PolicyEngine tutorials

Hands-on sessions for analysing tax and benefit policy reforms — household by household, or across the whole population. Everything runs in your browser.

IMA World Congress 2026 · Brussels

Analysing tax-benefit reform impacts with PolicyEngine

Friday 3 July, 10:30–12:00 · Room F (2300) · UCLouvain Saint-Louis Bruxelles · Max Ghenis (PolicyEngine)

Open the notebook in Colab Open the web app Setup instructions

Coming to the live session? Open the Colab notebook, save a copy to your Drive, and run the Setup cell beforehand — the install takes ~2–3 minutes, so doing it ahead means nobody waits on it together on the conference Wi-Fi.

Session flow

Live US and UK examples, in the order we'll run them — every link works on your own laptop, so follow along or explore ahead:

1 · Cold open

CliffWatch

Benefit cliffs and marginal tax rates across all 50 US states — every line is the full tax-benefit system.

2 · What PolicyEngine is

Scope of the models

Free, open-source tax-benefit microsimulation — full federal + state coverage in the US, full coverage in the UK, and growing country models.

3 · A real reform, end to end

UK Autumn Budget 2025

A real fiscal event: what it does to a single household and to the whole population — budget, distribution, poverty.

4 · Household by household

OBBBA explorer

The biggest US tax law in a decade, encoded within days — explore its impact on 40,000 real survey households.

5 · Under the hood

How the data works

Populace: multi-source fusion, distributional imputation, and calibration to 30,000+ administrative targets.

6 · Hands-on

Web app + Colab

Build your own reform in the browser — including elasticity-based labour-supply responses — then drive the same engine from Python.

7 · AI

UK chat assistant (demo)

Ask a UK policy question in plain language — answers computed by the PolicyEngine engine, not a language model guessing.

8 · The rules layer

Axiom (demos)

Where this is going: open infrastructure that encodes statutes as executable, verifiable rules — the layer the next generation of models builds on.

How the data works

Population results run on Populace, PolicyEngine's open US microdata. It isn't a single survey — it's assembled from primary sources and calibrated to administrative totals, which is what lets a national reform estimate match official aggregates while staying reweightable to any region.

1 · Sources

Primary data, combined

Household surveys (CPS, ACS), tax records (IRS), and wealth data (SCF) — each contributing the variables it measures best.

2 · Imputation

Fused, not averaged

Variables no single survey observes together are joined with zero-inflated quantile regression forests — sampling the full conditional distribution, so spread and the mass at zero survive.

3 · Calibration

30,000+ targets at once

Gradient-descent reweighting fits tens of thousands of administrative totals simultaneously, with L0 regularisation for a sparse, parsimonious set of households — inspect every target fit in the live dashboard.

4 · Geography

Reweighted, not rebuilt

The same engine reweights the ~340M-person national file to states and congressional districts — one dataset, filtered, not separate per-area files.

Full methodology, papers, releases, and validation evidence live at populace.dev. All of it is open source — Populace and the policyengine package — and the same calibrated microdata powers the web app, the Python API, and this notebook.

What you need to participate

Almost nothing. This tutorial is browser-based — a laptop with a modern browser and the conference Wi-Fi is enough. The steps below are optional, for going deeper.

Required · everyone

A browser

Confirm you can reach policyengine.org and app.policyengine.org. The household calculator, reforms, and population analysis all run there — no account, no install.

Optional · for the notebook

A Google account

To follow the Python notebook in Google Colab — runs in your browser, nothing installed. Before the session: open it, File → Save a copy in Drive, then run the Setup cell (installs in ~2–3 min).

Optional · run locally

Python ≥ 3.10

Prefer your own machine? pip install policyengine installs the US and UK models.

Optional · demo

The AI assistant

The Claude Code plugin is shown live — nothing to install to follow it. To try it yourself you'll need Claude Code and an Anthropic account.

Hands-on examples use the US model, whose microdata is open. The UK model works identically, but its population microdata isn't publicly downloadable, so UK population analysis is shown by the presenter.

Questions about this tutorial or the setup? Email max@policyengine.org.