How we predict what a damaged vehicle will fetch

A motor total loss in Germany or Italy is settled as the pre-accident value of the car less the residual value of the wreck, paid in cash, with the claimant keeping the vehicle. The residual value is therefore a deduction from the payout rather than an asset the insurer disposes of.

Until recently that deduction was whatever the available bidders happened to offer in the days after the assessment. We predict it instead. This is how.

What the model targets

The target is the highest bid the vehicle draws across all residual-value platforms — not an average of them, and not one buyer’s purchase price. That is the figure a settlement is built on, so it is the figure worth predicting.

The distinction matters more than it sounds. Averages are stable and comparatively easy to model. Maxima are not: whether a particular vehicle reaches a high figure depends on whether the buyer who would have paid most happened to see it, which depends on which platforms it reached, in which week, and in which country. A meaningful share of the variation in what these cars realise is not about the car at all.

That is also why we model buyer economics separately in fourteen markets rather than fitting one European curve. Labour rates, parts pricing, realistic repair scope and post-repair resale all differ enough that a single continental average is the wrong object.

What it learns from

Realised outcomes. Every training row is a vehicle that actually went through the process and produced a real number, running from January 2023 onward.

One decision inside that is worth stating because it is invisible from outside. An early version of the training set included records where the residual figure was an appraiser’s estimate rather than a market-settled outcome. They were removed. They looked like clean data, they improved the apparent fit, and they were teaching the model to predict the wrong thing.

Judgements like that — which rows are actually evidence — are most of the work, and they do not transfer with a data file.

Why it is not one model

Vehicles are routed by economic profile rather than by vehicle type — a car worth €3,000 and a car worth €22,000 are not the same problem scaled, and neither is a vehicle worth repairing versus one that plainly is not. A single fitted curve across all of them is wrong in a predictable direction on each.

There is a second split above that. Own-damage and liability claims get separately trained models, even though they target the same quantity. In liability, very few of these vehicles are actually sold, and bidders know it — so bids behave like options rather than commitments, and they run high and unreliable. One model serving both lines would be wrong in the same direction on every liability claim.

How we test it

Out-of-sample, which is the only test that means anything. A block of recent claims is removed entirely from the training data, the model is fitted on what remains, and it is then scored on claims it has never seen.

For the liability model, the holdout is a recent month of claims the model had never seen. We report four things from it: the share of vehicles predicted within 20% of the settled value, the mean and median absolute error in euros, and the portfolio bias.

Two of those matter more than the headline hit rate. The median error is the one to compare against the difference the product exists to recover — if the typical miss is larger than the typical gain, the prediction is not useful, whatever the hit rate says. And the bias tells you whether the model is systematically flattering itself in one direction; ours sits close to zero, which is what you want and is not automatic.

We report all of it by segment as well as in total, because a portfolio figure that averages away its own weak tail is not evidence. The low-value segment is where the model is least accurate, and we would rather say so than have it discovered in a pilot.

The figures themselves go to insurers evaluating the model, alongside a scoring run on their own history — which is a more useful test than any number we could put on a website.

What it does not do

It does not replace market price discovery in a disputed settlement. A prediction with this error band is accurate enough to set an expectation, to triage which vehicles are worth listing, and to tell you when a bid set has come in short. It is not accurate enough to be the last word on its own.

It does not score every vehicle. There is an eligibility range — very low-value vehicles are one example, electric vehicles another — and a claim outside it is returned as not scored rather than answered with a number we do not believe. We would rather decline than be confidently wrong on a file an insurer is about to settle.

It is reported in euros, per segment, not as one percentage. A single portfolio percentage is dominated by the cheapest vehicles and says almost nothing about behaviour on a €15,000 car. What decides whether the prediction is useful is the typical error in euros against the difference at stake on that claim, so that is what we report — segment by segment, with the hit rate beside it.

What happens after the prediction

A number an insurer cannot act on is an opinion. Under German practice an insurer cannot use a valuation to challenge a residual figure — only a concrete, binding purchase offer will do that.

So where an insurer wants to act on our figure, we make that offer ourselves: a fixed price, free collection from wherever the vehicle stands, payment on collection. If it is accepted, we buy the vehicle, and whatever it realises afterwards is our result. The insurer bears no price risk and no disposal risk.

That is the part that makes the model commercially real rather than merely interesting. Predicting the number is the easier half.