How zeb Implemented a Machine Learning-Powered Shipment Cost Prediction System for Data-Driven Freight Procurement

Manual Pricing Practices and Systematic Cost Deviations

Our client had shipment pricing decisions that were primarily driven by manual benchmarking and historical intuition. This approach created inconsistencies across procurement teams and limited the ability to scale pricing operations effectively.

The organization experienced measurable financial inefficiencies, with 4.98% overpayment and 8.28% underpayment across weekly shipments. Pricing accuracy also varied across equipment types and lane combinations, with limited standardization in handling scenarios such as team versus solo driver cost differences.

Additionally, the pricing process required auditability, confidence-based workflows, seamless system integration, and reliable performance during peak shipping periods. Without a predictive intelligence layer, it was difficult to ensure consistency, reduce financial exposure, and support real-time decision-making.

Machine Learning-Based Prediction Engine with Confidence-Aware Workflows

zeb designed and implemented a production-grade shipment cost prediction system that delivers real-time pricing intelligence using machine learning models, structured shipment data, and market rate signals.

Data Integration and Quality Framework The solution consolidated over 195,000 shipment records from multiple sources, including TMS data, DAT and GreenScreen market rate feeds, and headhaul/backhaul datasets.

A structured data quality framework was implemented with null handling strategies, fallback mechanisms, and filtering criteria to ensure reliable and consistent training data.

Feature Engineering and Optimization More than 35 candidate features were evaluated and refined to 21 high-impact features with minimal impact on accuracy.

Key feature enhancements included:

• Market rate composites combining multiple external data sources
• Dwell time categorization to capture operational variability
• Time-based features to identify pricing patterns across different periods

Model Selection and Optimization Multiple machine learning models were trained and compared on Amazon SageMaker using consistent datasets. A Ridge Regressor model was selected for production due to its high accuracy (3.37% MAPE), stability through regularization, and fast inference performance.

Confidence-Based Decision Framework A custom confidence scoring mechanism (70–92%) was developed based on lane density and data reliability. This enabled automated pricing for high-confidence scenarios while routing low-confidence predictions for manual broker review with defined threshold controls.

Production Deployment and Monitoring The solution was deployed as a scalable REST API using Amazon ECS and API Gateway, with data managed in Amazon S3 and Redshift.

Monitoring and observability included:

• Latency and error tracking through CloudWatch
• A comprehensive QuickSight dashboard with real-time metrics and multi-dimensional filters
• Performance tracking across teams, lanes, and shipment types

Controlled Release and Continuous Improvement A structured deployment lifecycle was established, covering validation, approvals, and production monitoring.

Three model versions were released with continuous performance improvements, reducing MAPE from 6.71% to 3.37%, supported by weekly performance tracking and retraining workflows.

Improved Pricing Accuracy and Scalable Procurement Intelligence

The engagement delivered measurable improvements across procurement operations and financial governance:

• High Accuracy with Production-Ready Performance: Achieved 96.63% prediction accuracy and 3.37% MAPE, exceeding defined thresholds (>90% accuracy, <10% MAPE), with 69.73% of predictions within 10% and 95.12% within 20% of actual shipment costs.
• Reduced Financial Exposure: Addressed 4.98% overpayment and 8.28% underpayment patterns across ~2,900 weekly shipments through consistent, data-driven pricing, improving margin control and carrier alignment.
• Confidence-Aware Automation: Implemented 70–92% confidence scoring to enable automated pricing for high-confidence cases and broker review for low-confidence scenarios, optimizing operational efficiency.
• Continuous Model Improvement: Delivered three production versions (V1.0–V1.2) with measurable accuracy gains, supported by weekly performance tracking and model versioning for auditability.
• Scalable and Observable Architecture: Deployed a REST API validated with 120 concurrent requests and zero errors, supported by real-time monitoring and a multi-dimensional QuickSight dashboard.
• Foundation for Advanced Use Cases: Established a scalable base for future capabilities, including overpayment detection, dwell time-based pricing optimization, and lane-level performance analysis.

By partnering with zeb, TranscendTrail successfully transitioned from manual pricing practices to a machine learning-driven approach that improves accuracy, consistency, and scalability.

The solution enables procurement teams to make informed, real-time pricing decisions while maintaining control, transparency, and operational efficiency.

Connect with zeb to design scalable, confidence-aware ML solutions tailored to your freight operations.