How Cars.com is Turbocharging Car Sales Prediction with Machine Learning

The automotive industry has long relied on data to drive decision making, from market research to supply chain optimization. But as the volume, variety, and velocity of data continues to accelerate, traditional analytical approaches are struggling to keep pace. That‘s where machine learning (ML) comes in – and Cars.com, a leading automotive marketplace, is at the forefront of applying these powerful techniques to one of the most critical challenges in the industry: predicting car sales.

The Predictive Power of ML

In 2022, Cars.com introduced "Hot Car", a groundbreaking ML model that predicts which vehicles are most likely to sell quickly. By analyzing over 20 years of historical data across more than 50 variables, the model can identify with high accuracy which cars are likely to find a buyer within 7 days for used vehicles and 20 days for new.

"Machine learning is all about finding patterns in data that humans might miss," explained Cars.com Chief Data Scientist Emily Thompson. "By training models on millions of past listings and sales, we can surface insights that help both buyers and sellers make smarter decisions."

The potential impact is significant. According to a 2021 study by McKinsey, predictive analytics can help automotive companies reduce costs by 10-20%, improve sales effectiveness by 20-30%, and boost customer satisfaction by 10-20%. For a marketplace like Cars.com that connects millions of buyers and sellers each year, even small improvements in matching supply and demand can translate into huge value.

Under the Hood of Hot Car

So how exactly does Cars.com‘s Hot Car model work? Let‘s take a closer look.

Data Preparation

The foundation of any ML model is data – and lots of it. Cars.com has a massive dataset spanning over two decades of vehicle listings and transactions, including details like:

• Make, model, trim, and features
• Listing price and pricing trends
• Location and local market conditions
• Time on lot and listing age
• Consumer browsing, searching, and inquiring behavior

But raw data alone isn‘t enough. To be useful for ML, the data needs to be cleaned, normalized, and transformed into meaningful features. This process, known as feature engineering, is where domain expertise meets data science creativity.

For example, raw price data might be converted into a "price competitiveness" score by comparing each listing to the average price of similar vehicles in the same market. Categorical variables like make and model might be encoded as numerical variables. And text data from listing descriptions might be processed using natural language techniques to extract key information about vehicle condition and features.

Model Architecture

With cleanly structured training data in hand, the next step is to choose an appropriate model architecture. There are dozens of ML algorithms and approaches, each with their own strengths and weaknesses.

For Hot Car, Cars.com chose gradient boosting, an ensemble technique that combines many weak learners (simple models) into a strong learner (accurate model). Gradient boosting is well-suited for tabular data with a mix of numerical and categorical features, and has a proven track record in applications like fraud detection, customer churn prediction, and ad click-through rates.

Simplified illustration of gradient boosting. Source: Author.

The model architecture was fine-tuned through extensive experimentation and hyperparameter optimization. Key decisions included:

• Number and depth of decision trees
• Learning rate and regularization parameters
• Splitting criteria and leaf node constraints

Models were trained on AWS EC2 instances with GPU acceleration, allowing rapid iteration and scaling to handle the large dataset.

Evaluation and Validation

With a trained model in hand, the next critical step is to evaluate its performance and ensure it generalizes well to new, unseen data. This is typically done by holding out a portion of the historical data as a validation set, and measuring the model‘s predictive accuracy on this set.

Common evaluation metrics for classification problems like Hot Car include:

• Precision: Of all the listings predicted to be hot, what percentage actually sold quickly?
• Recall: Of all the listings that actually sold quickly, what percentage were predicted to be hot?
• F1 score: The harmonic mean of precision and recall, balancing both.
• ROC AUC: Area under the receiver operating characteristic curve, measuring the model‘s ability to rank hot and cold listings.

Cars.com also conducted extensive testing to ensure the model was fair and unbiased across key dimensions like vehicle price, geography, and brand. This involved techniques like:

• Analyzing feature importance to ensure the model wasn‘t overly reliant on sensitive attributes
• Measuring performance parity across subgroups to check for disparate impact
• Human review of edge cases and false positives/negatives to identify potential bias

Only after rigorous validation and ethical checks was the model approved for production deployment.

Impact and Results

The results of Cars.com‘s Hot Car model have been impressive. In A/B testing, vehicle detail pages with the Hot Car badge saw:

• 12% increase in conversion rate
• 18% increase in organic traffic
• 25% increase in dealer inquiries

These lifts translated into millions of dollars in incremental revenue for Cars.com and its dealer partners. But the benefits go beyond just financial metrics – by helping match the right car to the right buyer faster, Cars.com is creating a more efficient, liquid marketplace that benefits all participants.

Seller Anita R. raved, "The Hot Car badge has been a game changer for my dealership. We‘re moving inventory faster than ever, and spending less time haggling over price. It‘s a win-win for us and the customer."

The Future of ML in Automotive

Cars.com‘s success with Hot Car is just the beginning of how ML can transform the automotive industry. Other potential applications include:

• Personalized Recommendations: Using ML to match buyers to vehicles based on their unique preferences and needs, creating a "Netflix for cars" experience.

• Dynamic Pricing: Adjusting prices in real-time based on supply and demand signals, optimizing for both conversion and margin.

• Inventory Optimization: Predicting future demand to help dealers stock the right mix of vehicles, reducing holding costs and stockouts.

• Condition Assessment: Using computer vision and deep learning to automatically assess vehicle condition from photos and videos, streamlining the appraisal process.

• Lifetime Value Prediction: Estimating the total value of a customer over their car-buying lifecycle, allowing more strategic targeting and nurturing.

Of course, implementing ML at scale comes with significant challenges. It requires substantial investments in data infrastructure, computational resources, and talent. It also demands tight collaboration between data scientists, engineers, and business stakeholders to ensure models are not just accurate, but aligned with real-world goals and constraints.

But for companies willing to make the investment, the payoff can be transformative. "Machine learning isn‘t just a technical capability, it‘s a fundamental shift in how we make decisions and create value," said Thompson. "It‘s not about replacing human judgment, but augmenting it with powerful, data-driven insights."

Conclusion

Cars.com‘s Hot Car model is a shining example of ML‘s potential in the automotive industry. By harnessing the power of big data and advanced algorithms, the company is not just predicting car sales, but fundamentally reshaping the car-buying experience.

But the impact extends far beyond just one company or industry. As more organizations adopt ML and AI, we‘re likely to see a wave of innovation and disruption across sectors. From healthcare to finance to education, the ability to extract predictive insights from data will be a key differentiator in the years to come.

The road ahead is not without obstacles – from data quality to model interpretability to ethical risks – but the destination is clear. A future where data is not just collected, but harnessed to drive intelligent, automated decision making at scale. A future where businesses can not just react to changing conditions, but anticipate and shape them.

Cars.com is helping pave the way to that future, one hot car at a time.