A transportation (travel demand) model is a quantitative framework that forecasts how people and goods move through a region under different land-use and network scenarios. Done well, a model becomes a decision engine: it helps test road and transit projects, pricing policies, and growth plans before spending real money.
This guide covers when you actually need a model, what data it requires, common software, the four-step modeling process, the GIS pieces, calibration/validation, costs, timelines, and practical tips.
When do you need a travel demand model?
Use a regional TDM when you must answer questions like:
- Which corridor alternatives best relieve congestion by 2035 or 2045?
- How do land-use changes (new housing, employment centers) shift travel patterns?
- What is the network-wide effect of a BRT/metro line, tolls, or parking pricing?
- How do policy scenarios (fuel cost, fares, telework) affect mode shares?
You do not always need a full TDM:
- Small towns/corridors: Sketch planning (HCM methods, ITE trip rates, simple spreadsheets) or microsimulation of a corridor may be enough.
- Project-level design: Use operational tools (HCM, SIDRA, VISSIM/SUMO) once the regional demand is known or assumed.
What a TDM can and can’t do
Can
- Forecast network-wide demand by O-D, time period, and mode.
- Compare alternatives consistently using the same assumptions.
- Produce link volumes, speeds, V/C, transit boardings, and accessibility metrics.
Can’t (without extra work)
- Predict driver behavior perfectly; models approximate reality.
- Replace detailed operations analysis at intersections.
- Work reliably without good data and careful calibration.
Core software stack (pick what fits your scope and budget)
| Purpose | Commercial | Open-Source / Low-Cost | Notes |
|---|---|---|---|
| Macroscopic 4-step models + transit | PTV VISUM, TransCAD, EMME, CUBE | AequilibraE (QGIS plugin), Emme Classic tools, Python libraries | Strong transit and assignment features in commercial tools; AequilibraE is improving fast. |
| Dynamic traffic assignment (DTA) | Aimsun Next, Dynameq, PTV Vissim + Visum/Vistro | DTALite, NeXTA | DTA captures time-dependent queuing/spillback; heavier data and calibration needs. |
| Microsimulation (ops-level) | PTV Vissim, Aimsun | SUMO | Use after TDM to test signal plans/queues at project level. |
| Activity-based / agent models | (Commercial ABM add-ons) | ActivitySim, MATSim | Higher fidelity (daily activity patterns), higher data and skill requirements. |
| GIS + data engineering | ArcGIS Pro | QGIS, PostgreSQL/PostGIS, Python, R | Essential for TAZs, network building, GTFS, ETL. |
| Visualization | Tableau, Power BI | Kepler.gl, QGIS, Python dashboards | For maps, screenlines, KPI dashboards. |
Data requirements (the part that makes or breaks your model)
| Data Category | Examples | Typical Sources | Notes |
|---|---|---|---|
| Base network | Road centerlines, lanes, speeds, capacities, restrictions | OpenStreetMap, local road agencies, GIS base maps | Conflate to a routable graph; code facility types and turn restrictions. |
| Transit supply | Routes, stops, headways, fares, access links | GTFS from operators, agency shapefiles | Import GTFS; check stop spacing, transfers, and walk access. |
| Zones & land use | TAZ boundaries; households, population, employment by sector | Census/Statistics office, planning depts., parcel data | Zones should align with barriers and major roads; keep intrazonal sizes reasonable. |
| Socio-economics | Income, car ownership, student/worker ratios | Household surveys, census | Critical for mode choice segmentation. |
| Travel surveys | Household travel survey (HTS), intercepts, RP/SP surveys | Commissioned studies, universities, consultants | Gold standard but expensive; sample should represent all market segments. |
| Counts & screenlines | 24-hr/peak link counts, turning counts, transit boardings | Traffic counts, APC/AVL data | Use for calibration/validation (GEH, RMSE, screenline checks). |
| O-D/Probe data | Mobile phone (CDR/GPS), app data, floating car travel times | Data vendors, Google/HERE APIs (travel times) | Useful to seed O-D matrices and validate speeds. |
| Future land use | Growth by TAZ, development plans | MPO/planning agencies | Drives forecast scenarios. |
The GIS pieces you cannot skip
- Define TAZs that respect physical barriers and land-use homogeneity; avoid very large zones in urban cores.
- Build/clean the network: conflate multiple sources; snap nodes; code lanes, speeds, capacities, turn bans, tolls, HOVs, centroid connectors.
- Transit import: ingest GTFS; verify headways, fares, access links, transfers.
- Skim geography: maintain consistent coordinate systems; precompute walk/bike access distances and impedances.
The classical 4-step modeling procedure (plus data needs)
Trip Generation
- Goal: Estimate productions/attractions by purpose (HBW, HBS, HBO, NHB, freight).
- Methods: Cross-classification (category analysis), linear/log-linear regression.
- Inputs: Households by size/income, employment by sector, auto ownership, school enrollments.
- Outputs: Trips by TAZ and purpose (P/A).
Trip Distribution
- Goal: Connect P/A to O-D flows.
- Method: Gravity model with impedance (time/cost) and friction factors; optional K-factors for special pairs.
- Inputs: Skims (time, cost, distance), productions/attractions, seed O-D (if available).
- Checks: Average trip length (ATL) by purpose; intrazonal share; reasonableness of flows.
Mode Choice
- Goal: Split O-D travel among auto, transit (local/rapid), walk/bike, TNC/taxi.
- Method: Multinomial or nested logit (often by market segments: income, car ownership, trip purpose).
