PHI403 Lecture 18 — Risky Predictions

On the epistemology of prediction. Predictions are always fallible: Hume's problem of induction makes them logically invalid; Popper's falsificationism accepts them only provisionally.

Even with a complete data set, a causal claim goes beyond the data — this is what external validity requires but data alone cannot supply. Without a theory linking data points, we have no reason to assume a within-sample pattern transfers to new samples or to the population.

Closed systems — labs, models, ideal conditions — allow near-100 % predictions (lunar eclipses, pressure in a gas container). But stepping outside the closed system into real-world complexity, predictions become fallible again.

A central warning: adding more factors to a model trades internal for external validity — classic overfitting. More realistic-looking models do not automatically produce better predictions.

The lecture's positive proposal: tendency predictions. Instead of saying what will, would, or must happen (which fails if any factor is missing), predict what tends to happen given current knowledge. Tendencies are not guaranteed, but they are more than accidental — they reflect dispositions that the cause genuinely has.

This lecture is the most directly relevant one for ML prediction and predictive process monitoring: overfitting, external validity, and the framing of predictions as tendency claims all map onto standard ML concerns.

Connections

Back-link: sources/2023-anjum-rocca-phi403-causation-in-science. Concepts: concepts/dispositionalism · concepts/rct-limitations · concepts/probabilistic-causation · concepts/predictive-process-monitoring.