Scooters where people want them: disentangling utilization from demand

blur-couple-daylight-2547627.jpg

by Evan Fields, lead data scientist at Zoba. Evan holds a PhD in Operations Research from MIT, believes aggressively in cities, bakes enthusiastically, and can be found on Twitter at @evanjfields.

Zoba provides demand forecasting and optimization tools to shared mobility companies, from micromobility to car shares and beyond.

At Zoba, we talk to lots of mobility service operators who are beginning to acquire data about their service that they want to leverage to improve operations. Typically, these conversations begin with a discussion of how the operator currently thinks about supply, demand, and vehicle utilization, so I thought I’d write a quick post about how we at Zoba think about demand and utilization. These related concepts are sometimes implicitly conflated, but it’s important to be clear about their differences.

For optimal operation of a mobility service, vehicles must be placed where users want to ride them. Accordingly, it’s important to distinguish between the demand for the mobility service and the utilization of existing vehicles. Since these words have subtly different meanings in different contexts, I’ll begin by giving the definitions we use at Zoba. Throughout this short post, I’ll talk about “rides” without loss of generality; the same demand and utilization concepts apply to reservation-based mobility services as well.

  • Demand is all the rides users would take in the absence of capacity constraints, i.e. the set of rides that would be performed if any user could get any vehicle any place any time. Typically, demand is measured in units of rides wanted per hour in a given region.

  • Utilization is all the rides users actually take, and is thus constrained by the available vehicles on which to take those rides. For business reasons, utilization is typically measured as the fraction of time that a given set of vehicles is used. E.g. if a given vehicle has a utilization of .25 on a given day, that means the vehicle was used for .25 * 24 = 6 hours that day.

Notice how utilization depends on the supply of available vehicles, but demand can happen anywhere! Intuitively, people can only take rides — and as data scientists, we can only observe rides — in places where there are vehicles to serve those users. In general, there are two reasons why demand differs from utilization (after appropriate unit conversions):

  • Censoring occurs when a user can’t find a vehicle they want at a time and place they want; that user’s demand is lost. This includes cases where there are no vehicles in an entire region, but people would use vehicles if they were there!

  • Substitution occurs when a user’s first-preference isn’t served by any available vehicle, but the user will accept a substitute vehicle instead. Typically the user accepts a vehicle nearby to their preferred location, creating spatial spillover.

As a consequence, if an area is totally saturated with vehicles, then the rides users actually take — the utilization — are an excellent proxy for demand. And this does happen in practice! For example, our analysis (stay tuned for forthcoming posts) shows that some football-field-sized regions of Austin, Texas average dozens of vehicles available at a time.

When thinking about demand and utilization at a high level, the intuitive picture to keep in mind is the following:

demandutil (1).png

When demand for a mobility service is zero, the vehicles comprising that service have utilization of zero as well. As demand rises towards infinity, the utilization also rises, but utilization is a concave function of demand; utilization can never exceed 1.

Consider why this figure and the associated intuition only applies to an entire mobility service, rather than any set of vehicles, such as those in a particular area. The vehicles in a given area may have utilization greater than zero, even if there is no demand in that area — if demand spills from adjacent areas! For example, suppose there are two regions as depicted in this cartoon:

twozone (1).png

Lots of users want to start rides in the left region, but there are no vehicles there. Many of these users will decide not to travel or use some other transportation mode — from the perspective of the mobility service, their demand is censored — but some users will walk to the right region and start rides there. Thus, even though the right region has no demand, it still can have nonzero utilization. Likewise, the left region has no utilization but does have demand!

In practice, the mismatch between demand and vehicle allocation is not usually so stark; mobility operators have some intuition about where users would like to start rides and can position their vehicles accordingly. But, it’s possible to do better than intuition. If we can estimate the demand for a mobility service, with data-driven methods and mathematical optimization we can discover vehicle allocations which optimally serve user’s desires — whether “optimal” means maximizing rides, maximizing revenue, maximizing accessibility, or any combination thereof.

The key here is estimating the demand: fleet operations should be optimized around what users want, not how users are already using the service. In short, disentangling the unobserved latent demand from the observed utilization unlocks more efficient mobility service operations.

Zoba is developing the next generation of spatial analytics in Boston. If you are interested in spatial data, urban tech, or mobility, reach out at zoba.com/careers.