Skip to content
Coordinates of ThoughtHome of Philosophy
Join
Question · Explain

Friedman’s billiard player: is predictive success really all a model owes us?

Wanjiru Njenga Fellow
· 4 replies
Economy & workAmericanAnalyticMilton FriedmanNancy Cartwright#as-if-assumptions#economic-models#instrumentalism#prediction
Follow

My models predict, quite well, who will repay a small loan, using patterns in phone and payment records. They do not describe why anyone repays. Some of the strongest predictors are things that have no plausible causal story at all.

A colleague told me that Milton Friedman defended exactly this kind of modelling in a famous 1953 essay, "The Methodology of Positive Economics". His example was an expert billiard player: you can predict his shots very well by assuming he calculates the optimal angles with complex mathematics, even though he does nothing of the sort. The assumption is false; the predictions are good; that, Friedman says, is all that matters.

It's a comforting argument for someone in my job. But I've seen models like mine fail suddenly when the world changes, and fail without any warning, because nobody knew what they were tracking.

Could someone explain Friedman's argument properly, and the main philosophical objections to it? I'd like to know whether I should feel comforted or worried.

LikeAnswer
Follow4 answers

Join the conversation: members reply here, and joining is free. Reading is always free.

Join free to reply Sign in

4 answers

Accepted answer

  1. Javier Ferrandis

    ✓ Accepted answer Fellow

    Some of each, I think.

    Friedman's claim is that the realism of a theory's assumptions is irrelevant to its assessment; the test is whether its predictions are accurate for the class of phenomena it is meant to explain. His examples were the billiard player and leaves on a tree, which are distributed as if each leaf sought to maximise its sunlight. Firms, he argued, behave as if they maximise profit, because firms that don't tend not to survive.

    Three main objections:

    1. Prediction isn't all we want. Economics is asked to explain and to guide policy. A model that predicts without describing a mechanism gives no warning when conditions change, which is exactly the failure you've seen.

    2. The Lucas critique. Robert Lucas argued in 1976 that the relationships in a model can shift when policy changes, because people respond to the policy. A purely predictive model, estimated on past data, can't anticipate that.

    3. Capacities. Nancy Cartwright argues that models work by isolating real tendencies or capacities, not by being convenient fictions. If that's right, unrealistic assumptions are fine only when they isolate something real. Your strongest predictors with no causal story are the worrying kind.

    Samuelson mockingly called Friedman's position the "F-twist", yet the essay is still much discussed.

    Helpful · 3
    1. Wanjiru Njenga

      Fellow

      Javier, the Lucas point lands hard. When we change our lending rules, borrowers change how they use their phones, partly because word gets around about what the model rewards. Our model predicts the old world and creates a new one. So "as if" lasts exactly until people learn what you're assuming.

Other answers

  1. Emeka Anyaoha

    Fellow

    From philosophy of science: Friedman is often called an instrumentalist, someone who thinks theories are just tools for prediction. But the billiard example only works because there is a real mechanism (skill, practice, selection) that makes the as-if story track something. That's closer to Cartwright than to pure instrumentalism. The as-if works when there's a hidden reason it works. Your mystery predictors may have no such reason, and that's the difference between a good idealisation and a lucky correlation.

    Helpful · 1
  2. Nur Aisyah Rahman

    Fellow

    And a fairness note from someone who audits models: a predictor without a mechanism often encodes history. If a phone pattern correlates with repayment because it correlates with wealth, and wealth correlates with ethnicity or region, the model reproduces past exclusion while looking neutral. Friedman's essay didn't have to think about that. Your job does.

    Helpful · 2