- Inputs: Skims per mode (in-vehicle time, wait, walk, transfer, fares, parking cost), socio-economics.
- Estimation: Use survey RP/SP data to estimate coefficients (value of time, transfer penalties).
- Outputs: Mode shares by O-D and purpose.
Assignment
- Highway: Static user equilibrium (UE) with volume-delay functions (e.g., BPR), or DTA for time-dependent effects.
- Transit: Multi-path assignment with crowding/transfer penalties where supported.
- Outputs: Link volumes, speeds, V/C, VHT/VKT, queue proxies; station/line loads for transit.
Skims (the glue)
- Before distribution/mode choice, compute skim matrices (shortest-path impedance by time/cost). Update skims iteratively as networks load.
Calibration and validation (how to make the model credible)
Key concepts
- Calibration: Adjust model parameters so base-year outputs match observed data.
- Validation: Test on hold-out data (or different time period/locations) to ensure transferability.
Typical calibration sequence
- Network & speeds: Ensure free-flow speeds and capacities by facility type are realistic; tune volume-delay (BPR) parameters.
- Trip generation: Match observed totals by purpose and area type; adjust rates/auto ownership models.
- Trip distribution: Calibrate friction factors to match average trip length and intrazonal shares; use K-factors sparingly for known anomalies.
- Mode choice: Fit to observed mode shares by segment and corridor; apply reasonable transfer and parking penalties.
- Assignment: Match link counts, turning counts, and screenlines; tune centroid connectors and turn penalties.
Common target metrics
- Link volume GEH: ≥85% of calibration counts with GEH < 5; most links < 7.
- RMSE / %RMSE by facility and volume bin within accepted ranges.
- Screenline totals within ±5–10%.
- Average trip length within ±5% by purpose.
- Mode shares within ±2–3 percentage points overall and by segment.
- Transit boardings/line loads within ±10–15% at key locations.
Tools & techniques
- Matrix estimation (ME/ODME) using counts and priors.
- Sensitivity tests: increase fuel price, change transit headways, add parking cost; verify directional and approximate elasticities make sense.
- Split data into calibration and validation sets to avoid overfitting.
How long does it take?
Indicative timelines (heavily scope-dependent):
| Scope | Typical Duration | Notes |
|---|---|---|
| Small city, classic 4-step, limited new surveys | 3–6 months | Reuse existing data; focus on calibration to counts/screenlines. |
| Mid-size metro, full update with some surveys | 6–12 months | Fresh HTS/OD data, transit network coding, policy scenarios. |
| Large metro, ABM or DTA + new surveys | 9–18+ months | Complex networks, multiple agencies, iterative calibration. |
Team often includes a PM, data/GIS engineer, modeler, survey specialist, and QA/QC.
What does it cost?
Very rough, order-of-magnitude guidance (varies by region, rates, data, licenses):
- Data: Household/OD surveys are the biggest ticket; probe data and counts also add up.
- Software: Commercial licenses can range from low five figures to higher, per seat per year; open-source reduces license costs but increases engineering effort.
- Delivery:
- Small city, 4-step with limited new data: ~US$50k–150k.
- Metro with new surveys and commercial stack: ~US$300k–$1M+.
Budget ongoing maintenance for annual updates to networks, land use, and counts.
Is a TDM necessary for every city or town?
Not always. Consider lighter methods when:
- The question is local/corridor-specific (e.g., one interchange or main street).
- The budget and timeline are constrained, and decisions don’t hinge on network-wide redistribution.
- You can defend a decision using HCM, ITE trip rates, targeted counts, and microsimulation.
Escalate to a TDM when cumulative, citywide effects matter or when agencies need a consistent forecasting platform for multiple projects.
Practical workflow you can reuse
- Scope: Purposes, time periods, geography, KPIs, scenarios.
- Data audit: What exists? What must be collected or purchased?
- GIS: TAZs, network, GTFS, centroid connectors.
- Base model: Trip gen → distribution → mode choice → assignment. Build skims.
- Calibration: Follow the sequence above; document every change.
- Validation: Independent counts/screenlines, hold-out geographies.
- Scenarios: Future land-use, network projects, policy tests.
- Documentation & handover: Model spec, user guide, versioned data, scripts.
Common pitfalls (and how to avoid them)
- Messy zones: Oversized TAZs hide short trips and distort intrazonals → refine in urban cores.
- Untuned speeds: Unrealistic free-flows or BPR parameters → garbage skims and misallocated demand.
- Overusing K-factors: Hide structural issues; fix network or data first.
- Ignoring transit access: Missing walk links or transfer penalties → inflated transit shares or misassigned paths.
- No version control: Use Git; tag calibration milestones; keep a change log.
- Poor documentation: Future users can’t reproduce results → create a data dictionary and runbook.
Deliverables clients and agencies expect
- Model files + networks + skims for base and forecast years.
- Data dictionary and metadata for every table and parameter.
- Calibration report with targets, diagnostics, and achieved metrics.
- Scenario book: assumptions, results, maps, KPIs.
- Run scripts (batch) to reproduce results end-to-end.
FAQ
- Four-step vs. ABM? ABM adds behavioral realism but needs more data and skill; 4-step is faster and adequate for many planning questions.
- Static vs. DTA? DTA for time-dependent queuing/spillback when peak-period dynamics matter; otherwise static UE suffices.
- SUMO’s role? Use SUMO/VISSIM after the TDM to test intersection control, lane use, and queuing for specific projects.
