title	On the Comparative Performance of Machine Learning and Economic Models for Risky Decision-Making
author	Mateus Hiro Nagata
date	6/1/2025
output	github_document

---

🧠 On the Comparative Performance of Machine Learning and Economic Models for Risky Decision-Making

This repository contains the reproduction package for the working paper "On the Comparative Performance of Machine Learning and Economic Models for Risky Decision-Making".

The project investigates how machine learning methods can model and predict human decision-making under risk, comparing traditional economic frameworks (like Cumulative Prospect Theory) with data-driven and hybrid approaches.

---

📘 Overview

The goal is to compare three broad classes of models:

1. Economic Models – grounded in behavioral economics (e.g., Cumulative Prospect Theory).
2. Pure Machine Learning Models – fully data-driven (e.g., random forests, regularized regressions).
3. Hybrid Models – neural networks that embed economic theory into their structure or loss function.

This comparison provides insights into how well each class captures utility curvature and probability weighting, two central features of decision-theoretic behavior.

---

📁 Repository Structure

Script	Description
DT_ML_Data.py	Prepares and preprocesses the dataset for training and evaluation. Includes feature construction and transformations related to decision-theoretic variables.
DT-ML MLE.py	Implements Maximum Likelihood Estimation (MLE) for parameter estimation in decision-theoretic models.
DT-ML NN.py	Defines and trains hybrid neural network models incorporating economic theory (e.g., constrained layers or custom loss terms).
DT-ML Regularized Regressions.py	Trains regularized regression models (LASSO, Ridge, Elastic Net) as baseline comparators.
DT-ML Trees.py	Implements tree-based ML models such as Random Forests and Gradient Boosted Trees.

---

⚙️ Requirements

This project is written in Python 3.14.0. Core dependencies include:

numpy
pandas
scikit-learn
tensorflow  # or pytorch
matplotlib
seaborn
scipy

To install dependencies:

pip install -r requirements.txt

---

🚀 Usage

To reproduce results from the paper:

1. Prepare the data

   python DT_ML_Data.py

2. Train and evaluate models

   python DT-ML_Trees.py
   python DT-ML_Regularized Regressions.py
   python DT-ML_NN.py

3. Analyze outputs

   Model metrics, parameter estimates, and figures are saved in the designated results folders or printed to the console.

---

📊 Outputs

The scripts generate:

• Comparative performance metrics (accuracy, RMSE, log-likelihood)
• Visualizations of utility and probability weighting functions
• Tables comparing economic, ML, and hybrid models

---

🧩 Research Context

This work contributes to bridging behavioral economics and machine learning, by:

• Testing empirical validity of decision-theoretic assumptions
• Using ML to discover flexible, nonparametric patterns in choice behavior
• Building interpretable hybrid models combining structure and prediction power

---

Data

• The data is from Vieider, Ferdinand M., et al. "Common components of risk and uncertainty attitudes across contexts and domains: Evidence from 30 countries." Journal of the European Economic Association 13.3 (2015): 421-452
• Unfortunately, I do not have authorization to publish the data.

---

📄 Citation

If you use this code or reproduce the results, please cite:

Mateus Hiro Nagata. On the Comparative Performance of Machine Learning and Economic Models for Risky Decision-Making. Working Paper, HEC Paris, 2025.

---

🪪 License

This project is released under the MIT License.

---

Project Structure

├── data/
├── code/
│   ├── DT_ML_Data.py #Cleaning data
│   ├── DT-ML_MLE.py  # MLE estimation of CPT
│   ├── DT-ML_NN.py   # NN predictions and hybrid models 
│   ├── DT-ML_Regularized_Regression.py # LASSO
│   ├── DT-ML_Trees.py # Random Forests and other variations
├── output/
│   ├── figures/       # Generated plots
│   └── tables/        # Results tables
└── README.md