AI Economist Agent: An Agentic Framework for Evidence-Based Economic and Financial Analysis with RAG, Knowledge Graphs, and Large Language Models

arXiv:2606.20041 · econ.GN, cs.AI, cs.LG, q-fin.EC, q-fin.GN · Submitted 2026-06-18 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "AI Economist Agent: An Agentic Framework for Evidence-Based Economic and Financial Analysis with RAG, Knowledge Graphs, and Large Language Models".

Jane: The paper was written by Masahiro Kato from Mizuho-DL Financial Technology, Company, Limited.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: We are starting today with a heavy hitter titled "AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models."

Jane: That title is a mouthful, Tom, but Masahiro Kato from Mizuho-DL is essentially proposing a way to make AI act like a professional economist.

Tom: Does that mean the AI is actually doing the economic thinking instead of just repeating what it read online?

Jane: Exactly, it's trying to move away from just generating pretty sentences and toward actually using logic and math.

Lu: I love the idea of an "agentic" framework because it suggests the AI has a sense of purpose and can follow a complex plan.

Tom: Lu, you think that autonomy is the real breakthrough here?

Lu: I do, because an agent can navigate through different layers of information to find a specific truth.

Meng: I wonder how much work it takes to set up that kind of environment for the AI to actually function.

Jane: It sounds like a lot of heavy lifting is happening behind the scenes before the AI even starts writing.

Meng: If the system has to manage its own planning and retrieval, the stability of that workflow must be a huge concern.

Lalam: This could change how we trust automated systems in high-stakes sectors like finance.

Tom: So we aren't just looking at a chatbot, we are looking at a digital analyst.

Lalam: It creates a culture where we can rely on AI to provide a clear trail of evidence for every claim it makes.

Jane: That makes sense, and it leads us directly into how this system actually manages all that information.

Summary: Tom: Now that we have the title of "AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models" out of the way, let's look at the mechanics.

Jane: The system uses a Knowledge Graph to keep everything organized, including text evidence and mathematical model specifications.

Tom: So the graph acts like a massive, structured memory for the AI?

Jane: Precisely, it's not just a pile of documents; it's a map of how different economic concepts and data points relate to one another.

Lu: The creativity in this design comes from how the agent uses that map to select the right mathematical tools for a specific problem.

Tom: Lu, you're saying the AI chooses its own equations?

Lu: It selects a model specification from the graph and then asks a separate computational layer to do the actual math.

Meng: That separation is smart because it keeps the LLM away from the actual arithmetic.

Jane: That way, the AI doesn't hallucinate a number; it just reports the result that the math engine produced.

Meng: I'm curious about the "write-back" part of the process where the results go back into the graph.

Tom: Is that how the AI stays consistent with its own previous steps?

Meng: It seems like the only way to ensure the final report is actually grounded in the data the model just calculated.

Lalam: This architecture builds a bridge between human-style reasoning and machine-style calculation.

Jane: It's a way to make sure the narrative and the numbers are always in sync.

Tom: That brings us to the big question: does this actually work better than what we have now?

Paper discussion segment 3: Tom: We've seen how "AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models" is built, so let's talk about the results.

Jane: The researchers compared three different ways of generating reports: an LLM alone, a standard RAG system, and this new model-grounded GraphRAG.

Tom: And they tested it on two very real-world scenarios, right?

Jane: Yes, they looked at U.S. inflation and Federal Reserve policy, as well as bank stress tests regarding commercial real estate.

Lu: The jump in quality is visible when you see how the GraphRAG version connects the scenario to specific models like DSGE-lite.

Tom: So the inflation report wasn't just guessing about interest rates?

Lu: No, it was tracing a path from the inflation shock through a specific model to the resulting yield changes.

Meng: I noticed they used a regime-switching model for the bank stress test application.

Jane: That was a key part of the second experiment to see if the AI could handle non-linear shifts in credit spreads and property prices.

Meng: The fact that the system can report specific stress metrics like the CET1 ratio or liquidity coverage shows it's doing real work.

Lalam: The most impressive part is the auditability, which the researchers measured with high scores.

Tom: You mean we can actually see the "why" behind every statement in the report?

Lalam: Exactly, it transforms the AI from a black box into a transparent collaborator that shows its work.

Jane: It's much more reliable when you can trace a claim back to a specific model run and a specific piece of evidence.

Tom: It's a massive improvement over just getting a fluent but potentially empty narrative.

Conclusion: Tom: We are coming to the end of our discussion on "AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models."

Jane: This paper really shows that grounding AI in formal models and structured graphs is the path forward for specialized fields.

Lu: I think this opens the door for AI to assist in complex policy design by providing a structured playground for different economic theories.

Meng: I'll be watching to see how easily these knowledge graphs can be integrated into existing financial databases.

Lalam: If we can master this, we'll move toward a world where complex economic truths are accessible and understandable to everyone.

Tom: It's a fascinating vision of the future for both finance and AI research.

Jane: It certainly makes you realize how much more there is to do with agentic workflows.

Tom: Thanks to the whole team for joining us today.

Jane: See you next time, everyone!

Tom: Goodbye!

