How San Francisco Schools Used Data Science to Let Teens Sleep In

San Francisco wanted to let teenagers sleep in. The science was clear, the law was on the books — and the logistics were a nightmare. How do you change start times across a school district, keep buses running and avoid ballooning the transportation budget?

That’s where a team of researchers came in. Working with the San Francisco Unified School District (SFUSD), Arthur Delarue from the University of Virginia’s Darden School of Business, along with co-authors Zhen Lian of Yale School of Management and Sébastien Martin of Kellogg School of Management, designed algorithms to find thousands of “reasonable” scheduling possibilities by sifting through more than a septillion options, then built them into a suite of interactive tools for policymakers.

The result: a data-driven redesign that allowed SFUSD to select new start times in the 2021-2022 school year, save more than $5 million in bus costs, comply with California’s new start-time law ahead of schedule, and give students the extra rest science says they need.

The findings appear in a forthcoming Management Science article, “Algorithmic Precision and Human Decision: A Study of Interactive Optimization for School Schedules.”

Why School Start Times Matter

Because of biological shifts in adolescence, most teenagers can’t fall asleep before 11 p.m., making early mornings a struggle. Yet thousands of American high schools still ring their first bell before students are fully rested — despite decades of evidence linking later start times to better health, mood and learning outcomes.

In 2019, California became the first state to mandate the school day start no earlier than 8 a.m. for middle grades and 8:30 a.m. for high schoolers.

For SFUSD, that meant a massive operational puzzle: shifting start times across 133 public schools without breaking the system that gets more than 50,000 students to class each day.

Lessons from Boston, Success in San Francisco

For Delarue, an assistant professor in the Data Analytics and Decision Sciences area at UVA Darden, the story began years earlier in Boston.

In 2017, as a Ph.D. student at MIT, he helped design an algorithm to optimize bus routes for Boston Public Schools — a project that took 50 buses off the road.

That success led to the next challenge: rethinking school start times across the district.

“That presents an operational question, which is: which school should start at which time?” says Delarue. “And that is a hard problem because it's impacted by things like the geometry of the district and where the students live and which buses are going to pick them up. So there's a connection between the start times and the bus routes.”

Boston Public Schools hadn’t changed its start times in decades, and, when you move a school from an 8:30 a.m. start to 7:30 a.m., for example, everyone notices. 

Even though the benefits of letting teens snooze a little longer were clear, the blowback was intense. Parents, community groups and city councilors complained. Two weeks after the Boston School Committee voted unanimously to approve the new schedule, the district announced it was shelving the plan.

Fast forward to December 2020 — the height of the COVID-19 pandemic — when the chief of operations for San Francisco’s public schools contacted Delarue: the district needed help overhauling school schedules to comply with Senate Bill 328 by fall 2022.

Leading up to the pandemic, depending on the campus, public school students in San Francisco could start class at as many as 18 different times.

“San Francisco had legal pressure to change their starting times because they were not in compliance, and they had an opportunity to do so because they were currently online for the 2020-2021 school year and were planning to reopen in fall 2021,” says Delarue. “It was an opportunity to start with a clean slate.”

It was also a chance for Delarue and his colleagues to apply what they had learned from the pushback in Boston to changing school start times.

“One of the lessons was that optimization is important — getting a schedule that won’t blow up your bus costs is really important,” says Delarue. “But what's more important is the ability for stakeholders who are not researchers, but who are experts in how their school district works, to explore what’s possible and what they want.”

Armed with their prior experience, and a new set of challenges, the researchers got to work.

“What we wanted to do differently from Boston was create solutions that were not necessarily ‘the best’, but simply ‘good enough’, that were not going to blow up the bus budget,” says Delarue. “And so we came up with a much simpler approximation of transportation. And what that meant was that we could, given other priorities, produce an optimized schedule in tens of seconds, versus 30 minutes in Boston.”

That allowed the researchers to narrow down the universe of possibilities to just a few thousand good options. Next, they made a tool — using a Google spreadsheet — that allowed policymakers to check boxes with various constraints and generate the best options.

By April 2021, SFUSD had new start times for its public schools.

“The district could assign schools to one of three start times (7:50, 8:40, 9:30),” the researchers write. “The solution aligned afternoon end times across many nearby elementary schools to allow for teacher professional development. The new start times granted the district nearly $5.5 million in annual transportation savings.”

What’s more, a survey of elementary school parents and staff a year after the change suggested a majority approved of the changes.

Beyond School Schedules

In their work with SFUSD, the researchers found that giving policymakers interactive tools they could use independently helped them identify their priorities and constraints far more efficiently.

“By introducing a simple mathematical model describing the interaction between researchers and policymakers, we found that this insight may apply in other operational settings,” they say.

“More precisely, we find that when two teams with complementary expertise work towards a shared goal, their dependence on each other’s progress can create bottlenecks. To mitigate these bottlenecks, it is more important to reduce the coupling between the two teams than to increase either team’s productivity. In practice, one takeaway is that it is more important to invest in better interactive tools than faster optimization algorithms.”

The forthcoming paper provides methodology, including both algorithms and decision support tools, that school districts can use to tackle the complex problem of adjusting school start times districtwide without incurring elevated bus costs.

“Our success in SFUSD relied in part on the observation that the problem is difficult not just mathematically, but also from a policy standpoint, where experts face many competing objectives that are not even well-defined, let alone prioritized,” the researchers say.

“In this setting, optimization is much more than a tool to find the ‘best’ solution — it helps policy makers understand what is possible and empowers them to find the best compromise.”

When Algorithms Meet Policy

At its core, the research asks how to make powerful optimization tools truly usable by policymakers — so human judgment and algorithmic precision can work together on the hardest problems.

There is an even bigger takeaway in the age of artificial intelligence, as many grapple with combining human expertise and algorithmic power.

“There are all these questions about how to balance human experts with AI agents,” says Delarue. “This is a very tangible example of designing a process that tries to make the most of the human skills and the algorithmic skills. Let’s let the algorithms excel where they excel, but then let’s let the humans have the input on all the things that matter.”

As he puts it, “Where else are there messy problems that need both computers and humans — and what’s the best way to get them to collaborate?”

Arthur Delarue is co-author of “Algorithmic Precision and Human Decision: A Study of Interactive Optimization for School Schedules,” with Zhen Lian of Yale School of Management and Sébastien Martin of Northwestern University’s Kellogg School of Management, forthcoming in Management Science.