STAC Report: Extending STAC-ML with Gradient Boosted Tree Models

POC results of STAC-ML with Gradient Boosted Tree Models

7 April 2025

STAC has completed a Proof of Concept (POC) benchmark, extending the STAC-ML™ Markets (Inference) to include Gradient-Boosted Tree (GBT) models (“El Popo Suite”). This POC evaluates GBT inference performance for real-time market data analysis.

Benchmarks were conducted on AWS bare-metal instances, comparing Intel Xeon (X86) and AWS Graviton4 (ARM) processors using the ONNX runtime. Three GBT models with varying complexities were tested, focusing on 99th percentile and maximum latency measurements.

Key findings include latency differences between X86 and ARM platforms and performance variations across GBT model complexities. The reports summarize the benchmark results and model specifications, and are available here: Intel (STAC250314a) and ARM (STAC250314b). STAC premium subscribers can access full configuration data, plus detailed tables and visualizations of audit results at the same links. To learn about subscription options, please contact us.

Feedback is requested to inform the formal specification of the GBT benchmark suite within STAC-ML™. This POC provides data relevant for optimizing machine learning infrastructure for time-sensitive financial applications.