Automating and Optimizing Financial Signal Discovery with Multi-Agent Systems

In quantitative finance, researchers build algorithms to trade assets, derivatives, and other financial instruments. A key part of that work is finding signals: patterns in messy market data that may help predict future returns. These signals can come from price and volume data, economic indicators, fundamentals, or alternative sources like news sentiment.

For years, quant firms have uncovered and tested signals largely by hand. Traditionally, quantitative researchers have to manually hypothesize, code, backtest, and refine hundreds of potential signals. The workflow is often fragmented, moving between data scientists, developers, and analysts, which creates significant lag in a market that moves in milliseconds. Now, AI can help automate parts of this workflow and speed up the research cycle. The quantitative signal discovery agent developer example from NVIDIA shows how to automate signal discovery using agentic architecture built with the NVIDIA Nemotron family of open models and the NVIDIA NeMo Agent Toolkit open-source library.

In this blog post, we walk you through a complete example of building an agentic system for signal discovery using the NeMo Agent Toolkit.

Agentic AI is bridging the gap between human expertise and automated efficiency. We designed a system of signal-discovery agents using NeMo Agent Toolkit to change signal discovery from a manual grind into a self-evolving, autonomous loop. The result is a painless process for designing, testing, and optimizing the agentic system.

Signal agent: Identifies potential alpha signals from market data.

Code agent: Translates signal descriptions into executable Python code.

Evaluation agent: Runs backtests, applies logical evaluation, and iteratively refines signal suggestions.

The NeMo Agent Toolkit provides the orchestration layer to build intelligent agents that plan complex research tasks, use tools to execute Python code, and reflect on backtesting results to refine their own hypotheses. Our proposed system consists of three specialized agents that interact in a continuous cycle of creation, execution, and refinement. The NeMo Agent Toolkit manages the "handoffs" between these agents, ensuring that context (such as signal definitions or backtest results) is preserved across the workflow.

Figure 1. Architecture of the quantitative signal discovery agent developer example

This agent acts as the creative brain, using nemotron-3-nano-30b-a3b through NVIDIA NIM to hypothesize new signal expressions. We designed the following system prompt: "You are a senior quantitative researcher at a top hedge fund. Generate {config.num_signals} unique stock selection signals based on the request" to define the core objective of the agent.

To ensure it produced logically sound results rather than "hallucinating" math, we provided it with a structured library of mathematical calculators as building blocks. Each calculator is a dictionary containing a name, signature, and meaning. In our example, there are 66 different operators covering operations from arithmetic, math (abs, clip, power), rank, and time series (log return, momentum, delta). For example, Rank_Add, normalizes two different sets of data (like price and volume) into a shared 0-to-1 percentile scale and sums them together. We provide the interpretation and the Python code, enabling the agent to combine different market signals into a single, balanced score.

By supplying these reference operators, the agent can reason through its "toolbox" to select the best combination of operators to measure market trends, ensuring the resulting signal is both high-quality and theoretically sound.

By asking the agent to output all the details, we ensure the interpretability and reproducibility of the signals generated by the agent; also, tracking this information will also facilitate the evaluation step further down the loop.

Once a hypothesis is formed, the code agent transforms natural language ideas into executable Python code. In the NeMo Agent Toolkit ecosystem, this agent is specifically tuned for tool-calling and code generation. It takes the blueprint from the signal agent and produces Python code that calculates the signal against historical price-volume data—with the operator implementations from calculator.json inlined automatically so the resulting module is self-contained and portable.

This agent processes the raw data from the code agent to calculate the Information Coefficient (IC) and Rank IC metrics that quantify how well the signal predicts price movement. In quantitative research, Rank IC is the correlation between the ranks of your signal values and the ranks of subsequent asset returns.

Unlike the standard Pearson IC, Rank IC is more robust against outliers and focuses on the relative ordering of assets—an essential feature for building ranking-based trading strategies.

Usually, a "good" institutional-grade signal maintains a mean Rank IC between 0.02 and 0.05. Anything consistently above 0.05 is considered a very strong signal, often found in high-frequency or niche momentum strategies.

The user can define an acceptance threshold for the evaluation agent to decide when to accept a generated signal. If a signal has suboptimal performance in backtesting, the evaluation agent generates optimization suggestions that are fed back into the Signal Agent's next iteration. This creates a self-improving loop where each subsequent generation of signals is more refined than the last.

Why NeMo Agent Toolkit for automating signal discovery?

The toolkit helps shift the project from a rigid script to a flexible research platform. Instead of hard-coding the interactions between agents, you define the system's logic—including personas, tools, and constraints—entirely within a YAML configuration. This modularity makes it trivial to swap models for different tasks. For example, you can assign a high-reasoning model to handle hypothesis generation while using a faster, more cost-effective model for the code agent without modifying the underlying source code.

In our implementation, the config.yml acts as the single source of truth for the entire signal discovery experiment. It centralizes every critical parameter, from the IC thresholds to the depth of the optimization loop. This enables quantitative researchers to iterate at high velocity—adjusting the forward-returns period or tightening the IC requirements—simply by editing a YAML file.

Arize Phoenix tracing adds observability to NIM-powered signal discovery agents in long-running quantitative workflows. Since signal discovery involves high-latency loops—generating complex price-volume formulas, running vectorized backtests, and performing fitness evaluations—Phoenix enables us to visualize the reasoning traces of the LLM in real time.

Figure 2. NeMo Agent Toolkit Tracing provides a visual trace of the LLM agent's reasoning across the signal discovery workflow

By tracking reasoning traces and token usage across the ChatNVIDIA() calls, we can pinpoint whether a bottleneck lies in the model's logic generation and gain deeper insights into the agent's thinking process at each step. This transparency is essential for debugging "silent failures" where an agent might produce a mathematically valid but economically meaningless signal, enabling optimization for both the prompt strategy and the computational costs of the discovery cycle.

Momentum-based signals are one of the most commonly used in trading. They are based on the empirical observation that asset