Transport model learns from unused routes—but matching flows does not guarantee reliable welfare estimates

Designing a coordinated mobility market requires understanding both sides: travelers choosing routes and operators deciding where to deploy capacity. In a preprint, Joseph Y. J. Chow of New York University extends an economic matching framework to networks where a journey can involve several providers. Computational experiments show that unused routes contain valuable information—but also that predicting passenger flows well does not establish that a model measures economic benefits accurately.

Matching travelers with capacity

The framework builds on assignment games, which describe matches between buyers and sellers together with payments that support stable outcomes. Here, the buyers are traveler groups, while the sellers are mobility operators. Travelers choose journeys; operators allocate limited capacity across their services.

Chow develops two formulations. One explicitly represents complete travel paths. The other uses individual network links and a perturbed utility route choice model, avoiding the need to enumerate paths. This link-based version allows some routes to receive no flow and accommodates overlapping journeys.

It also distinguishes between capacity deployed on individual links and capacity deployed across an entire service. A train, for example, provides seats along every link it traverses. The model can therefore combine rail services with link-level providers such as parking lots or local feeder transport.

Within the framework, market-clearing payments support an allocation that maximizes the model’s regularized social-surplus objective. Existing fares need not produce that outcome. At fixed fares, scarce capacity introduces a separate rationing price: a shadow value affecting travelers’ choices, not additional revenue received by operators.

Working backward from observed flows

The inverse model estimates traveler preferences and operator-side quantities from observed outcomes. Its estimation problem separates into two convex quadratic programs: one for travelers and another for operators.

An important design choice is to retain constraints associated with every observed zero-flow combination of traveler group and link. A genuinely unused alternative helps restrict which preferences could explain the observed choices. Rationing prices are permitted only on links observed at capacity, with bounds and a penalty limiting their ability to absorb unrelated discrepancies.

This is not estimation from a few passenger counts alone. The setup assumes exact traveler-group flows on sampled links, aggregate counts across all operated links, and observed deployed service capacities. Fares, network structure, demand totals and capacity limits are also known, while uncertain operating costs are estimated relative to a prior.

A synthetic Long Island test

The larger experiment uses a synthetic morning-commute market based on the Long Island Rail Road network. It contains 897 nodes, including 671 census tracts and 123 stations, connected by 8,320 directed links. Three modeled operators provide rail, local feeder transport and parking.

Public data inform the network, demand, fares, seat supply and operating costs. However, the flows used to test estimation are generated by a model, not collected passenger movements serving as independent validation.

The data-generating model includes 97 station-specific constants that the estimated model deliberately omits. This creates a specification mismatch. The experiment asks whether a simpler model can still recover flows, scarcity patterns and economic measures when its behavioral specification differs from the generator’s.

Across sampling rates of 10% to 50% of links, with five replications at each rate, flow fits had R-squared values between 0.96 and 0.98. Weighted mean absolute percentage error on held-out links declined from 28.9% to 19.2%. Correlation with the simulated true rationing prices rose from roughly one-half to about three-quarters.

Removing the zero-flow constraints more than doubled held-out error at the sampling rates tested in the ablation study. At 50% sampling, error increased from 19.2% to 44.6%. Suppressing rationing prices also worsened held-out error and pricing-wedge correlation.

Where the economic estimates diverged

High flow accuracy did not carry over uniformly to welfare-related outputs. At 50% sampling, estimated gross surplus at the model’s social optimum was 5.8% below the synthetic benchmark. Gross surplus at existing fares, however, was underestimated by 33.4%, and the resulting deadweight-loss estimate was 17.3% too high.

The paper attributes this divergence to omitted station constants being absorbed mainly into the outside-driving alternative. Observed flows reveal differences between alternatives without necessarily establishing the correct surplus level relative to that outside option.

The model’s remapped optimal fares were also lower than published fares, but the paper explicitly cautions that this example excludes capital costs and considers only inbound travel to New York City. Those results are not a demonstrated fare prescription for the actual railway.

The contribution is a tractable estimation framework with encouraging synthetic results. Its central warning is equally important: reproducing transport choices is not, by itself, sufficient validation for economic benefit calculations.