Predicting Vehicle Demand Before Acquisition

The Challenge

Automotive retailers make inventory acquisition decisions before knowing how much customer interest a particular vehicle will generate.

Two vehicles with similar prices, age and mileage can perform very differently once they enter inventory.

The objective was to develop a data-driven approach for estimating vehicle demand before acquisition, helping decision-makers distinguish between vehicles likely to generate stronger customer interest and those that may require more time to sell.

A key analytical metric was Leads per Day (LPD), measuring customer interest relative to the amount of time a vehicle remained available.

For example, an analysis might show that vehicles generating 0.40+ leads per day move materially faster than vehicles generating below 0.15 leads per day.

These figures are illustrative and do not represent client data.

Our Approach

We developed a machine-learning framework using historical inventory, vehicle characteristics and customer engagement data to identify patterns associated with vehicle demand.

The modeling incorporated signals across several dimensions.

Vehicle Characteristics

Make, model, year, trim, mileage, body type, drivetrain and other vehicle-specific attributes.

Pricing & Market Position

Vehicle pricing, valuation information, market positioning and external marketplace signals.

Vehicle Configuration

For specialized vehicle categories, additional attributes such as fuel type, roof/top configuration, number of doors and cab type were incorporated into the analysis.

Customer Demand

Historical lead activity, inventory age and Leads per Day provided the behavioral signals required to understand how different types of vehicles performed.

Finding the Right Modeling Strategy

An important part of the engagement involved determining how the prediction problem should be structured.

The initial approach explored predicting the exact Leads per Day value for an individual vehicle.

However, vehicle-level demand can be sparse and highly variable. Predicting an exact value was therefore not always the most useful representation for operational decision-making.

The problem was subsequently reframed around demand classification. Instead of asking:

"How many leads per day will this vehicle generate?" the model can answer:

"Is this vehicle expected to generate Low, Medium or High customer demand?"

Illustrative categories might look like:

• Low Demand: Below 0.15 LPD
• Medium Demand: 0.15–0.40 LPD
• High Demand: Above 0.40 LPD

Thresholds are illustrative and do not represent client data.

Improving Model Reliability

Model evaluation revealed that vehicles clearly exhibiting high- or low-demand patterns could generally be distinguished more reliably than vehicles positioned close to category boundaries.

This led to experimentation with a two-stage modeling strategy.

The first stage focuses on identifying vehicles with strong high- or low-demand signals. More ambiguous vehicles can then be evaluated separately using additional features and decision logic.

The objective was not simply to maximize a model metric, but to create predictions that were useful in real acquisition decisions.

From Prediction to Decision Intelligence

The resulting demand classifications were designed to support practical inventory decisions.

Vehicles with stronger predicted demand could be prioritized for acquisition, while uncertain vehicles could be flagged for additional review and lower-demand vehicles could be deprioritized.

Once a vehicle entered inventory, its actual lead velocity could also be compared with its predicted demand category.

This creates a continuous learning loop:

The Outcome

The project transformed historical lead and inventory data into a repeatable vehicle demand intelligence framework.

Instead of relying exclusively on experience, valuation information or broad market averages, acquisition teams can incorporate vehicle-specific demand predictions into their decision-making process.

Capabilities Demonstrated

Predictive Modeling • Classification • Feature Engineering • Demand Analytics • Model Validation • Threshold Optimization • Decision Intelligence • Continuous Model Improvement