A model predicts that a new transport link will save commuters time. But did it count all walking and transfers, model neighbouring congestion, and test its demand assumptions against real journeys? A sophisticated diagram is not a verified forecast.
The gold standard of transport modelling is an evidence-backed representation of travel behaviour and network effects, calibrated and validated for the actual decision. This 21-chapter guide explains origins, destinations, mode choice, assignment, uncertainty and post-opening tests.
Quick information: a model needs observable foundations
| Model element | Evidence | Tempting shortcut |
|---|---|---|
| Demand | Origins, destinations and travel purposes | One traffic count equals every trip |
| Choice | Usable modes, times and costs | All users can choose every mode |
| Network | Links, capacity and transfers | Straight line is a usable route |
| Calibration | Observed flows and travel times | A convincing animation |
| Forecast | Relevant scenarios and uncertainty | One central case is guaranteed |
| Evaluation | Independent checks and post-opening return | Approval equals real success |
Worked model example: origin-destination matrix
| Fictional movement | Trips |
|---|---|
| A to B | 120 |
| A to C | 80 |
| B to A | 90 |
| B to C | 110 |
| Total trips | 400 |
Origin totals A and B both equal 200; destination totals are A 90, B 120 and C 190. Both sets reconcile to 400, but the invented data do not establish real mode availability or time of travel.
Worked modal shift example
| Fictional measure | Before | After 48-trip shift |
|---|---|---|
| Total trips | 800 | 800 |
| Car journeys | 480 (60%) | 432 (54%) |
| Public transport journeys | 320 (40%) | 368 (46%) |
The arithmetic assumes constant total demand and a specific shift. The model still needs evidence that the change would actually happen and that any network or environmental consequences are credible.
The 21 transport modelling controls
Define the travel system
Chapters 1–6: The model must answer an actual decision · Define the geographic boundary and interactions · Model the correct time periods · Represent people rather than only vehicle counts · Establish origin-destination demand · Worked origin-destination matrix example.
Movement and mode choices
Chapters 7–12: Separate observed demand from a forecast · Understand trip generation and attraction · Mode choice depends on the real alternatives · Worked mode-share shift · Journey costs need complete paths · Network assignment depends on capacity.
Calibration, validation and scenarios
Chapters 13–18: Distinguish static and dynamic modelling · Calibrate parameters before evaluation · Worked calibration discrepancy · Validation needs information outside fitting · Behavioural assumptions and induced response · Include walking and accessibility barriers.
Operational results
Chapters 19–21: Treat uncertainty as scenario evidence · Monitor actual journeys after deployment · Govern models as auditable decision tools.
- The model must answer an actual decision
- Define the geographic boundary and interactions
- Model the correct time periods
- Represent people rather than only vehicle counts
- Establish origin-destination demand
- Worked origin-destination matrix example
- Separate observed demand from a forecast
- Understand trip generation and attraction
- Mode choice depends on the real alternatives
- Worked mode-share shift
- Journey costs need complete paths
- Network assignment depends on capacity
- Distinguish static and dynamic modelling
- Calibrate parameters before evaluation
- Worked calibration discrepancy
- Validation needs information outside fitting
- Behavioural assumptions and induced response
- Include walking and accessibility barriers
- Treat uncertainty as scenario evidence
- Monitor actual journeys after deployment
- Govern models as auditable decision tools
CONTROL 1 OF 21 · All controls
1. The model must answer an actual decision
Transport modelling represents travel demand and networks to explore conditions or compare options. It is not a digital copy of a city that automatically knows the future. A well-defined decision, spatial boundary, time period and required measure come before software choice.
A proposed bus link, road redesign and accessible station approach call for different scales of modelling. Identify what must change for people or freight, who is affected and which output would support the choice. Link modelling to independent observation rather than attractive animation.
CONTROL 2 OF 21 · All controls
2. Define the geographic boundary and interactions
A model must include the roads, services and neighbouring districts likely to be affected by the intervention. An improved junction can divert traffic into residential streets, or a new service can change trips beyond one corridor. A conveniently small boundary may exclude harm or benefits.
Describe internal zones and relevant external movements. For journeys across the boundary, explain how travel enters and leaves the study system. A road outside the visible map may still control the results through congestion or transfer behaviour.
CONTROL 3 OF 21 · All controls
3. Model the correct time periods
Morning peak, evening peak, weekend and holiday periods differ in purpose and demand. A model calibrated for one hour cannot automatically represent a whole day. Some networks change their service frequency or restrictions between periods.
