How the Legend Cars model works.

01 · Data

What the tracker observes

The current dealer is a live state: availability, price and, where supplied, days remaining come from the latest dealer feed. Historical dates form a separate ledger. The Legend Cars ledger records dated dealer appearances and the observed availability period.

A live fact always wins over a model estimate. If a car is currently listed, the page says so; a return forecast only describes what may happen after it leaves.

02 · Appearances

Turning daily snapshots into cycles

An appearance starts when a car is absent in one observed dealer state and present in the next. The interval is the number of calendar days between two genuine starts. Gaps shorter than 14 days are discarded as duplicate or reconciliation records, not treated as a new rotation.

This distinction matters: the model predicts a return, not a sold-out transition or an edited snapshot. Missing source periods remain uncertain rather than becoming invented appearances.

03 · Expected return

The centre is a cadence estimate, not a promised date

expected interval = round(full-history median × 0.7 + latest interval × 0.3)

Legend Cars use the full-history median for stability, with a small pull toward the latest completed cycle for recency. The median resists an unusual early or late return better than a simple average.

Cars without enough usable history receive the dealer-wide fallback of 89 days. That is an explicit population estimate, not an invented personal pattern.

04 · Windows

Why a range is more honest than a date

GTRBit measures each car’s historical miss around its expected interval. The inner likely period uses the 50th-percentile miss; the outer calibrated range uses the 80th percentile. The inner band is useful for attention, while the wider range is the reliability claim.

Sparse cars borrow variation from the whole dealer through shrinkage. As a car accumulates its own completed cycles, its own evidence carries more weight. Relative guardrails stop a fast and a slow rotation from receiving the same arbitrary number of days.

05 · Confidence

Reading a score and a label

The score measures how mature the wait is against the car’s expected cycle. It is not a probability that the car will appear on a particular day. “Approaching”, “In window” and “Very likely” begin at scores of 45, 60 and 75 respectively.

Confidence also depends on history. A car with a deep record can earn a narrow range; a thin record remains broad even when its midpoint looks close.

06 · Evidence

Calibration is tested before it is shipped

Each test rebuilds the model using only dates that existed before a known return, then compares the resulting window with what actually happened. The public accuracy page reports forecasts that were published before their outcome was known.

07 · Updates

What changes when the dealer refreshes

A sync updates the current dealer first, then merges verified historical appearances, recomputes every car’s cadence and refreshes windows. A car entering the dealer immediately stops being shown as an upcoming return.

Calibration settings are not silently tuned per car. A change to GTRBit is measured across the lot and is only retained when the recorded validation supports it.

08 · Limits

What the model cannot know

GTRBit cannot see an unpublished game rule, an event-driven selection, a future patch or a source outage. A schedule change can place a return outside both windows. That is a limitation of the information available, not a hidden certainty the model chose not to display.

Estimated dealer exit is a separate calculation from return timing. When a source gives an exact availability value, that value is displayed instead of an estimate.

09 · Glossary

Terms used across the tracker

Appearance

A verified start of a dealer listing.

Interval

Calendar days from one appearance start to the next.

Expected interval

The model’s cadence centre; never a guarantee.

Likely period

The tighter, actionable band around the expected interval.

Calibrated range

The wider empirical range designed to contain most returns.

Fallback

The dealer-wide estimate used when a car lacks usable history.