💰 quantitative finance

QRAFTI: An Agentic Framework for Empirical Research in Quantitative Finance

QRAFTI is a multi-agent framework designed to emulate a quantitative research team by integrating panel data toolkits with MCP servers to facilitate equity factor research, enabling the replication of established factors, testing of new signals, and generation of standardized reports with narrative analysis and computational traces.

Original authors: Terence Lim, Kumar Muthuraman, Michael Sury

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Terence Lim, Kumar Muthuraman, Michael Sury

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a machine that can predict which stocks will go up or down. In the world of finance, this is called Quantitative Research. Traditionally, this job is done by teams of highly paid experts (quants) who spend years doing three main things:

Having an idea (e.g., "Maybe companies with low debt do better?").
Doing the math (cleaning messy data, running complex formulas, and checking if the idea works).
Writing the report (explaining why it worked and proving it wasn't just luck).

This process is slow, prone to human error, and often hard to double-check.

Enter QRAFTI. Think of QRAFTI not as a single robot, but as a virtual research team built entirely out of advanced AI agents. The paper introduces this system as a way to automate and improve this entire research process.

Here is how QRAFTI works, explained through simple analogies:

1. The Team: Specialized AI Agents

Instead of one giant AI trying to do everything (which often leads to confusion), QRAFTI splits the work among three specialized "employees," each with a specific job:

The Research Analyst (The Planner): This agent listens to your idea (e.g., "Find me a factor based on Warren Buffett's style"). It breaks the big idea down into a step-by-step recipe. It doesn't do the math itself; it just figures out what needs to be done.
The Quant Developer (The Builder): This agent is the coder. If the Analyst's recipe requires a tool that doesn't exist yet, this agent writes the specific computer code to build it. It's like a carpenter who builds a custom shelf because the store didn't sell the right size.
The Risk Manager (The Auditor): This agent is the strict boss. It checks the work, ensures the math is standard, and writes the final report. It asks, "Did we check for bias? Is this result real, or just a fluke?"

2. The Toolbox: The "MCP"

Imagine the Analyst and Developer need to use a giant library of financial data (like a massive warehouse of stock prices). Instead of the AI guessing how to open the warehouse door, QRAFTI uses a Model Context Protocol (MCP).

Think of MCP as a universal remote control. Instead of the AI trying to learn how to operate every single machine in the warehouse, the remote gives it a simple button for "Get Stock Prices," "Clean Data," or "Calculate Risk." This prevents the AI from getting lost or making up fake data.

3. The "Self-Reflection" Trick

One of the biggest problems with AI is that it can confidently give you the wrong answer. QRAFTI solves this with a technique called Reflection-Based Planning.

Imagine you are writing an essay.

Normal AI: You ask it to write the essay, and it types it out immediately.
QRAFTI: You ask it to write the essay.
- Step 1: It drafts an outline.
- Step 2: It stops and critiques itself: "Wait, is this outline logical? Did I miss a step? Is this data available?"
- Step 3: It fixes the outline before writing the final essay.

This "pause and think" step drastically reduces errors, especially for complex, multi-step tasks.

4. The "Paper Trail" (Traceability)

In science, if you can't show your work, nobody believes your results. QRAFTI creates a Computation Graph.

Think of this like a GPS history log for a road trip. If you arrive at a destination, the GPS doesn't just say "You are here." It shows the exact route taken, every turn, and every stop.

If QRAFTI says, "This stock factor works," you can click a button and see the exact path it took to get there.
You can see exactly which data was used, how it was cleaned, and which formula was applied. This makes the research reproducible (anyone can do it again and get the same result) and auditable (you can check for mistakes).

5. The Results: Better than Just "Coding"

The paper tested QRAFTI against two other methods:

Just using tools: Letting the AI pick from the toolbox.
Just writing code: Letting the AI write Python code from scratch without a toolbox.

The Finding:

For simple tasks, writing code works fine.
For complex, multi-step tasks (like building a full investment strategy), QRAFTI's combination of tools + self-reflection was far superior.
When the AI was forced to just write code without the "toolbox" or the "reflection" step, it made many more mistakes and produced results that didn't match the real-world benchmarks.

Why Does This Matter?

Currently, the world of finance is full of "Factor Zoos"—thousands of investment strategies that look good on paper but fail in real life because they were built on messy data or statistical tricks (p-hacking).

QRAFTI aims to be a quality control filter. By standardizing how research is done, forcing the AI to check its own work, and keeping a perfect record of every step, it helps researchers:

Find real patterns in the market.
Avoid wasting time on dead ends.
Trust the results because the "math homework" is visible and correct.

In short: QRAFTI turns the chaotic, messy process of financial research into a structured, transparent, and highly reliable assembly line, managed by a team of AI specialists who double-check each other's work.

1. Problem Statement

Quantitative financial research is a complex, multi-stage workflow involving idea generation, data acquisition/cleaning, signal construction, backtesting, risk adjustment, and robustness evaluation. This process is highly sensitive to implementation details (e.g., lag conventions, universe screens, breakpoint definitions), leading to a "factor zoo" where many published predictors are fragile, non-reproducible, or statistical artifacts of data mining.

