Machine learning, LLM enrichment, and human review. Each doing its job.

Price prediction starts with a model race, not a model assumption. Nine algorithms are trained and evaluated on every run: linear regression, Lasso, Ridge, support vector regression, K-nearest neighbours, decision tree, bagging, random forest, and XGBoost. Each goes through two tuning stages: a randomised search across a broad parameter space, then a grid search focused on the best region. The model with the lowest RMSE on the held-out test set is selected.

AutoOracle trains two separate models: one for each major NZ car-market source. The platforms differ structurally. One captures seller type (dealer or private) and region from the listing tile; the other does not. Keeping the models separate means each one uses the signals actually available for that source, and a data quality problem on one platform does not affect the other.

Training uses a time-based split: the oldest 80% of listings by first-seen date form the training set, and the most recent 20% form the test set. A random split would allow the model to see future prices during training, since a listing from April 2025 in the training set can be compared to an otherwise identical one from October 2024 in the test set. A time-based split is the only honest simulation of a task that is always about predicting future listings from past patterns. At Iteration 2, XGBoost won both sources.

The price prediction on its own is not enough. A car listed below the estimate is only a genuine deal if the gap is meaningful, not statistical noise. The value score makes that judgment explicit on a 1-to-10 scale.

So how does the model decide what a car should cost? Not with a simple average. A median lumps a tired older car in with a nearly new one. Instead, a machine learning model trained on around 600,000 New Zealand listings predicts the going rate for that exact car, weighing its year, mileage, make, model, engine, fuel, transmission and region together. That predicted figure is the benchmark every asking price is measured against.

Take a 2019 Toyota Corolla, 60,000 km, automatic petrol, in Auckland. The model predicts a going rate of $21,000. Three sellers, three very different scores.

Asking price	Versus estimate	Score	Verdict
$18,000	14% below	9.0	Good value
$21,000	At estimate	5.5	Fairly priced
$25,000	19% above	1.6	Overpriced

Under the hood, the gap is turned into the score with a sigmoid function. The listed price is compared to the estimate as a percentage deviation; that percentage passes through a logistic curve calibrated with steepness constant k=7. A 0% deviation (listed at exactly the estimate) maps to 5.5, the neutral midpoint. Underpriced listings curve toward 10; overpriced ones toward 1. The sigmoid handles extreme outliers naturally: a car listed at 50% below the estimate approaches 10 asymptotically rather than being clamped.

Deviation	Score	Band
+20% underpriced	9.5	Green
+10% underpriced	8.3	Green
+5% underpriced	7.0	Green
At market	5.5	Amber
-5% overpriced	4.0	Amber
-10% overpriced	2.7	Red
-20% overpriced	1.5	Red

After training, k=7 was validated against 596,517 active records. The two market-source medians came out at 5.90 and 5.80. Green, amber, and red bands came out roughly balanced at one-third each. The slight upward skew is consistent with a competitive market where records tend to price fractionally below model estimates. No adjustment was required.

Each vehicle family in AutoOracle has a rich intelligence profile built from two sources: internal data gathered from 18 months of listings and the NZTA motor vehicle register, and LLM-generated content reviewed section by section before it is written to the database.

The pipeline for each family runs in a fixed sequence: resolve all taxonomy aliases so that data from MINI, BMW Mini, and Mini Countryman all aggregate together, gather internal stats (listing counts, median price, price trend across three time windows, MVR fleet size, regional distribution, top variants by volume), run a pre-enrichment variant quality audit, build a structured prompt from all of that data, call the language model, then step through the six output sections one at a time: accept as generated, edit in the terminal, or skip. Nothing is written to the database until the review is complete.

The edit log is part of the data model. Every field records whether the generated content was accepted, edited, or skipped. That log is a training signal for future automation of the review process and a running record of prompt quality over time.

The same pipeline then iterates variants. Each active variant for the family gets its own listing statistics, is compared against the family context already written, and receives LLM-generated notes reviewed before writing. At the time each family was first enriched, that meant 18 active variants for the Mini Countryman, 38 for the Delica, and 51 for the Cayman. The taxonomy keeps evolving, so those counts shift as variants are added and consolidated; the families enriched earliest stay the most deeply documented in the catalogue, and each enrichment pass deepens the record further.