For most of the twentieth century, parking demand modeling meant opening the ITE Parking Generation Manual, finding the appropriate land-use code, and multiplying the project’s square footage by a demand ratio. That approach produced the oversupplied parking that defines much of American urban form today. Its methodological weaknesses are now well understood.
Contemporary practice has largely moved to one of three alternative methods. Each is an improvement on ITE-style factor-based modeling. Each has its own failure modes that planners should recognize before committing to a forecast.
Method One: Utilization-Based Empirical Modeling
The most widely adopted approach. Planners collect actual occupancy data at comparable existing sites — usually via license plate recognition sweeps or wireless sensor deployments — and scale the observed peak utilization to the proposed site based on shared trip generation characteristics.
Strengths: Grounded in real observed behavior. Sensitive to local transit access, walkability, and alternative mode availability in ways ITE factors ignore.
Failure modes: The comparables problem is severe. A downtown mixed-use development compared to a suburban mixed-use development will produce a utilization estimate that reflects the comparable’s context, not the project’s. Planners often select comparables for convenience rather than genuine similarity, and the resulting forecast systematically biases toward the comparable’s context.
Method Two: Shared Parking Models
Popularized by the Urban Land Institute’s Shared Parking publications and now standard in mixed-use zoning contexts. The method accounts for the fact that different uses peak at different times — office during weekday midday, restaurant evenings, residential overnight — and calculates the minimum parking supply that satisfies the true combined peak rather than the sum of individual peaks.
Strengths: Produces dramatically smaller parking recommendations than non-shared methods in true mixed-use contexts. Well-validated for combinations that appear in the ULI data.
Failure modes: The publication’s tables reflect specific land-use combinations studied in specific markets. Applying them to novel combinations — coworking + gym + restaurant, for example — means interpolating where data doesn’t exist. The resulting recommendations are often accepted as authoritative when they’re actually extrapolations.
The second failure mode is more subtle: shared parking models assume users will willingly walk between uses. In actual deployment, users often don’t, and the “shared” pool fractures into de facto dedicated pools that can both be full simultaneously.
Method Three: Agent-Based Simulation
The newest and most data-intensive approach. The site is modeled as a set of arriving and departing agents with individual trip purposes, dwell time distributions, and mode-choice probabilities. The simulation produces not just a peak demand estimate but the full temporal distribution, allowing operators to plan for capacity management rather than simple supply.
Strengths: Captures dynamics the static methods miss — queuing at entrances during peak arrival, exit congestion after events, the effect of pricing changes on mode choice.
Failure modes: Calibration data requirements are enormous. Most projects don’t have access to the detailed arrival-rate and dwell-time data the models require, so analysts substitute assumed distributions. The model’s apparent rigor can obscure that its outputs depend heavily on those assumptions. “GIGO with extra math” is a fair critique when calibration data is weak.
The Common Failure That Spans All Methods
Every method described here forecasts demand for the use pattern that existed in the data-collection period. None naturally captures structural change in parking demand — remote work, ride-hailing displacement, e-commerce delivery, automated vehicle scenarios.
The practical implication: any single-point forecast for a 30-year parking investment should be paired with sensitivity analysis on structural assumptions. “Demand will be 400-900 spaces depending on remote-work persistence” is a more honest planning output than “demand will be 650 spaces.”
Which Method to Use When
Single-use suburban development: Utilization-based with carefully selected comparables. ITE factors as a cross-check only.
Urban mixed-use development: Shared parking models, with the ULI tables as the primary reference but with documented sensitivity on combinations not in the published data.
Major capital investment (garages, structured parking, transit-oriented development): Agent-based simulation if the calibration data can be obtained; otherwise utilization-based with explicit sensitivity ranges on the structural assumptions.
Policy analysis (parking maximums, minimum reductions, removal of parking requirements entirely): Utilization-based modeling of existing conditions combined with stated-preference surveys for behavioral response. This is where the field is most methodologically unsettled.
Frequently Asked Questions
Is the ITE Parking Generation Manual still useful?
As a sanity check, yes. As a primary forecasting tool, no. The manual’s factors are averages over data collected predominantly in suburban automobile-oriented contexts between 1980 and 2010. Applying them to urban contexts or to post-pandemic travel patterns produces systematic overestimation.
How much does ride-hailing change parking demand?
The published literature suggests 5-15% reduction at mature deployment, concentrated in urban cores. The range is wide because the effect depends heavily on local ride-hailing pricing, service density, and alternative mode quality. Planners modeling greenfield developments should treat the ride-hailing reduction as a sensitivity variable, not a point estimate.
Do zoning changes reducing parking minimums produce the predicted demand?
Mostly yes. Portland, Seattle, and Minneapolis policy evaluations show that actual developer-chosen parking supply in the absence of minimums tracks measured demand within 10-15% for residential and mixed-use developments. Commercial developments show larger variance, likely driven by tenant-specific preferences not captured in aggregate demand modeling.
What’s the expected change from widespread autonomous vehicle adoption?
Genuinely unknown. Projections range from 30% demand reduction (if autonomous vehicles displace private vehicle ownership) to 10% increase (if they induce more trips). The professional consensus is to avoid baking autonomous-vehicle scenarios into planning-horizon demand forecasts for at least another decade.