How CarCast Works: The Technology Behind Our Predictions

Used car price forecasting is a hard problem. Here's exactly how we solve it.

The Data

Every week, CarCast ingests pricing data from over 200 vehicle segments covering 84,000+ dealer listings nationwide (via MarketCheck). We track weekly median prices, listing volumes, and days-on-market trends for each Year/Make/Model combination.

We also incorporate macroeconomic signals: BLS CPI data for used vehicles, gasoline prices, and overall inflation -- factors that historically correlate with used car demand shifts.

The Model: Google proprietary AI

At the core of CarCast is a proprietary AI forecasting stack built around a 200-million-parameter foundation model for time-series forecasting.

Unlike traditional statistical models (ARIMA, Prophet) which require per-segment fitting, AI forecasting is a zero-shot forecaster -- it can generate accurate forecasts on vehicle segments it has never explicitly trained on, by learning universal temporal patterns from billions of time-series data points across domains.

How We Generate Forecasts

1. Data preparation: We take the full weekly price history for a segment (typically 100-200+ weeks) and apply a log transform to stabilize variance.
2. Inference: The transformed series is fed to AI forecasting running on dedicated GPU infrastructure via cloud GPU infrastructure. The model generates point forecasts and quantile predictions.
3. Post-processing: We extract P10, P50 (median), and P90 quantiles to create confidence bands, then reverse the log transform to get dollar values.
4. Trend classification: If the predicted 30-day price change exceeds +1.5%, we classify the segment as Rising. Below -1.5%, Softening. Everything in between is Stable. These classifications are informational analytics, not financial advice or a recommendation to purchase any vehicle.
5. Confidence scoring: We combine the prediction band width, historical volatility, and context length to produce a 0-100% confidence score.

Accuracy Tracking

We run walk-forward backtesting on every segment: training on historical data, predicting forward, and comparing against actuals. Our current metrics across all tracked segments:

• Mean Absolute Error: ~$400-800 per segment (varies by price level)
• Directional Accuracy: ~72% (the model correctly predicts whether prices go up or down about 72% of the time)

Limitations

No model is perfect. CarCast forecasts are probabilistic estimates, not guarantees. They work best for:

• Segments with 50+ weeks of history
• Normal market conditions (not black swan events)
• 30-day horizons (accuracy decreases with longer horizons)

See the forecasts in action. Try CarCast free