Business Context

RouteOps manages 18,000 delivery vehicles across North America. The fleet operations team wants a model to predict next-month maintenance cost per vehicle and justify whether a Random Forest is a better choice than Linear Regression for this fleet problem.

Dataset

The training data contains monthly vehicle-level records aggregated from telematics, maintenance systems, and driver logs.

• Size: 240K vehicle-month rows, 34 features
• Target: Continuous — maintenance_cost_next_30d in USD
• Distribution: Right-skewed with many low-cost months and a long tail of expensive repairs
• Missing data: ~12% missing in driver behavior fields, ~6% in weather/environment features

Success Criteria

A good solution should outperform a simple linear baseline by at least 15% on MAE and explain when nonlinear models are justified. A production-ready answer should also identify the most important cost drivers for fleet managers.

Constraints

• Nightly batch scoring for 18,000 vehicles must finish in under 10 minutes
• Model outputs should support feature importance review by operations analysts
• Retraining should be feasible monthly on standard cloud CPU instances

Deliverables

1. Build a baseline Linear Regression model and a Random Forest regressor.
2. Explain why Random Forest may outperform Linear Regression on this fleet dataset.
3. Design preprocessing for mixed feature types and missing values.
4. Evaluate both models using appropriate regression metrics and compare results.
5. Recommend a production model, retraining cadence, and monitoring plan.