Choose periods that match the public decision and explain whether demand is averaged or represented in shorter intervals. For a new school bus service, enrolment days and arrival deadlines may matter far more than general weekday totals.
CONTROL 4 OF 21 · All controls
4. Represent people rather than only vehicle counts
Vehicle flow is an important engineering measure, but it can conceal how many people are travelling. One bus can carry many passengers while a private car may carry one or several. A policy seeking access needs measures of people reaching actual destinations.
Use person trips, occupancy, travel-purpose distinctions and mode-specific journeys where relevant. A high vehicle-throughput score does not automatically mean better access for travellers without a car. The choice of unit is an important modelling assumption.
CONTROL 5 OF 21 · All controls
5. Establish origin-destination demand
An origin-destination matrix estimates or records trips from each origin zone to each destination zone during a stated interval. Row and column totals must be interpreted consistently. A matrix can be mathematically coherent yet fail to reflect actual travellers when the underlying survey or measurement is biased.
A useful matrix identifies trip purpose, time window, mode representation and data source. It also records whether flows are observed, inferred or projected. The smaller the zones, the greater the possible privacy and sampling problems in some datasets.
CONTROL 6 OF 21 · All controls
6. Worked origin-destination matrix example
In a fictional two-origin system, zone A sends 120 trips to B and 80 to C; zone B sends 90 to A and 110 to C. The origin totals are 200 and 200, yielding 400 trips overall. The destination totals are B 120, C 190 and A 90.
The row and column sums reconcile to 400, but the figures are invented. They do not say whether the journey destinations are actually useful or whether every trip occurs at the same hour. Verify matrix definition and data before treating sums as evidence of network demand.
CONTROL 7 OF 21 · All controls
7. Separate observed demand from a forecast
A traffic count, travel survey and mobile device dataset observe different slices of activity. Forecast demand uses assumptions about population, land use, prices and travel choices. A model should not label projected trips as actual measured movements.
Keep the forecast base year, input data vintage and scenario assumptions visible. An increase in passenger demand can reflect new trips, shifts from another mode or a different service area. These are different mechanisms with different public consequences.
CONTROL 8 OF 21 · All controls
8. Understand trip generation and attraction
Travel models may estimate how many trips are produced and attracted by zones under the chosen land-use and demographic assumptions. The number of residents alone is not enough; employment, schools, services, traveller characteristics and time periods also matter.
Describe data and calibration behind any rates. A residential zone and industrial zone can generate very different travel patterns despite similar mapped area. When new development is proposed, distinguish approved commitments from speculative plans.
CONTROL 9 OF 21 · All controls
9. Mode choice depends on the real alternatives
Walking, cycling, bus, rail, car and other modes differ in journey time, frequency, cost, reliability, accessibility and comfort. Modelling mode choice requires representing viable options available to the relevant user, not inventing a universal preference for one mode.
A household without a car does not face the same choice as a commuter who has one. A short bicycle trip may be unusable without a safe route. The model should include constraints and documented behavioural data to avoid overstating mode changes.
CONTROL 10 OF 21 · All controls
10. Worked mode-share shift
Suppose 800 fictional daily trips consist of 480 car and 320 public-transport journeys, giving a 60% car share and 40% transit share. If 48 of those car trips shift to transit with no change in total trips, the new counts are 432 and 368.
The new shares are 54% and 46%. This assumes a purely redistributive shift, not induced demand or suppressed travel. The calculation does not prove that any intervention would cause 48 real commuters to change modes.
CONTROL 11 OF 21 · All controls
11. Journey costs need complete paths
Travel demand depends on more than in-vehicle minutes. Access walks, waiting, transfer inconvenience, fares, reliability and accessibility can change relative attractiveness. A model that gives a train only its ride time while including the bus walk is biased.
Use identical origin-destination boundaries and explicit rules for route choice. Where generalised cost is monetised, document time values and other terms, using authorised local methodology. A headline mode score without components makes independent challenge difficult.
CONTROL 12 OF 21 · All controls
12. Network assignment depends on capacity
Assigning a fixed number of trips to shortest paths can ignore congestion or service crowding. In a realistic network, travel choices may respond to link times, capacities and route alternatives, which can in turn change with assigned demand.
Choose methods suitable to the problem and explain equilibrium or dynamic assumptions. A static route model may answer a simple geometry question but not a peak-hour queue spillback question. The model’s complexity should follow the decision and data quality.
CONTROL 13 OF 21 · All controls
13. Distinguish static and dynamic modelling
Static models often represent aggregate conditions for a period, while dynamic models represent time-varying conditions such as queues and interactions. Microscopic simulation can represent individual vehicles or agents under specified behavioural assumptions. More detailed animations do not guarantee better answers.
