Automotive · Predictive analytics
Anchored neural forecasting across three vehicle markets
200K+ repriced, every cycle.
Across three vehicle markets, weekly and monthly.
- Client type
- Vehicle Valuation Provider
- Stack
- PyTorch + Databricks + MLflow
Forecast the next published wholesale value for every vehicle in three heterogeneous markets, and prove each forecast against the future before it ships.
A vehicle-valuation data provider publishes the benchmark values that dealers, lenders, and insurers depend on. Extending that judgement across hundreds of thousands of vehicles, from high-volume pickups to RVs and niche powersports machines, demanded models that could read thin evidence without inventing a price.
The Challenge
Three markets that refuse to behave alike
Sparse, Uneven Evidence
Light duty auction lanes stream rich but noisy signals; powersports evidence arrives as manually uploaded files; and for RVs, no usable auction data exists at all, leaving published history as the only anchor.
Divergent Market Rhythms
Weekly repricing for cars, monthly for RVs and powersports. Seasonal powersports demand, long stretches of unchanged RV values, and constant light duty auction churn all had to live under one discipline.
Accuracy That Survives the Future
A model can score well while quietly republishing last period's value. When most values barely move, that persistence baseline is punishing, so every forecast was scored on honest, future-facing splits.
The Approach
Anchored by design
Turon AI built all three engines on one conviction: in benchmark valuation, stability is a feature. Rather than regress prices from scratch, each model anchors on the vehicle's last published value and predicts the change.
The shared blueprint is embedding-rich neural networks, market-evidence blending, and hard QA gates, specialised per market for cadence, clustering, and data depth.
Validation
Multi-regime
Baseline
Persistence
Backtest
Rolling
Predict the move, not the price
Each engine outputs a bounded adjustment to the vehicle's anchor value. Level correctness is inherited by construction; model capacity is spent on direction and magnitude.
Validate the way production works
Holdouts and time-based splits matched to each market, plus rolling one-step-ahead backtests over a full held-out year, always benchmarked against persistence.
One blueprint, three specialisations
Cars get weekly auction-blended forecasting; RVs, history-anchored dynamics; and sparse powersports segments, hierarchical clustering, each tuned to its market's physics.
System Architecture
Two layers, one valuation pipeline
Layer 01
Forecast Core
Embedding-rich neural forecasters, specialised per market and trained on large multi-year panels.
- Categorical embeddings over make, model, and class, extended with a learned cluster hierarchy for thin powersports segments.
- Dedicated input streams for price history, trend, seasonality, and auction evidence where a market provides it.
- Anchored targets that predict each vehicle's next move rather than its price from scratch, kept robust to outliers.
Direct impact: stable, bounded forecasts that track each vehicle's own trajectory.
Layer 02
Market Reconciliation
Raw forecasts are reconciled with observable market structure before a single value is published.
- Light duty: auction-confidence blending weighted by volume, spread, and agreement, so the model keeps authority.
- Powersports: cluster-level pooling that stabilises thinly traded segments.
- RVs and all markets: hierarchy-integrity checks and publish-time hard gates.
Direct impact: values that respect market structure and survive expert scrutiny.
One Blueprint, Three Markets
Because every engine descends from one anchored core, there is a single discipline to maintain, and a new market can be added by specialising the blueprint rather than starting over.
Validated Against the Future
Every candidate is scored on future-facing splits and rolling backtests before it earns deployment, never on a flattering in-sample number.
Trust by Construction
Leakage-proof features, hierarchy checks, and hard publish gates mean nothing reaches client tables that has not passed its tests.
Deployment & Iteration
Scheduled, gated, and observable
The engines run as scheduled Databricks jobs, weekly for light duty and monthly for RVs and powersports, with every run tracked in MLflow and anomalies flagged before the client sees them.
Each generation was rebuilt from the last one's failure modes, trading brittle complexity for lean, maintainable pipelines.
Retraining on merit, not on a timer
A hash of the training code is compared against the registered model's fingerprint; retraining triggers only on real change, and staleness curves proved the anchored design tolerates long intervals gracefully.
Selected on backtests, not splits
Deployment candidates are chosen on rolling multi-month backtests rather than their best validation split, the honest measure of how an engine will behave next cycle.
Publish-time hard gates
Runs check ordering across trims and configurations, flag conflicting values, and abort on duplicate keys or non-positive prices before anything reaches the client's tables.
Outcome
Accurate, anchored, and audit-ready
Light Duty: Weekly, Evidence-Blended Values
The full car and truck panel reprices every week from a rich engineered feature set, with auction evidence blended in proportion to its confidence and hierarchy checks on every publish.
Recreational Vehicles: Monthly Values From History Alone
RVs are priced monthly with no auction data at all. On its rolling backtest, the deployed engine placed 94.3% of values within ±5% of the published benchmark.
Powersports: A Month of Lead Time
Machines are valued a month ahead of publication, with cluster-pooled evidence stabilising sparse segments and every cycle re-scored against a naive last-value baseline.
“The result is a valuation platform that predicts the move rather than the price, is benchmarked against held-out history in every market, and publishes nothing that has not passed its gates.”
Technical Strength
Deep expertise across the forecasting stack
Time-Series & Panel Modelling
Anchored targets, lag-safe feature design, and seasonality handling across large multi-year vehicle panels.
Validation & Backtest Design
Multi-regime split matrices, persistence baselines, rolling one-step-ahead backtests, and band-coverage scoring.
PyTorch Model Development
Embedding-rich, multi-stream architectures trained, tracked, and hardened to production readiness.
Data Forensics & QA Engineering
Cadence profiling, leakage audits, outlier policy, hierarchy-integrity checks, and publish-time hard gates.
One discipline across three markets
The engagement reflects Turon's core operating model: understand the full technical system, connect research to production constraints, and improve the layers that actually decide field performance. For this provider, that meant three forecasting engines that share one discipline: anchored on published history, validated against the future, and gated before every publish.
Work with us
Have a problem worth publishing?
If it's hard enough to end up in this record, it's the kind of work we want. Tell us what to build, and we'll show ours.