The classical four-step model — trip generation → trip distribution → mode choice → trip assignment — implemented end-to-end on real, open data for Milan, and pushed past the textbook baseline: real GTFS transit times, user equilibrium via MSA, and a policy scenario. Pure Python (OSMnx + GeoPandas + SciPy + NetworkX), no modelling software, no bundled datasets — everything downloads at run time.
→ Read the notebook · → interactive maps live at abadeanlou.com/four-step
User-equilibrium car assignment on Milan's real network; below, the Area-C-style through-traffic-ban scenario (red = more traffic, green = less):
| Stage | Method |
|---|---|
| Network | OSM major roads, largest strongly connected component, assumed speeds/capacities per class, compressed to SciPy sparse form for fast repeated assignment |
| Zones & land use | ~1.2 km hex grid; population ∝ OSM residential area, jobs ∝ OSM workplace features, scaled to stated control totals |
| 1. Generation | Production/attraction rates, attraction balancing |
| 2. Distribution | Doubly-constrained gravity model, exponential deterrence, Furness/IPF |
| 3. Mode choice | Binary logit on real scheduled transit times — a frequency-based transit graph built from Milan's official GTFS feed (boarding waits from AM-peak headways, scheduled in-vehicle times, walking transfers) |
| 4. Assignment | User equilibrium via the Method of Successive Averages with BPR volume-delay, relative-gap convergence plot, v/c comparison against all-or-nothing |
| Scenario | Through-traffic ban in a 2.5 km inner cordon (an Area-C-style measure): two-class equilibrium assignment, difference map of where banned trips relocate |
| Extras | Select-link analysis at equilibrium, car + PT skim matrices, GeoJSON exports, folium interactive maps |
- Road network, boundary, land-use proxies: OpenStreetMap (via OSMnx)
- Public transport: the official GTFS feed published on the Comune di Milano open-data portal (produced by AMAT from ATM scheduled service data)
- Trip rates, deterrence and logit parameters, BPR coefficients, and the population/jobs control totals are stated assumptions, not values calibrated to Milanese surveys or counts. The notebook flags every one.
- The transit graph is frequency-based (headway/2 boarding waits), not a timetable router; the closing section lists what a production model adds.
- I have applied this same workflow to a real metropolitan planning network (~58,000 links with 2030/2040 scenario attributes); that dataset is not distributable, hence this open-data rebuild.
python -m venv .venv && .venv/Scripts/activate # or source .venv/bin/activate
pip install -r requirements.txt
jupyter lab notebooks/four_step_model_milan.ipynbFirst run downloads the Milan network, OSM features, and the ~26 MB GTFS feed (cached afterwards). Full execution takes several minutes.
- routing-engine-osmnx — FastAPI shortest-path service on OSM networks, live demo
- Accessibility-using-Transit — GTFS transit accessibility for seven European cities, live maps
- bikeflow dashboard — 24/7 forecasting pipeline on live bike-share data