The FHWA Traffic Analysis Toolbox includes guidance for selecting methods and calibrating microsimulation. Report the spatial and temporal scale, process assumptions and data requirements rather than using a visual simulation as an independent source of truth.
CONTROL 14 OF 21 · All controls
14. Calibrate parameters before evaluation
Calibration adjusts plausible model parameters to match selected observed conditions. It should follow a documented plan and use appropriate data, such as flows, journey times or route choices. A model tuned to one attractive corridor can fail in adjacent locations.
Record calibration target, instrument source, time period, parameter limits and changes. Do not tune away unexpected evidence by rewriting data without justification. A calibration fit is not independent proof of the model’s forecasting power.
CONTROL 15 OF 21 · All controls
15. Worked calibration discrepancy
A fictional model predicts 5,460 vehicles in a defined period where an appropriate count observes 5,200. The model exceeds the observation by 260, or 260/5,200 = 5% relative to the observed value.
The result is a descriptive error for one count. It does not decide whether the model is fit for purpose or how its queues, travel time and mode share perform. A calibrated model needs additional checks appropriate to the decision.
CONTROL 16 OF 21 · All controls
16. Validation needs information outside fitting
Validation examines whether the model can reproduce relevant conditions or support intended predictions beyond parameters tuned during calibration. Using the same selected observations as both the sole calibration target and sole validation proof can inflate apparent success.
Choose independent data when feasible and document any limitations. For a corridor, checks may include travel-time distribution, queue behaviour and external movements rather than only one total volume. Explain where model confidence remains limited.
CONTROL 17 OF 21 · All controls
17. Behavioural assumptions and induced response
A new road, faster rail route or fare change can alter demand, departure time, destinations and mode choices. Treating future trips as fixed may overstate benefits or misdescribe impacts. Different induced responses can occur over different time horizons.
Document the behavioural evidence and uncertainty. Compare scenarios with a genuine no-change alternative and consistent land-use assumptions. An increase in modeled capacity is not a guarantee of permanent travel-time savings.
CONTROL 18 OF 21 · All controls
18. Include walking and accessibility barriers
A station-centre dot does not convey usable walking paths, street crossings, stairs or lift status. Travel models that use straight-line distance can create unrealistic access for some users. Accessibility matters at the level of the complete journey.
Model the permitted path where needed and verify critical links with appropriate field evidence. The result should distinguish an accessible vehicle from a journey that can genuinely be completed. The goal is useful access, not merely more route kilometres.
CONTROL 19 OF 21 · All controls
19. Treat uncertainty as scenario evidence
Population growth, development, weather, travel behaviour and costs may differ from central assumptions. One precise forecast cannot reveal all this uncertainty. Scenario modelling tests coherent alternative futures, while sensitivity tests explore individual assumptions.
The UK Department for Transport updated its TAG unit M4 forecasting and uncertainty guidance in May 2026. Its methods provide UK context and not automatic Singapore regulatory rules. Show which decisions change under plausible alternatives.
CONTROL 20 OF 21 · All controls
20. Monitor actual journeys after deployment
A model is an estimate; the opened road, bus service or public space is the actual outcome. Post-implementation evidence can reveal where forecasts succeeded, failed or concealed unequal impacts.
Define before-and-after measures, data collection dates and comparison limitations. Keep counts of people, service reliability and safety separate. A construction milestone cannot substitute for observed journey outcomes or justify selectively reporting only favourable links.
CONTROL 21 OF 21 · All controls
21. Govern models as auditable decision tools
Model runs depend on source data, network versions, code, parameters and scenario choices. A credible report must preserve these so independent reviewers can reconstruct the claim. Important failures and untested regions belong in the review, not just the preferred forecast.
Keep an assumptions register and versioned outputs. The immutable Clementi Secondary 1 Mathematics reference models the transferable reasoning habit: define quantities, calculate consistently and test the result when conditions change.
Independent transport model audit
| Stage | Required record | Misleading failure |
|---|---|---|
| Question | Users, geography and time periods | A tiny model hides network impacts |
| Demand | Trip observations and projection basis | Forecast called measured volume |
| Behaviour | Mode availability and choice assumptions | Every person assigned unrealistic alternatives |
| Calibration and validation | Separate checks and error distributions | One count matches so everything is correct |
| Uncertainty | Scenarios, alternative futures, update dates | One precise answer presented as inevitable |
| Real return | Measured outcomes and correction ownership | Construction completion replaces evaluation |
Frequently asked questions
What is transport modelling?
