On the Monday the dashboard recommended cancelling the children's sessions, the library room was almost empty. Outside, though, a metal barrier took over half the pavement, the bus terminated two neighbourhoods early, and a temporary sign could not be seen from the underpass exit. For months, water-main work had closed the walking route many families used. The dashboard recorded none of it. It showed a falling attendance curve and an apparently sensible verdict: “forecast demand is too low.”
On a spreadsheet, the proposal looked efficient: remove Wednesday story time, reduce reading-support sessions, and move the budget to an activity with fuller seats. The problem was not that someone had produced a forecast. The problem was allowing a forecast built from an abnormal period to remove opportunities before checking whether families still had a reasonable way to use them.
Forecasting is not the same as deciding in advance who deserves a service. A model can organise signals, flag change, and help people ask better questions. It cannot, on its own, pronounce that a neighbourhood no longer needs an activity when its data include roadworks, disrupted transport, or hours that have become impossible. That distinction matters because cancellation is not neutral: it reduces chances to read, play, and meet, then lets the lower attendance created by that withdrawal look like a preference.
When a number records a barrier#
Attendance is observable, but it is not a complete explanation. Twenty empty chairs may mean the format holds little interest. They may also mean the only bus arrives late, the station lift has been out of service, or crossing a construction zone with a small child has become unsafe. The same count can arise from very different causes. Treating it as a stable expression of desire turns a temporary condition into a label for the people who live nearby.
Attendance is observable, but it is not a complete explanation. Twenty empty chairs may mean the format holds little interest.
In this case, the dashboard used eighteen months of registrations, cancellations, and actual attendance. That was a long window, but it included ten months of diverted buses and five months when the main entrance was partly closed. It also counted a cancellation sent by text in much the same way as a place that was never booked. It could not distinguish a family that lost interest from one that stopped attempting the journey after travel time rose from twelve minutes to forty. A forecast can be internally consistent while answering a question that is much too narrow.
This cannot be fixed with a decorative footnote saying “interpret with care.” The team needs to state which conditions changed, when they changed, who they affected, and how they appear — or fail to appear — in the data. The analyst should be able to identify that route 6 no longer stops close to the building, that the works made the entrance harder to find, and that winter sessions coincided with occasional closures. That context sheet does not discredit the model. It gives the model the ground it needs so that it does not mistake an obstacle for a preference.
The library should also ask what its success measure is actually measuring. If the goal of a children's library is simply to fill a room on a particular afternoon, a headcount may be enough. If the purpose is to make a nearby place where children can read, play with language, and meet other people, the record must be broader. It needs travel distance, reported barriers, repeat participation, and the alternatives families genuinely have. A metric decides what becomes visible. It should therefore never be the sole argument when a local public resource is at stake.
Letting the neighbourhood answer back#
Before withdrawing an activity, the library can hold a short, focused conversation. It does not need an abstract survey about whether people “like” the programme. It needs to hear from people who are not getting there: families who cancelled, the nearby school, the residents' association, front-desk staff, and people familiar with the altered routes. The questions should be about the obstacle. At what point does the journey become too long? Which day is unworkable? What information was missing? What change would make a return possible?
Qualitative answers are not sentimental decoration around serious data. They are evidence about mechanisms the administrative record did not capture. A parent may say that her child loved the session but that the diversion required walking through an unlit street. A school liaison may explain that recently arrived families do not know there is a side entrance. A librarian may have noticed attendance fall whenever the passage is blocked and recover when an alternative route is explained. None of those accounts automatically replaces the data series. Together, they make it possible to understand what the series means.
The conversation should return to the team with a traceable record. A simple table can separate what the model estimates, what has been observed locally, what remains uncertain, and what decisions should not yet be made. For example: forecast of eleven participants; disruption on the usual route confirmed; longer travel time reported by twelve families; no evidence yet about demand after reopening. Writing this down makes a useful fact unavoidable: “low demand” is still a hypothesis, not a verdict.
Test access before cutting provision#
The responsible alternative is not to keep every activity forever without examining results. It is to run a limited, checkable access experiment before removing the service. For six weeks, the library could retain one weekly session, move it temporarily to a time that connects with the school run, and mark a safe path from the temporary bus stop. It could also allow booking by phone or in person, because a confusing booking platform is another barrier that a dashboard commonly reads as lack of interest.
The experiment should say in advance what it is meant to learn. Any rise in attendance is not automatically a success. The team should compare arrival, repeat visits, reasons for absence, and the effort the change asks of families. A public one-page record can name the trial dates, access changes, responsible person, information sources, review threshold, and decision date. This prevents the pilot from becoming a vague promise or a single number from later being used to support the opposite conclusion.
The comparison group does not have to be a community denied an activity. It can be the service's own history before the works, a different time slot, or the difference between people who receive information about the alternative route and those who do not. It examines the institution's access conditions instead of treating families as a behaviour that needs to be corrected.
During the trial, someone with authority must be able to make a quick correction. If signage disappears, another transport change lengthens the trip, or the time clashes with school collection, waiting for a monthly report makes no sense. Oversight means having room to act on a specific condition. The team can publish short adjustments, record why they made them, and look again at the results. That memory makes it less likely that the next construction project will produce the same interpretive mistake.
Efficiency must not become a quiet withdrawal#
Libraries have real obligations to manage rooms, staff, and budgets. The choice is not between data and local knowledge; it is about which kind of efficiency matters. Closing an activity may save money today and make it more expensive tomorrow to rebuild a relationship with families who no longer count on the service. The cost is uneven as well. A family with a car, spare time, and another library nearby may find a workaround. A family dependent on the disrupted route loses an opportunity without having chosen to do so.
For that reason, the forecast should appear as one part of a conversation, alongside its uncertainty and assumptions. The dashboard can say, “with the available data, we expect lower attendance if conditions remain unchanged.” It should not say, “this neighbourhood no longer wants children's activities.” The first statement invites investigation and practical tests. The second fixes a label and makes it easier for an access failure to become permanent.
A useful safeguard is a no-withdrawal rule during events that distort a series: prolonged works, transport disruption, school closures, or major schedule changes. The rule does not forbid programme review. It requires that any reduction be supported by evidence gathered after the disruption and by an accessible alternative. It should also identify who signs the decision, what appeal is possible, and when the decision will be reconsidered. Accountability becomes real when it has names, dates, and a route to correction.
The opening scene provides a straightforward test. Once the barriers are gone, the stop returns, and families know a usable route, will the dashboard still say the activity is unnecessary? If nobody has tried to find out, removing story time is not a conclusion about demand. It is a bet placed on a neighbourhood after obstacles have been put in its way. The library can use a forecast to direct attention, but it must retain the responsibility to look beyond the screen, hear what the data leave out, and test an improvement before closing a door.
