Q-Logos

How to Read ETA Percentiles: Median, p90, p95 and What a Simulated Arrival Range Tells You

A plain-language guide to ETA percentiles: why one arrival time hides a spread, what the median and the high percentiles mean, how a Monte Carlo simulation produces them, and how far to trust a range that comes from a synthetic noise model.

One arrival time hides a spread

A single ETA looks precise. It is one number, often down to the minute. But a trip does not take one fixed amount of time. The same stops in the same order can take longer on one day than another, because of small delays that nobody can list in advance: a slow junction, a blocked loading bay, a long wait at a door.

So the honest picture of a trip's duration is not a point but a distribution: a spread of possible durations, some more likely than others. A single ETA is one summary of that spread, and it throws away the part you most need when being late matters, which is how wide the spread is and how far the late side reaches.

Percentiles are the standard way to put that missing part back.

What a percentile is

Imagine the same trip made a large number of times. Write down each duration and sort the list from shortest to longest. A percentile is a position in that sorted list.

  • p50, the median. The middle of the list. Half of the trips are shorter and half are longer.
  • p90. Nine trips in ten are at or under this duration. One in ten is longer.
  • p95. Only one trip in twenty is longer.
  • p99. Only one trip in a hundred is longer. This is close to the late edge of the spread.

Two things follow straight from the definition. The percentiles always come in order, so p50 is never above p90. And the gap between them tells you about the shape of the spread: if p90 sits close to p50, trips are consistent; if it sits far above, the late side has a long tail.

Median versus average

The median is not the same thing as the average. The average adds every duration and divides by the count, so a few very long trips pull it upward. The median only cares about the middle position, so it barely moves when the worst trip gets worse.

Trip durations are usually lopsided. There is a floor on how short a trip can be, and no real ceiling on how long. That shape pushes the average above the median. When a tool gives you both, the gap between them is a quick sign of how lopsided the spread is.

Why planning to a high percentile is the cautious choice

If you promise a customer the median, you should expect to be late about half the time. That is what the median means. Whether that is acceptable depends on what lateness costs.

Planning to a high percentile turns the question around. Instead of asking "when will I probably arrive", you ask "what time can I commit to and miss only rarely". Planning to p90 means accepting that roughly one trip in ten runs over. Planning to p95 tightens that further, at the price of more slack on ordinary days.

The trade is simple and worth stating plainly:

  • A higher percentile means fewer late arrivals and more idle time.
  • A lower percentile means a tighter schedule and more late arrivals.
  • No percentile removes lateness. It only sets how often you should expect it, and only if the spread behind the numbers resembles real trips.

That last condition is the important one, and it is where most misreadings begin.

Where the spread comes from: Monte Carlo in one paragraph

A planner rarely has thousands of recorded runs of your exact trip. So it builds the spread by simulation. A Monte Carlo simulation takes one estimated duration, replays the trip many times, and on each replay adds random variation drawn from a chosen noise model: rules that say how large the random delays can be and how they are distributed. Each replay yields one simulated duration. Sort them, read off the positions, and you have percentiles. The method is sound, but notice what it does and does not contain: the randomness is whatever the noise model says it is. If the model was fitted to recorded trips, the percentiles reflect those trips. If the model is a set of chosen ranges, the percentiles reflect those choices.

Three questions to ask of any ETA range

Before you lean on a p90, ask:

1. What is the starting estimate? Percentiles are spread around a base duration. If that base ignores the road network, the whole range inherits the error. The page on straight-line versus road distance explains why that matters.

2. Where does the variation come from? Recorded trips, live feeds, or a synthetic noise model. Only the first two contain information about the real world on the day you travel.

3. Which route does the range describe? If a tool shows several orderings, check whether the range belongs to one of them or to each.

A tool that answers these on the page is easier to trust than one that prints a range with no explanation.

A worked example: what the Q-Logos route page shows

The Q-Logos web route planner is a useful example because it answers those three questions in its own output. Here is what the current live version does, and nothing beyond it.

The starting estimate is modelled. The planner orders your stops on straight-line distance and derives an estimated duration from that model. It is not a measured drive time. The page on how a multi-stop route planner works covers what the planner computes and what it leaves out.

The variation is synthetic. After the ordering step, the server runs a Monte Carlo simulation on the estimated duration. Each simulated trip is that duration plus random variation: one random term that scales with the length of the trip, and two further random terms that the server's code names weather and traffic. Those names describe the role each term plays in the model. They are draws from fixed ranges, not readings. The simulation consults no traffic feed and no weather data. The ranges lean toward delay rather than being centred on zero, so the median of the simulated trips usually sits above the single estimated duration on the route card. That is a design choice in the model, not a measurement.

The planner also lets you pick a traffic level, and it may apply the current weather at your first stop. Those settings adjust the estimated duration by a fixed planning rule before the simulation starts. They do not make the random terms in the simulation any more real.

The percentiles are printed under the results. When the server sends them, the route page shows a line labelled ETA percentiles with p50, p90, p95 and p99 in whole minutes. If the server does not send them, the line is absent; the page does not fill in a figure of its own. The percentile line describes the first of the three orderings the planner returns, not all three. Because the random draws behind it are not fixed in advance, the line can shift slightly if you run the same stops again. The page on simulated annealing for route ordering explains why the orderings themselves can also differ between runs.

The on-time confidence figure comes with its definition. Each route card carries a figure labelled On-time confidence. It is the share of simulated trips that finish within that route's estimated duration plus a buffer, under the same noise model. Unlike the percentile line, this figure uses a random source tied to the estimated duration, so the same estimate gives the same figure. The server sends a sentence that defines the figure, including how many simulated trips it used and the ranges of the random terms, and the route page prints that sentence under the results. Read it there rather than taking the figure on its own.

How many simulated trips sit behind the percentile line depends on your plan; the pricing page lists it.

How to read the Q-Logos numbers in practice

  • Treat the percentiles as a description of the model's spread around a modelled duration. They show how the planner's assumptions play out, not how your streets behave.
  • Use the gap between p50 and p90 as a reminder to add slack, not as a measured late-arrival rate.
  • Do not read the on-time confidence figure as a delivery success rate. No recorded deliveries are behind it.
  • If the base duration looks wrong for your area, for example because of a river crossing or limited-access roads, the percentiles are wrong by at least as much.
  • For a commitment to a customer, check the trip against your own experience of the roads. The planner does not model road networks, turn-by-turn directions, delivery time windows, vehicle capacity or more than one vehicle.

The short version

A single ETA is the middle of a story. Percentiles tell the rest: the median is the typical trip, the high percentiles are the time to allow when lateness is costly, and planning to a high percentile is the cautious choice. But a percentile is only as informative as the spread it was read from. Ask what the starting estimate is, where the variation comes from, and which route the range describes.

To see a percentile line and its definition on your own stop list, create a free account and run the planner. If your stops live in a spreadsheet, the page on the CSV import path explains what the planner reads.

Create a free Q-Logos account, run a stop list, and read the ETA percentiles and the server's own definition of its on-time confidence figure under the results.

Get started