Using Predictive Modeling to RevealHidden Curb Demand in Boston, MA
The City of Boston partnered with Turnstone on a 24-month pilot to answer a deceptively simple question: could a model built from parking payment data alone tell the city what is really happening at the curb? It mattered because payment data tells only part of the story. During the pilot, paid occupancy made Back Bay and Beacon Hill look barely half full, around 52 to 55%. The model, which accounts for the drivers who never pay, underpay, or overstay, showed the same curbs running at 78 to 88%. The neighborhoods were not underused. The data the city had been looking at was just incomplete.
Boston's parking pressure has kept rising even as car ownership has fallen, and the most common complaint the city hears is simply that there is not enough parking. The hard part is knowing whether that is true on any given block, because the city could not see real occupancy without sending people out to count it. An earlier performance-parking effort relied on manual counts and sensors, and a more ambitious demand-based pricing program stalled in large part because collecting, analyzing, and reporting that data by hand was too slow and too expensive to sustain.
Underneath the cost problem was a measurement problem. Payment data, the one continuous signal the city already had, systematically understates real demand, because a meaningful share of the curb never shows up in a transaction. Reading payments as if they were occupancy would point the city toward exactly the wrong conclusions, lowering rates on blocks that were actually full.
Turnstone built Boston's occupancy picture from data the city already had, its existing meters, mobile payments, and parking inventory, with no new sensors in the ground. The pilot tested two models against each other: a simple one that treated payments as occupancy, and Turnstone's model, which adds the unpaid, underpaid, and overstaying vehicles that payments miss. To check them, both were validated against manual ground-truth counts, people physically counting parked cars block by block over several hours. The results were decisive. Turnstone's model predicted real occupancy with 86 to 89% accuracy, while the payments-as-occupancy model performed worse than simply guessing the average, it had no real predictive power at all. Once the model was trusted, the city no longer needed constant manual counts. Turnstone's drift monitoring watches for shifts in parking behavior and flags when a fresh sample is actually warranted, which in the pilot reduced the manual collection burden by roughly 90% while holding accuracy steady. Because the model runs on transaction data that is anonymous by design, with any personal information dropped on receipt, the approach fit Boston's open-data and privacy commitments without extra work.
The pilot's clearest lesson was how much a city cannot see when it reads payments as occupancy. At the evening peak, payment data accounted for barely over half of the vehicles actually at the curb, hiding up to 38% of real demand. That gap is the difference between a neighborhood that looks like it needs lower rates and one that is genuinely full. Validated against manual counts, Turnstone's model held 86 to 89% accuracy across the study, while the payments-only model was less reliable than a flat average, confirming that modeling, not payment data alone, is what reflects the real curb.
It was also far cheaper than the alternative. Continuous modeling cost under $0.50 per blockface per day, against roughly $100 for a single day of manual counting, while delivering data every day of the year rather than a one-time snapshot. For the city, the value was less about any single number and more about the conversation it made possible: with continuous, trusted occupancy data, Boston could respond to 'there is not enough parking' with what was actually happening block by block, engage residents and businesses with evidence, and evaluate changes as they happened instead of waiting on the next expensive study. Along the way the pilot also surfaced practical findings the city had not had visibility into, including a distinct high-demand core within Back Bay that behaved differently from its quieter edges, and a vendor data feed sending incorrect timestamps that had gone unnoticed until the model exposed it.
Boston's pilot was a study, not a deployment, and that is exactly what makes it useful. An independent city ran the method for two years, validated it against the manual counts it was meant to replace, and confirmed it was both accurate and far cheaper. The takeaway travels: the curb a city manages is not the curb its payment data shows, and seeing the real one, every day, is what lets a city manage it with evidence instead of guesswork.
