Data Engineering · Machine Learning · 2024
Nomad: Supply Chain Spatio-Temporal Demand Forecasting
Hierarchical Bayesian forecasting model coupled with distributed PySpark aggregations, predicting inventory requirements across 450 regional hubs.
- Role
- Quantitative Data Engineer
- Timeline
- 10 weeks
- Year
- 2024
- Disciplines
- Data Engineering, Machine Learning
Hierarchical Time Series Forecasting
Forecasting SKU-level demand requires reconciling aggregate national forecasts with volatile local hub fluctuations. Nomad implements a hierarchical reconciliation framework (MinT) that guarantees mathematical consistency across all regional levels.
Models produce probabilistic quantile forecasts (p10, p50, p90), enabling inventory planners to make risk-weighted procurement decisions.
Scalable Spark Execution
Running daily forecasts across 15 million time-series combinations required fine-grained partition management in PySpark. We optimized Delta Lake z-ordering on SKU and location keys, reducing read I/O by 78%.
The batch pipeline completes in 35 minutes, providing refreshed forecasts before morning logistics shifts commence.