While Large Language Models (LLMs) have advanced, they often struggle with long, interdependent workflows common in empirical finance. They tend to lose track of prior results, fail to maintain consistency across steps, and lack the rigorous structure required for reproducible scientific research. Existing single-turn or static code-generation approaches often fail to handle the iterative planning and error correction necessary for robust factor discovery.

2. Methodology

The authors introduce QRAFTI (Quantitative Research Assistants with Financial Tools and Intelligence), a multi-agent framework designed to emulate a quantitative research team. The system integrates an empirical research toolkit with Model Context Protocol (MCP) servers to expose data access and factor construction as callable tools.

Core Architecture

QRAFTI employs a specialized multi-agent architecture with three distinct roles:

Factor Research Agent ("Research Analyst"): Interprets natural language requests, decomposes tasks, selects tools, and maintains the factor construction pathway.
Standardized Reporting Agent ("Risk Manager"): Generates standardized diagnostics and narrative research reports.
Code Writing Agent ("Quant Developer"): Dynamically writes and executes custom Python code when built-in tools are insufficient, ensuring auditability by recording code and outputs.

Key Technical Components

Panel Data Abstraction: The system encapsulates financial data as a $T \times N$ panel (Time $\times$ Asset) within a Panel class. This standardizes operations like winsorization, quantile binning, and rolling transformations, reducing implementation variance.
Tool-Augmented Execution: Instead of relying solely on generative text, agents invoke specific tools (primitive math, cross-sectional transforms, time-series transforms, portfolio construction) via MCP. This grounds the reasoning in executable code.
Reflection-Based Planning: To mitigate planning errors in multi-step tasks, the Research Agent follows a three-phase loop:
1. Drafting: Propose a sequential step plan.
2. Critique: Audit the plan for technical feasibility and logical gaps.
3. Refinement: Revise the plan before execution.
Context Compaction: For long workflows, a specialized LLM call distills chat history into a structured summary of computed identifiers, inputs, and dependencies to prevent context loss.
Traceability: The system generates computation graphs (Directed Acyclic Graphs) visualizing the flow from raw inputs to final outputs, alongside structured logs for auditability.

3. Key Contributions

Framework for Reproducibility: QRAFTI shifts the paradigm from "chat-style" LLMs to tool-grounded, multi-agent systems specifically designed for empirical asset pricing, emphasizing reproducibility and explainability.
Standardized Diagnostic Protocols: The framework implements a rigorous reporting standard (adapted from Novy-Marx and Velikov, 2023) that evaluates factors across coverage, spread statistics, risk-adjusted alphas, and size quantiles, reducing researcher degrees of freedom.
Hybrid Execution Strategy: It effectively combines a constrained set of reliable, pre-validated tools with dynamic code generation, balancing safety/standardization with flexibility.
Evaluation Metric (Sim@k): The authors propose Sim@k (Cosine Similarity at k), a metric that measures the expected quality of the best result among $k$ attempts by comparing generated factor returns to benchmark references via cosine similarity, rather than requiring exact numerical matches.

4. Results

The authors evaluated QRAFTI on two replication tasks (Fama-French HML value factor and Jensen et al. JKP price momentum factor) and a dynamic code generation task.

Replication Accuracy: QRAFTI successfully replicated established factors, with scatter plots showing strong agreement (observations near the 45-degree line) between constructed panels and benchmark returns from Ken French and JKP factors.
Ablation Studies (Table 1):
- Tool Access: Access to the full toolset significantly outperformed "coding-tool only" configurations, especially in 2-step and multi-step tasks (e.g., Sim@1 dropped from 1.000 to 0.591 for 2-step tasks without tools).
- Reflection Planning: For multi-step tasks, adding reflection-based planning improved performance from a Sim@1 of 0.7435 (tools only) to 0.9858 (tools + reflection).
- Consistency: The system achieved near-perfect performance (Sim@5 $\approx$ 1.0) on simpler tasks and high consistency on complex workflows when using the full architecture.
New Factor Discovery: The system successfully generated a "Buffett-style" factor (combining quality and value signals) and produced a standardized report highlighting its performance characteristics, including a negative Sharpe ratio but positive alpha after risk adjustment.

5. Significance

QRAFTI represents a significant step toward autonomous, reproducible scientific discovery in finance.

Workflow Transformation: It suggests a future where human researchers shift from manual coding and data manipulation to designing and managing agentic research assistants.
Mitigating "P-Hacking": By enforcing standardized diagnostics, transparent computation graphs, and strict adherence to replication protocols, the framework helps distinguish genuine economic signals from statistical noise.
Scalability: The modular design allows for the integration of new data sources and tools without retraining the core models, making it a scalable solution for the evolving "factor zoo."

In conclusion, QRAFTI demonstrates that combining specialized agents, tool integration, and reflection-based planning creates a robust system capable of executing complex empirical finance workflows with a level of accuracy and explainability that surpasses dynamic code generation alone.