From New York City to Eugene, bike sharing systems are more visible than ever. In just one example, in 2015, the Bay Area bike share system, known as Ford GoBike, included 70 stations spread across five cities. It has now expanded to 540 stations. Nationwide, shared bicycle ridership has grown dramatically in recent years, from over a quarter-million in 2010 to nearly 30 million in 2016.
Bike sharing programs allow customers to rent a bike on a short-term basis. The systems comprise networks of stations at which the bikes are docked. Customers check out a bike then drop it off at another station near their destination. Most systems are equipped with electronic sensors on both the bikes and the stations that make collecting granular data easy and inexpensive.
The use of bike sharing systems helps improve public health, lessens the environmental impact of transportation, and increases accessibility to other forms of public transportation. Pretty convenient, but Pradeep Pendem, assistant professor of operations and business analytics at the Lundquist College of Business, noticed a problem.
“Imagine if a customer comes to a station and doesn’t find a bike,” he said. The customer could find another station, often by using an app, or instead take another form of transportation.
“In the short-term, that’s fine,” he said, “but this experience could affect long-term ridership,” while creating negative word-of-mouth about the reliability of the bike share network as a whole.
Pendem saw this as a resource-allocation issue. He wanted to figure out how best to allocate resources in a network setting. His findings are outlined in his working paper, “Demand Estimation and Allocation of Bikes Using Real-Time Trip and Inventory Data to Maximize Ridership in Bike-Sharing Systems.”
Pendem reasoned the first step in effectively allocating resources in a network setting was to calculate true demand. Easier said than done in a bike sharing system. If 10 people go to station X but only eight people find bikes—forcing two customers to choose either a different station or another mode of transportation—the demand data at station X reflects eight riders, but in reality, it was 10.
“The data we see is not really a true demand,” Pendem said. “Unless you take a bike from one station to another station, that information is not tracked. The information you see in real-world data is always censored.”
“That’s the primary aspect of what my paper focuses on. How you estimate that 10?” he said.
The key to solving this issue, Pendem found, was to look at where the bike share stations are placed and how far apart they are located.
Pendem has built a stochastic (randomly determined) demand model which covers customers choosing between neighboring stations or another form of transit when faced with no bikes upon arrival. Using mathematical models and complex calculations to crunch a vast amount of data, Pendem’s model calculates the optimal number of bikes to be allocated across stations in the network to increase ridership and satisfaction.
Ultimately, he hopes his research will provide a tool that can be used to better forecast and determine bike share demand.
Preliminary results are encouraging, he said. Using a massive dataset from the Bay Area bike sharing system, Pendem’s model predicts an 8 percent increase in ridership and a 2 percent increase in satisfaction. Of course, there are more datasets to test and analyze to see if his model and its results are applicable to other bike share systems in other locations. But for his part, Pendem is confident his approach has merit.
“If you don’t incorporate the process that I use in the study to estimate demand, you are going to underestimate demand in some stations and overestimate demand in other stations,” he said. “Predicting incorrect demand at stations leads to incorrect resource allocation and hence lower ridership and customer satisfaction.”
—William Kennedy