Client Background

A medium-sized supply chain and logistics company managing warehouse operations and delivery across multiple Australian states had built a basic machine learning model to forecast product demand and optimize inventory distribution. Developed in-house, the model offered some directional value but lacked the precision and consistency required for confident supply chain planning.

Challenge

The client’s existing machine learning model struggled to deliver actionable insights:

  • Forecast accuracy was limited to 71%, often triggering overstocking or stockouts across regional warehouses.

  • The initial feature set was narrow, relying only on historical sales and seasonality patterns.

  • There was no model benchmarking to determine whether more effective algorithms or tuning strategies were available.

  • Stakeholders had low confidence in using the model to make operational decisions due to unpredictable outputs and lack of interpretability.

The client engaged Data Gravity to strengthen the entire ML pipeline—from data foundation to decision support.

Solution by Data Gravity

Service Provided: Machine Learning Audit, Feature Engineering & Model Optimization
Technologies Used: Python, pandas, scikit-learn, Random Forest, XGBoost, LightGBM, SHAP

1. ML Pipeline Review & Data Profiling

Data Gravity initiated the engagement by reviewing the existing forecasting pipeline:

  • Identified gaps in the feature selection logic, training/testing splits, and performance evaluation metrics.

  • Detected underutilized data sources such as delivery delays, supplier lead times, weather impacts, and promotions.

  • Introduced version control, pipeline modularity, and proper logging to enhance reproducibility and future updates.

2. Strategic Feature Engineering

To improve the predictive power of the model, Data Gravity:

  • Extracted new, high-impact features including:

    • Regional delivery delays and historic lag patterns

    • Supplier reliability scores and fulfillment variability

    • Promotion flags, holiday effects, and rolling demand averages

    • Inventory turnover ratios and warehouse congestion indicators

  • Used Random Forest feature importance to isolate the most informative variables and eliminate redundancies.

  • Validated engineered features through correlation analysis and impact on model performance.

3. Model Benchmarking & Hyperparameter Optimization

  • Benchmarked existing model against Random Forest, XGBoost, Random Forest, Neural Network, MLP, and LightGBM algorithms.

  • Tuned hyperparameters using GridSearchCV and Bayesian optimization, focusing on:

    • Forecasting precision

    • False negative rate (stockout risk)

    • Model generalization across regions

  • Applied cross-validation and custom scoring metrics tailored to operational thresholds and cost implications.

4. Explainability & Business Integration

  • Applied SHAP explainability tools to identify which features influenced specific forecast outcomes.

  • Built visual dashboards to help operations teams explore forecasts, understand confidence levels, and drill into key drivers (e.g., weather impact on demand).

  • Delivered clear documentation and a modular framework to ensure the solution could be scaled or transferred internally.

Results

  • Model accuracy improved from 71% to 89%

  • Stockout incidents reduced by 38%, and overstocking by 31%

  • Forecasting precision increased, enabling more agile replenishment strategies

  • Business users gained full transparency into predictions, fostering greater trust and usage in planning cycles

  • The new pipeline enabled the client to standardize forecasting across warehouses, improving overall supply chain responsiveness

Client Testimonial

“Working with Data Gravity completely changed how we approach forecasting. Their feature engineering and model tuning didn’t just improve performance—it helped us see the ‘why’ behind each prediction, giving our team confidence to act on it.” Head of Logistics & Planning

Key Takeaways

  • Data quality and feature richness are the cornerstones of accurate forecasting in supply chains.

  • Model benchmarking and tuning unlock performance gains that basic pipelines often miss.

  • Explainability tools like SHAP are vital in operational contexts where trust and clarity matter.

  • Partnering with a team like Data Gravity turns analytics from a back-office tool into a strategic enabler.

Skills