Overview

EconIRL estimates models where agents make repeated choices while anticipating future consequences. It combines classical dynamic discrete choice estimators with inverse reinforcement learning estimators behind one Python interface.

Use EconIRL when the question is structural rather than only predictive. The outputs include reward objects, policies, value functions, and counterfactual policy calculations where the estimator supports them.

Read this overview first if the menu names are unfamiliar. It gives the basic problem shape before the estimator pages start using method-specific notation.

What are DDC/IRL models even about?

Dynamic discrete choice models study repeated decisions. A user, firm, driver, or operator chooses an action today while anticipating how that choice changes future states. The model connects observed choices to payoff tradeoffs.

Inverse reinforcement learning starts from a similar decision problem but asks for the reward model behind observed behavior. Instead of only predicting the next action, IRL estimates what the decision maker appears to value and how behavior changes under new incentives or environments.

EconIRL brings these two traditions into one workflow. Fit an estimator, inspect the fitted rewards and policies, then run supported counterfactuals.

Installation

pip install econirl

Quick Start

This example uses NFXP because it is the reference tabular structural estimator. The code is meant to show the workflow, not to decide that NFXP is right for every dataset.

from econirl.datasets import load_rust_bus, rust_bus_reward_spec
from econirl import NFXP

df = load_rust_bus()
model = NFXP(n_states=90, discount=0.9999, utility=rust_bus_reward_spec(90))
model.fit(df, state="mileage_bin", action="replaced", id="bus_id")

print(model.params_)
cf = model.counterfactual(replacement_cost=4.0)
print(cf.policy[50, 1])

Output

{'operating_cost': 0.001002924937407198, 'replacement_cost': 3.072263682263484}
0.055196266692073837

Where to Start

Question	Documentation entry point
I have my own panel of decisions.	Read using your own data.
I need a tabular DDC baseline.	Use NFXP.
I need to compare methods.	Use Choosing and Comparing Estimators.
I need the signature of a class.	Read the API reference.
I need to understand simulation-study evidence.	Read simulation studies.

Latest Updates

The documentation separates available methods from simulation-study evidence. Each estimator page states the target, the reported evidence, and the conditions under which the method should be used.

NFXP is the reference structural estimator for tabular dynamic discrete choice.