Marketplace Reliability
I analyzed booking and cancellation records from an anonymized healthcare staffing marketplace. Late cancellations were concentrated in shifts booked 24 to less than 72 hours before start, so I used that window to define a pilot population and test a history-based eligibility rule.
Project provenance
This page is based on an anonymized portfolio edition of a take-home case. MarketplaceCo and Metro A are aliases. Source and cleaned row-level records are not published because redistribution permission was not provided.
Scroll horizontally to inspect the full workbook screenshot.
Overview
A late cancellation means a healthcare-professional cancellation recorded less than 24 hours before the scheduled shift start.
Across 41,040 Metro A shifts, 2,486 had a late professional cancellation, a 6.06% shift-level rate. No-call/no-show affected 1,253 shifts, or 3.05%.
Late cancellation became the primary outcome for the pilot. No-call/no-show stayed as a guardrail.
Finding the target window
I grouped shifts by how far ahead they had been booked.
Shifts first booked 24 to less than 72 hours before start had the highest observed late-cancellation rate: 205 of 1,505 shifts, or 13.62%.
Because some shifts had multiple booking events, I repeated the comparison using only shifts with exactly one recorded booking. The same window ranked highest: 163 of 1,427 shifts, or 11.42%.
Using history available at booking time
For each booking in the 24-to-72-hour window, I rebuilt the professional's prior history using only shifts that had already resolved before that booking. The booked shift was excluded from its own history.
The target-window analysis covered 1,742 booking decisions across 1,701 shifts and 393 healthcare professionals.
With at least 10 prior resolved shifts and a prior late-cancellation rate below 10%, 589 decisions met the screen. Forty-four later had a same-professional late cancellation, a 7.47% historical rate. The other 1,153 target-window decisions had 181 late cancellations, or 15.70%.
589 decisions · 44 later late cancellations
At least 10 prior resolved shifts and prior late-cancellation rate below 10%1,153 decisions · 181 later late cancellations
Historical comparison groupHistorical comparison only; not an estimated treatment effect.
Choosing the starting rule
I compared minimum-history requirements of 5, 10, and 15 prior resolved shifts.
The 10-shift rule covered 589 decisions, or 33.81% of the target-window sample, at a 7.47% historical late-cancellation rate. Five shifts increased coverage to 43.57% at an 8.04% rate. Fifteen shifts reduced coverage to 27.67% at a 7.88% rate.
I used 10 prior resolved shifts as the pilot starting rule. It kept roughly one-third of the target-window decisions without narrowing the group as far as the 15-shift rule.
Scroll horizontally to inspect the full workbook screenshot.
Designing the pilot
The proposal uses a 16-week randomized Metro A pilot for open shifts entering the 24-to-72-hour window.
Professionals meeting the history screen receive 30 minutes of preferred access. If the shift remains unfilled, it returns to the wider qualified pool.
The shift follows the current booking process.
The 3.5-percentage-point threshold is a pilot-design assumption. The historical 7.47% rate is a benchmark, not a forecast.
Limits
The historical comparisons are observational. They were used to choose what to test, not to estimate the effect of preferred access.
The supplied records may omit earlier professional activity, and some shifts have multiple booking events. The analysis is also limited to Metro A; another market would need its own baseline and history check.
Reflection
The main shift in this project was moving from a descriptive pattern to a testable operating rule.
I had to separate information that was available at booking time from information that appeared later, choose how much prior history was enough for a starting rule, and define outcomes that could not be allowed to worsen during the test.
The historical analysis narrowed the question. The randomized pilot would answer whether the rule actually improves reliability.