An arrival time should come with the odds of missing it

I think carriers should stop sending shippers a single arrival time. My position, after twenty-nine years of making and receiving such estimates on water and on the road, is that every arrival forecast a carrier sends should include a window it expects to hold for a stated share of arrivals, that the carrier should publish how often its windows actually held, and that every human override of a forecast should be recorded and reported to the same audience. A carrier that does those three things will look less precise for a season and become more useful for good.

This is not a technical argument about forecasting models. Most of the models are adequate. The difficulty is that a single number invites everyone downstream to treat it as a commitment, while allowing the carrier to treat it as a hope.

What a single time conceals

At Morvaise we carry bulk cargo on the Great Lakes and the St. Lawrence. A laker bound for a grain elevator near Quebec City may pass through the Welland Canal and the Seaway locks, cross two lakes and wait for a berth. Each step carries its own uncertainty. A queue at the canal in October can add a day; a north-easterly on Lake Ontario can add another. When we sent a single estimated time, we compressed all of that into a figure that looked exact to the hour.

The terminal receiving the vessel then made exact decisions with it. It called a labour gang for a shift, ordered rail cars for the outbound grain and told its own customers when their product would move. When we missed, the terminal bore most of the cost, and we saw only the fraction that came back as a waiting-time claim.

When I analysed our 2021 season, the median arrival came nine hours after the first estimate we had sent. Nearly a third of arrivals missed their estimate by more than twelve hours. The errors were not random. First estimates were set by the commercial desk, where an optimistic time could help win a cargo, and they were revised later much more often than earlier. Nobody set out to mislead anyone. A single number had simply given an optimistic organisation nowhere to put its doubt.

A window gives the doubt somewhere to go. Since the 2023 season we have sent a likely arrival time and a window we expect to contain the actual arrival nine times in ten. In calm summer weather on a short route the window may be six hours wide. On a long voyage in November it can reach thirty. The width is information in its own right: it tells a terminal how much slack to plan for, and it tells us where our forecasting is weakest.

Judge the window by how often it holds

The measure that matters to a shipper is not how close our likely time came on average. It is calibration: whether a window we describe as holding nine times in ten actually holds about nine times in ten. A forecaster who is right on average but whose windows hold only six times in ten is misleading the people who plan against them, however good the average looks.

So we publish calibration. Every shipper on our portal receives a monthly report showing, for its own routes and for the whole fleet, how often arrivals fell inside the window, how often they came early and how often late. In the 2025 season the fleet figure was 88% across 1,510 arrivals. That is slightly short of the design, and the report says so. It also names the routes where we fall furthest short, which in 2025 were the long autumn voyages from Lake Superior.

Publishing the record changed our internal conversations more than our external ones. When a forecast missed, the question was no longer whether the customer had noticed. It was whether the miss belonged in the tenth of arrivals we had told customers to expect outside the window, or showed that we had misjudged a route or a season. That is a question a data team can actually work on.

There are fair objections. Some customers say they want one number and will plan with it whatever we send. Some will. But we found that terminal schedulers, given a window and a record, began to use both, and then asked for refinements. One asked for a firm time before which the vessel would not arrive, because that was the figure that decided when to call labour. Another objection is competitive: a rival quoting firm times looks more confident. It does, until its customers keep their own record of its arrivals, as the better ones already do.

Who may move a forecast, and how we find out

A forecasting system can be honest and still be overridden. Commercial and operations staff often know things the model does not: a berth that has just come free, a master's report of heavier ice than forecast. I do not want to stop them sharing that knowledge. I want it written down.

At Morvaise anyone with the authority may change a forecast, but the system keeps the original, the change, the person, the stated reason and, later, the actual arrival. Every month my team reports how overrides performed against the model, by person and by reason. In the first season the pattern was plain. Overrides citing operational information, such as a revised berth time, improved the forecast. Overrides with no stated reason, most of them entered the day before a customer call, made it worse. We disciplined nobody. We showed the figures, and unexplained overrides fell by two-thirds within a season.

I would make three requests of any carrier's leadership, on water or on land. Send a window with every arrival time, and say what share of arrivals it is designed to contain. Publish to your customers how often it held, by route and season, including the seasons you would rather not discuss. And record every override, so that you can learn whose knowledge improves the forecast and whose hope degrades it.

None of this requires a better model. It requires accepting that an arrival estimate is a statement about probability, and that the people who plan against it are entitled to know how good our probabilities are. A carrier that publishes its misses will lose some arguments with customers who wanted certainty. In my experience it wins the larger argument, which is whether its numbers can be planned against at all.