Representing travel demand and networks to investigate conditions and compare policies or infrastructure options.
Does a traffic count equal demand for all modes?
No. Counts at one location or mode do not describe every person trip or route choice.
What is calibration?
Adjusting appropriate model parameters to reflect selected observed conditions.
Why validate separately?
A fitted model can match its calibration observations yet fail in other conditions.
Does a more detailed simulation guarantee accuracy?
No. Fidelity depends on data, methods, uncertainty and the intended question.
Can one projected mode split be taken as fact?
No. Behavioural assumptions and alternative scenarios must be tested.
Authority, guidance and limits
The UK Department for Transport Transport Analysis Guidance was updated 28 May 2026. Its TAG M1-1 Principles of Modelling and Forecasting and TAG M4 Forecasting and Uncertainty provide methods in the UK context. The US Federal Highway Administration Traffic Analysis Toolbox covers tool selection and microsimulation. These references are not Singapore project approvals. Numerical cases are fictional.
Connected Gold Standard subjects
Transport Planning · Transport Policy · Transport Economics · Transport Geography · Traffic Engineering · Traffic Management · Transportation · Urban Planning · Systems Engineering · Data Science · Statistics · Forecasting · Transport Accessibility · Risk Management · Education
Additional worked comparison: shift in mode versus induced demand
In a simplified fictional corridor, 800 person trips occur during a defined period: 480 by private car and 320 by public transport. If exactly 48 travellers switch from car to public transport and no additional trips occur, car trips decline to 432 and public-transport trips rise to 368. The total remains 800. Car share decreases from 60% to 54%, a six-percentage-point change, not a six-percent reduction in total travel demand.
Now consider another hypothetical future in which the improved service also enables 40 additional trips, all by public transport, while the initial 48-trip switch still occurs. Total person trips become 840. Car trips remain 432, while public-transport trips become 408. Car share is 432/840, approximately 51.4%, and public-transport share is approximately 48.6%. The numbers tell a different story because the total demand denominator has changed.
| Metric | Mode shift only | Mode shift plus 40 new trips |
|---|---|---|
| Total person trips | 800 | 840 |
| Car trips | 432 | 432 |
| Public transport trips | 368 | 408 |
| Car mode share | 54% | ≈51.4% |
Neither scenario proves real passengers would behave this way. The model needs observations or credible parameters about trip generation, alternatives, access and timing. Counting new public-transport trips as diverted car trips would exaggerate a road-congestion benefit. In a real network the new journeys could also create different crowding or service requirements; the comparison must include relevant costs and outcomes.
Additional validation exercise: matching one count is not enough
Suppose a calibrated model predicts 5,460 vehicles for a specified corridor period against an observed count of 5,200. Relative error is 260/5,200 = 5%. That number may be acceptable for some exploratory purposes and unacceptable for others, depending on the intended decision and validation criteria. It cannot establish how well the model predicts queue lengths, bus passenger delays, pedestrian access or routes elsewhere.
An independent validation plan would check different conditions and measures, such as additional traffic-count sites, representative journey-time distributions and relevant transit outcomes. It would explain which parameters were calibrated, which data were held aside and whether the observed system changed during collection. If the model was repeatedly altered until every validation count fitted, the validation set may no longer be truly independent.
Make safety, accessibility and uncertainty visible
A route model may indicate a station is two hundred metres from a residential entrance, yet a river crossing, steps or an out-of-service lift make the nominal path unusable for some travellers. The model must distinguish a straight-line distance from a real, legally permitted, accessible journey. Include first- and last-mile paths and verify important links where the decision depends on them.
Where a model guides high-consequence decisions, apply appropriate professional transport safety standards and responsible public governance. A favourable mean travel-time saving does not offset a mandatory safety breach. The UK Transport Analysis Guidance and US FHWA toolkits provide useful modelling principles but do not authorise a particular Singapore infrastructure design or operational change.
Release checklist for a reproducible transport model
An audit should locate the study purpose, area and period; demographic and development assumptions; origin-destination data; mode alternatives; network and capacity model; source data and their dates; calibration changes; independent validation; scenarios for uncertainty; and planned post-opening monitoring. Each statement about an improvement should be traceable to a model output and the evidence supporting its assumptions.
Maintain the model and data as versioned artifacts. A decision maker should be able to see what happens if demand is lower, delays are higher or a key transfer becomes inaccessible. When real results become available, compare them honestly with the original forecast and document corrections rather than presenting the model as infallible.