How many kilometres a year is normal, and what to expect

A used car's recorded mileage tells you its past, but the cost of owning it depends on its future. Two cars sitting at 60,000 kilometres today can sit at wildly different figures three years from now, and that fork in the road decides maintenance bills, depreciation and resale. Our mileage forecasting model turns the question of how hard this car will be driven from here into a probability distribution rather than a guess.

Why a single number is the wrong answer

Asking for one predicted figure invites false precision. Driving patterns are noisy, so we output a forecast range with a confidence band, and that band itself is informative: a wide band flags a vehicle whose usage history is erratic and therefore harder to price, while a tight band signals a car whose future is genuinely predictable.

The model blends the car's own mileage trajectory with cohort behaviour, weighting recent intervals more heavily because the most recent owner's habits dominate the near future. A car that did 30,000 kilometres a year under a sales rep but has crawled along under a retiree for the last two years is forecast on the retiree's pattern, not the headline lifetime average.

The features that carry the most weight

After feature-importance analysis across the training set, a short list of inputs does most of the work. Each one is something a buyer can reason about, which keeps the forecast interpretable rather than a black box you are asked to trust on faith.

• Historical annual average derived from inspection and service intervals.
• Body type and powertrain, which correlate strongly with typical use.
• Regional driving norms, since rural cars accumulate distance faster.
• Recent trend slope, capturing whether usage is rising or tapering.

These inputs interact in ways a flat average would miss. A diesel estate in a rural region with a rising recent trend is a high-mileage forecast almost regardless of its current reading, while a small petrol hatchback in a city with a tapering slope points firmly the other way.

How we validate the forecast

We hold back a slice of vehicles, predict their next year, then compare against what actually happened. On the held-out set the median forecast lands within 11% of observed mileage, and we publish that error rather than hiding it, because a model that will not state its own accuracy is asking you to trust it blindly.

A forecast you cannot challenge is just an opinion in a confident font. We would rather show you the error bar than pretend it does not exist.

Where the model is deliberately cautious

For imports, very new cars and vehicles with sparse histories, the model widens its band rather than forcing a number. A thin record genuinely is harder to forecast, and saying so honestly is more useful than a precise figure built on almost nothing.

Forecasts as a fleet tool, not just a single-car one

The same model that prices one car scales to a whole fleet. A business comparing two candidate vehicles for a high-mileage role cares less about today's odometer than about projected accumulation, because that is what drives servicing schedules, residual value at disposal and the point at which a vehicle stops earning its keep. Aggregated across a fleet, an accurate mileage forecast turns replacement planning from a guess into a schedule.

What the forecast changes for a buyer

A low-mileage car that our model expects to be driven hard is a different proposition from one likely to stay gentle. The forecast feeds directly into our total-cost estimates, so you see future wear reflected in expected maintenance rather than discovering it later in a workshop invoice.

It also reframes the classic low-mileage premium. A car is only worth paying up for its low reading if the future is likely to stay low too; if the forecast says the next owner will pile on the kilometres, that premium is paying for a past that is about to end.