Large-scale agentic quant research with Weights & Biases

In this video, we walk through how to use Weights & Biases to improve reliability, reproducibility, and explainability in large-scale, agent-driven quantitative research. *We cover two key use cases:* *1. Alpha research with backtesting* Learn how to build, debug, and iterate on a multi-agent system for market analysis. A market event is routed through a classifier to specialized agents (macro, historical, sentiment, and quant), each making tool calls to gather live data. A synthesis agent then generates a probabilistic forecast. You’ll learn how to use W&B Weave to inspect agent traces, including tool calls, latency, cost, and outputs and drill into low-confidence predictions to identify root causes and improve agent behavior. *2. Strategy testing* See how to move beyond manual tuning with an automated optimization loop. You’ll see how to run multiple trials with different agent weight configurations and use a meta-optimizer to iteratively minimize metrics like Brier score. You’ll also learn how to use track convergence and performance using W&B Models. With the visualization tool in W&B Models, you can easily explore how weight configurations impact outcomes using parallel coordinate plots. W&B Weave and Models are fully integrated, giving you a complete audit trail for research, validation, and compliance. Weights & Biases enables end-to-end visibility across your quant research workflows, allowing you to move from hypothesis to back-testing and compliance-grade reproducibility while optimizing strategies at scale. *Chapters:* 00:00 Introduction 00:35 Use cases overview 00:50 Event-driven pipeline architecture 01:38 Weights & biases UI demo 02:18 Debugging model output 03:44 Fixing model & improved forecast 04:26 Evaluation at scale 05:07 Comparing model performance 06:23 Automated optimization loop 08:00 Insights from optimization