Why we show a window,
not a date.
Every number on this site comes from one measured dataset and one model. This page explains exactly how that model works, including the parts where it is weak.
What is actually being predicted
Gran Turismo 7's Legend Cars dealer holds a small line-up of classic cars that rotates continuously. A car appears, stays available for a while, sells out, and is replaced. Some time later it comes back. There is no published schedule: Polyphony Digital's selection logic is opaque, and nothing in the game tells you when a specific car will return.
What we can do is measure. Every appearance is recorded with its date, so for each car we hold a series of intervals: the gaps, in days, between one appearance and the next. That series is the raw material for every forecast on this site, and what it shows is a process that clusters loosely around the two-to-three month mark without ever settling on a fixed period.
That is the whole problem. A process that scatters across weeks cannot honestly be reduced to a single day.
This page previously stated that a car essentially never returns within 14 days of leaving, and called that a hard fact about the game. It was neither hard nor a fact. Our appearance log contains no gap shorter than 14 days because the importer discards them, so the log could never have shown otherwise, and citing it as evidence proved nothing at all.
The raw daily snapshots, which are not filtered that way, show the opposite. In June 2026 the Pontiac GTO ‘The Judge’ '69 left the dealer on the 1st and was listed again on the 6th; the Plymouth Superbird '70 was gone for six days and the Suzuki V6 Escudo '98 for eight. On each of those days the dealer was otherwise full, so these were real removals, not gaps in our own record.
What the 14-day rule actually is, and all it ever was, is a modelling choice of ours: a car that disappears briefly and comes straight back is treated as one interrupted stay rather than a new rotation, because counting it as a fresh five-day cycle would drag that car's median far below its real rhythm. That choice is defensible. Presenting it as a law of the game was not.
The case against a precise date
Earlier versions of this site published one estimated return date per car. It was the obvious thing to do and it was quietly misleading. Measured against what actually happened, that single date missed by about 15 days on average. A date is a promise the data cannot keep: it looks certain, it is usually wrong by a fortnight, and when the car fails to show up on the day, the whole forecast reads as broken.
The fix is not a better date. It is admitting that the answer is a range, and then being careful about how wide that range has to be.
Step one: the centre
Each forecast starts from an expected interval, blending the car's full-history median with its most recent completed cycle:
expected = round( median × 0.7 + latest × 0.3 )
The median is the stable backbone. It shrugs off the occasional six-month disappearance that would drag an average around. The 30% weight on the latest cycle adds a light touch of recency, because walk-forward testing showed it measurably reduces error. The result is floored at 14 days, our own threshold, not the game's, for the reason given in the correction above.
This centre still exists behind the scenes, but the site does not lead with it on purpose. It anchors the maths; it is not a date you should plan around.
Step two: the width, earned not assumed
This is where most trackers would pick a round number, plus or minus a week say, and apply it to everything. We do not, because a fixed width is a fiction: some cars rotate like clockwork and some wander badly, and giving them the same window misrepresents both.
Instead the width is derived from each car's own record of being wrong. For every past cycle we compute how far the actual interval fell from what would have been expected, producing a distribution of that car's own misses. The window is then drawn from that spread. Two windows are drawn from it:
The likely period uses the median miss: the tight band where returns cluster most. The calibrated range uses the 0.80 mark: the wider, honest band that is meant to actually contain the return.
Cars without much history would produce nonsense quantiles from three or four data points, so their spread is shrunk toward the typical miss observed across the whole dealer, weighted by how much history they have. A car with a dozen recorded cycles is trusted almost entirely on its own evidence; a car with two borrows most of its uncertainty from the population. Finally, guardrails trim only pathological tails, and they are expressed as a fraction of each car's own cycle rather than a fixed number of days, so a fast-rotating car and a slow one are never forced into the same box.
The consequence is visible across the site: a consistent car gets a narrow window because it has earned one, and an erratic car gets a wide one because anything narrower would be a guess dressed up as a forecast.
Leaving the dealer is a different question
Predicting when a car returns and predicting when it leaves are not the same problem, and they do not share a method. Departure is far more tractable. Once a listing stops being purchasable it spends exactly one dealer refresh in a sold-out state before its slot is replaced, a pattern that holds in essentially every departure we have observed, with the handful of exceptions all falling on days when our own sync failed to run rather than on any change in the game.
Where GT7Info's public feed exposes a car's remaining availability, we use that number directly: it is a fact, not an estimate, and it is never widened into a window. Only when that feed is unavailable do we fall back to estimating from observed stay durations.
What this model cannot do
It cannot see intent. If a car is deliberately brought back early for an event, a promotion or a Special Pick, nothing in its interval history contains that information, and no amount of calibration will recover it. We recently watched a car return after 33 days when its own 22-appearance record pointed to roughly 79, outside not just the likely period but the full calibrated range. That is not a tuning failure; it is a car that broke its own pattern, and the model is structurally blind to it.
It is also only as good as its inputs. Cars with one or two recorded appearances have windows that are mostly borrowed from the population, and their pages say so rather than projecting false confidence.
Everything here is a likelihood, never a guarantee. The honest question is not whether the forecast is ever wrong. It is how often, and by how much. We publish that answer in full on the accuracy page.