Please rotate your device to landscape mode to view the charts.

Predicting and Understanding Initial Play

Journal: American Economic Review

Date: 20191201

Author: Fudenberg, Drew; Liang, Annie

Abstract:
We use machine learning to uncover regularities in the initial play of matrix games. We first train a prediction algorithm on data from past experiments. Examining the games where our algorithm predicts correctly, but existing economic models don't, leads us to add a parameter to the best performing model that improves predictive accuracy. We then observe play in a collection of new "algorithmically generated" games, and learn that we can obtain even better predictions with a hybrid model that uses a decision tree to decide game-by-game which of two economic models to use for prediction.

Link: Google Scholar

Key Findings

Level-1(α) Model Improves Predictions

Adding a risk aversion parameter α to the level-1 model significantly improves prediction accuracy of modal actions in experimental games from 72% to 79%

Performance Varies by Game Type

The level-1(α) model performs exceptionally well on random games (92% accuracy) but less well on algorithmically designed games (38% accuracy)

Hybrid Model Outperforms Base Models

A hybrid model combining level-1(α) with Pareto-dominant Nash equilibrium (PDNE) achieves 79% accuracy across all games, better than either model alone

Model Performance Comparison

  • Level-1(α) achieves 69% completeness compared to 58% for basic Level-1
  • The hybrid Level-1(α) + PDNE model reaches 73% completeness on lab games
  • Random guessing baseline achieves 33% accuracy

Performance Across Game Types

  • Level-1(α) achieves highest accuracy (92%) on random games
  • Performance drops to 38% on algorithmically designed games
  • Lab games show intermediate performance at 79% accuracy

Hybrid Model Decision Process

  • Model first checks if level-1 action is part of Pareto-dominant NE
  • Then evaluates presence of symmetric NE with high payoffs
  • Finally considers whether action maximizes both players' payoffs

Contribution and Implications

  • Demonstrates how machine learning can improve economic models through systematic identification of patterns in experimental data
  • Shows the value of combining behavioral models with algorithmic approaches for better predictive accuracy
  • Provides a methodological framework for developing hybrid prediction models that maintain interpretability

Data Sources

  • Model Comparison Chart: Based on Table 7 showing accuracy comparisons between prediction models
  • Game Type Performance Chart: Constructed using data from Tables 2, 4 and 5 showing Level-1(α) performance across different game types
  • Hybrid Model Flow Chart: Based on Figure 5 showing the decision tree for model assignment