NVIDIA demonstrates a multi-agent system for automating quantitative finance signal discovery using the NeMo Agent Toolkit and Nemotron models. The system coordinates three specialized agents: a signal generator that hypothesizes alpha signals from market data using a library of 66 mathematical operators, a code agent that translates signal blueprints into executable Python, and an evaluation agent that backtests signals using Information Coefficient metrics and feeds optimization suggestions back into the loop. The YAML-driven config allows researchers to swap models per agent, tune IC thresholds, and plug in custom data or operators. An end-to-end example mining momentum signals against S&P 500 data illustrates the iterative refinement process, with Arize Phoenix providing observability into LLM reasoning traces.
Table of contents
Agentic systems for signal discoveryNeMo Agent Toolkit multi-agent signal discovery loopWhy NeMo Agent Toolkit for automating signal discovery?End-to-end example: Mining momentum-based signalsQuantitative signal discovery agent developer exampleGet startedSort: