Random Utility Models: MIT Power of Three improves predictions

MIT researchers on June 11, 2026 published an upgrade to random utility models that models each alternative's utility as the average of three independent draws rather than a single draw. The modification, dubbed the "Power of Three," produces consistently better prediction and calibration on standard discrete-choice datasets used across economics, transportation, and marketing.

The change targets a nearly century-old framework for modeling preferences, where a decision maker is assumed to associate a random utility with each option and then choose the maximum. The Power of Three keeps that maximum-choice architecture but replaces the single random draw per alternative with three independent draws whose mean is used as the option's utility. The researchers present mathematical arguments and empirical tests showing the new specification alters the distributional tails and correlation structure in ways that improve rank-order predictions.

How the "Power of Three" works

At its core, the Power of Three modifies the stochastic component of a standard random utility model. Instead of U_i = V_i + e_i, where e_i is one random shock per alternative, the new formulation sets U_i = V_i + (e_{i1} + e_{i2} + e_{i3})/3. That averaging reduces variance and changes the effective shape of the shock distribution, while preserving independence across alternatives when the base shocks are independent.

The paper shows that three draws is a practical sweet spot: one draw is the classical approach and many draws converge back toward a deterministic limit; three draws produce meaningful changes to choice probabilities without requiring large computational overhead. The authors explore identification, likelihood computation, and maximum-likelihood or simulation-based estimation routines compatible with standard discrete-choice toolchains.

Results and tests

The team evaluated the Power of Three on a mix of simulated and real-world datasets including consumer product choices, transportation route selection, and voting-style choice problems. Across those datasets the three-draw model improved out-of-sample accuracy and calibration metrics relative to single-draw baseline RUMs. Improvements were most pronounced on datasets where alternatives had closely clustered mean utilities and decisions were sensitive to tail behavior.

Computational cost increased modestly because each alternative requires three draws per simulated evaluation, but the authors note that estimation scales well with modern simulation and parallelization techniques. They also tested variants (two draws, four draws) and reported that two-draw variants yielded smaller gains while larger numbers of draws diminished practical benefits.

The paper discusses limitations and diagnostic checks. The Power of Three changes implied substitution patterns and can interact with common identification strategies. Where data are sparse or the deterministic utility V_i dominates stochastic noise, the practical gains are smaller. The authors provide guidance on model selection, cross-validation, and when the three-draw form is likely to help.

Why it matters

The Power of Three is a modest modification to a foundational model that yields consistent predictive gains with limited added complexity, making it accessible for applied researchers and practitioners. For anyone building discrete-choice forecasts—transport planners, product managers, policy analysts—a better-fitting stochastic specification can improve scenario forecasts and policy counterfactuals without overhauling established workflows.