Skip to content

Bayesian Regret Minimization

Why It Matters

Outcome regret makes you regret good decisions that happened to fail - and prevents learning. You abandon strategies that were correct but unlucky. Bayesian regret minimization separates outcome luck from process quality. It asks: "Was my decision-making process flawed, or did a low-probability outcome occur?" This enables learning from genuine errors (miscalibrated probabilities, wrong values) without being misled by outcome noise.

The Key Move

After a decision with known outcome, resist immediate emotional reaction (regret or satisfaction). Instead, decompose:

  1. Remind yourself what probabilities and utilities you assigned ex-ante.
  2. Ask: "Was my probability estimate wrong?" Check if evidence you had predicted differently than you estimated.
  3. Ask: "Was my utility estimate wrong?" Check if outcomes feel more/less valuable than you'd predicted.
  4. If probabilities and utilities were right but outcome was bad, you experienced correct-decision-bad-outcome - no regret.
  5. If estimates were wrong, you have learnable regret - update your models.

Classic Application

Poker hand analysis. You bet with 70% probability of winning (correct estimate given cards you could see). You lost (30% outcome). Outcome regret says: "I shouldn't have bet." Bayesian regret analysis asks: "Was 70% the right estimate?" Review: given what you could see, yes. "Would I make the same bet again?" Yes - 70% is good odds. Conclusion: no regret, just variance. Compare to: you bluffed with 20% probability of success (incorrect estimate - you didn't account for opponent's tells). You lost. Bayesian analysis: probability estimate was wrong, bet was bad. Update: incorporate opponent tells in future estimates.

Surprising Application

Career path decisions looking back. You chose a stable corporate job over a risky startup (90% probability of stability vs. 40% for startup). The startup became a unicorn (rare outcome). Outcome regret: "I missed out on millions." Bayesian regret analysis: "Given information at decision time, was my 40% estimate wrong?" If no (startups really do mostly fail, this one was ex-ante indistinguishable), then no regret - you made the right decision and got unlucky. If yes (you ignored strong signals this startup was exceptional), then update your startup evaluation process. This prevents either excessive risk-taking (chasing rare wins) or excessive risk-aversion (regretting correct decisions).

Failure Modes

Hindsight bias - "of course I should have known" distorts memory of what you believed ex-ante. Probability creep - retroactively claiming you "always knew" the unlikely outcome was more likely. Cherry-picking outcomes - only analyzing decisions with bad outcomes, ignoring lucky bad decisions with good outcomes. Paralysis from process regret - even good processes have mistakes; the goal is fewer errors, not zero.

Go Deeper

Savage, "The Theory of Statistical Decision" (1951) on regret functions; Tetlock & Gardner, Superforecasting, Chapter 3 on distinguishing outcome from process.