AV demand forecasting

Approach

To approximate real customer behavior before launch, we adapted Miami's regional travel demand model to include a new AV ride-hail option. The travel model is a city-planning tool that simulates how a synthetic population of residents makes daily travel decisions. This gave us a realistic pool of ~100,000 simulated customers and 170,000 trip requests to work with.

The model represents user choices probabilistically: each agent weighs the expected cost and travel time of every available mode (car, transit, walking, ride hail) for their specific trip. This made it well-suited for sensitivity testing. We could adjust a single variable, like fare level, and observe how customer behavior shifted across segments.

We validated the model's spatial and temporal distributions against real-world traffic and ride-hail data to confirm the outputs were reasonable before drawing conclusions.

Scenario Design

I designed a scenario matrix testing three service levers at three levels each (25%, 50%, 75%):

• Discounted fares

• Reduced wait times

• Faster in-vehicle travel times

This produced a structured set of comparisons that let us isolate which factors actually drove customer choice, and by how much.

Results

The most consequential findings revealed what riders actually care about when deciding whether to travel by AV.

Fares matter far more than wait time. A 50% fare discount tripled ride demand. A 50% reduction in wait time increased demand by only 50%. This was a significant signal: customers choosing AV ride-hail are more price-sensitive than convenience-sensitive.

In-vehicle travel time was essentially irrelevant. Customers were not deterred by longer journey times once they were in the vehicle. They cared about getting a ride at an acceptable price, not about how fast it arrived at the destination.

Demand varied meaningfully by segment. Higher-income households were the most responsive to wait time reductions (not fares). Work trips increased as wait times decreased. Car-free households represented a significant pocket of latent demand, a finding with implications for equity-oriented service design.

Fleet size hits diminishing returns. As fleet size scaled up, per-vehicle utilization efficiency declined, ranging from 4 trips/hour at smaller fleet sizes to 2.2 trips/hour at larger ones. More vehicles did not translate linearly to more value.

Decisions this work informed

These findings directly shaped Ford's pre-launch planning:

Fleet configuration: Simulations showed the most efficient fleet for this service area and fare structure was 100 vehicles operating on local streets. Highway operation would have required a significantly larger fleet to achieve equivalent ride volume, due to faster trips generating lower per-minute fares.

Pricing strategy: The fare sensitivity finding shifted pricing conversations, as the data supported a competitive launch fare rather than premium positioning.

Operational targets: Recommendations I developed were tied to a 16% improvement in fleet operational efficiency and 24% higher per-mile revenues compared to baseline scenarios.

The credibility of this work contributed to my being offered a role on Ford's Corporate Strategy team focused on AV development strategy for the Board of Directors.

Limitations

The model's behavioral assumptions were calibrated from existing travel data, not primary user research with AV-specific attitudes. Respondents in the underlying travel surveys had no direct experience with autonomous vehicles, so novelty effects, safety concerns, and trust-building dynamics were not captured. In retrospect, pairing the quantitative modeling with even a small qualitative study, such as interviews or a stated preference survey with AV-specific scenarios, would have strengthened the behavioral validity of the findings. That's a design choice I'd make differently today.