.

The philosophical framework for this distinction was sharpened by Carl Hempel's Deductive-Nomological model (1948): genuine explanation is a deductive argument from laws plus initial conditions to the explanandum. On this view, prediction and explanation have the same logical structure — they differ only in epistemic context. Critics have challenged this symmetry: explanations require the cited regularities to be genuinely causal, not merely statistical, and they require the regularities to be non-accidentally true. A systems-level view adds a further constraint: explanation must be adequate to the system's level of organization, not merely its micro-level components. See also: Mechanism versus Statistics, Causality, Scientific Realism.

The Asymmetry in Scientific Practice

Despite formal philosophical efforts to equate prediction and explanation — most notably through the Deductive-Nomological model, which treats both as derivations from laws plus initial conditions — the two are asymmetric in practice in ways that matter for how science develops.

A prediction that fails is falsifying: it tells you the model is wrong, but not which assumption failed. An explanation that fails is diagnostic: a mechanistic model that predicts incorrectly can be interrogated — which mechanism was misspecified? Which parameter was out of range? — in ways that a correlation model cannot. A pure prediction engine, when it fails on an out-of-distribution case, offers no principled direction for improvement, because it has no mechanistic commitments to revise.

This asymmetry has concrete consequences for scientific fields that rely heavily on predictive models. Epidemiological contagion models, trained on past outbreak data, fail outside their training distribution in ways that are uninformative — the model fails, but the failure does not tell you which assumption about transmission dynamics was wrong, because the model has no explicit transmission dynamics. AlphaFold fails on intrinsically disordered proteins in ways that do not diagnose the underlying physics, because the model has no explicit physics.

The epistemological consequence: prediction engines tend to produce terminal knowledge — knowledge that ends inquiry rather than advancing it. When a field acquires a sufficiently accurate prediction engine, the incentive structure shifts away from mechanistic research. This is not a conspiracy; it is a consequence of how funding, attention, and prestige track measurable performance benchmarks. A benchmark measuring prediction accuracy does not and cannot measure explanatory depth. Optimizing for the benchmark optimizes away from explanation.

The Role of Training Distributions

A structural feature of statistical prediction models — neural networks, machine learning systems broadly — is that their predictive accuracy is relative to a training distribution. Predictions outside that distribution are not merely less accurate; they are unprincipled. The model has no basis for estimating its own uncertainty in genuinely novel regimes, because it has no model of what makes a case novel.

This creates a systematic failure mode: high predictive accuracy on in-distribution benchmarks is taken as evidence that the model 'understands' the phenomenon. But understanding — in the sense of having a model that transfers to novel conditions — requires explanatory content that the training distribution does not supply. The Protein Data Bank is a training distribution for protein structure prediction; the proteins that are biologically most important (IDPs, novel folds, evolutionarily distant sequences) are systematically underrepresented. Measuring prediction accuracy against the same distribution that generated the training data measures interpolation, not understanding.

The distinction between prediction and explanation is, in this sense, the distinction between interpolation and extrapolation — not in the geometric sense, but in the causal sense: does the model encode causal relationships that transfer when conditions change, or does it encode correlations that hold only within the distribution of conditions under which it was trained? A causal model can predict behavior under interventions; a correlation model cannot. The test of understanding is always: does the model remain accurate when the world changes in a way that breaks the training correlations?

Any field that cannot distinguish its prediction accuracies from its causal knowledge has not yet earned the right to claim it understands the systems it models. The benchmark is not understanding. The benchmark is evidence that more work remains.