Mizuho-DL Financial Technology, Company, Limited

econ.GN, cs.AI, cs.LG, q-fin.EC, q-fin.GN

Submitted: 2026-06-18

Updated: 2026-09-10

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 84/100

The gist: The paper introduces the "AI Economist Agent," an agentic framework designed for conducting rigorous, evidence-based economic and financial analysis by integrating Retrieval-Augmented Generation

Key concepts

Knowledge Graph: The graph functions as a structured memory for the AI. It organizes information by mapping how different economic concepts and data points relate to one another, rather than treating them as isolated documents. This structure allows the agent to understand complex relationships.
Agentic Framework: This framework suggests that the AI has a defined purpose and autonomy, allowing it to follow a complex plan. It enables the system to navigate through multiple layers of information, moving beyond simple responses toward purposeful analysis.
Model-Grounded Analysis: This process requires the AI to base its conclusions on formal mathematical models and logic, rather than just generating fluent text. The agent selects a model specification from the graph and uses a separate computational layer to perform accurate calculations.

Terminology

RAG — 61× in this paper · explained in 3 episodes
Grounding — 28× in this paper · explained in 3 episodes
Knowledge Graph — 24× in this paper · explained in 7 episodes
Knowledge graphs — 12× in this paper · explained in 2 episodes
Agentic Framework — 8× in this paper · explained in 6 episodes
Trace — 7× in this paper · explained in 2 episodes
AI Agents — 5× in this paper · explained in 2 episodes
Retrieval-Augmented Generation — 5× in this paper
Graph Construction — 3× in this paper · explained in 3 episodes
Core — 2× in this paper · explained in 2 episodes
Large Language Models (LLMs) — 2× in this paper · explained in 6 episodes
Retrieval-Augmented Generation (RAG) — 2× in this paper · explained in 16 episodes
Agentic Architecture — 1× in this paper · explained in 1 episode
Agentic Retrieval — 1× in this paper · explained in 1 episode
Agentic Workflows — 1× in this paper · explained in 1 episode
Consistency — 1× in this paper · explained in 3 episodes
Generative AI — 1× in this paper · explained in 2 episodes
Reinforcement Learning — 1× in this paper · explained in 4 episodes

Summary

The paper introduces the AI Economist Agent, an agentic framework designed for conducting rigorous, evidence-based economic and financial analysis by integrating Retrieval-Augmented Generation (RAG), Knowledge Graphs, and Large Language Models. This framework is crucial because it moves beyond simple text generation by providing a traceable mechanism to link qualitative narrative reasoning with quantitative model outputs and retrieved factual evidence, thereby enhancing the reliability of complex economic forecasting.

The Agentic Execution Pipeline

The system operates via a structured, multi-step agent execution trace designed to ensure comprehensive data integration before report generation. This process begins with initial graph setup, involving the loading of Evidence rows, temporal facts, model catalog, and the scenario node. Key operational steps include:

Planning: An LLM planner determines the necessary sequence of actions.
Retrieval: The system executes both Text retrieval (for RAG) and a Graph retrieval before execution to gather context.
Model Selection and Execution: The process utilizes an LLM model selector to choose the appropriate specification, followed by the Model execution layer.
Graph Write-back: After computation, a dedicated Graph writer records the results into structured objects such as ModelRun, ForecastOutcome, and BankMetric.
Report Generation: Finally, an LLM report generator uses the accumulated graph contents to produce the final analysis.

Integration of Quantitative Modeling (Model Grounding)

The framework explicitly incorporates quantitative stress testing by linking scenario narratives to executable models. The model-grounded approach is necessary because scenarios often involve non-linear transitions, such as moving from normal refinancing conditions to stress conditions. This integration involves:

Selecting a specific model specification, such as the Regime Switching spec- Regime Switchification 005 for Application 2.
Executing the model against the defined scenario shock.
The resulting outputs are not merely textual claims; they are structured metrics, including calculated values like stress CET1 of 5.984 and stress LCR of 143.877, which are recorded as traceable objects within the graph structure.

Comparative Report Generation Conditions

The framework tests three distinct report generation conditions to isolate the value added by each component, ensuring transparency in the information set provided to the LLM:

LLM-only: Receives only the Scenario narrative.
RAG-only (Document-RAG): Receives both the Scenario narrative and Retrieved evidence text, grounding its stress narrative in external documents.
Model-grounded (GraphRAG): Receives the most comprehensive set of inputs, including the Scenario narrative, Retrieved evidence text, pre-run graph paths, a selected model specification, and the structured execution request.

Evaluation Criteria and Auditability

The efficacy of the generated reports is judged against rigorous criteria that mandate traceability and consistency. A high score requires adherence to multiple dimensions:

Evidence grounding: Ensuring the report uses retrieved evidence rather than unsupported claims.
Graph reasoning: Requiring the use of supplied graph paths or relationships.
Model grounding: Demanding that the report is linked to a selected model specification and executed output, noting that simply mentioning a model name is insufficient.
Numerical discipline: Verifying that all numerical statements are consistent with the information supplied.

This comprehensive structure ensures that the final analysis is not only economically coherent but also fully auditable, exposing enough information to trace the source of every claim made.

Improvements for AI systems

Improvement: Decoupling Narrative Reasoning from Quantitative Computation via an Agentic Execution Layer

What the improved AI system can do: It eliminates numerical hallucination in high-stakes domains (finance, macroeconomics, engineering) by restricting the LLM to qualitative interpretation and delegating all mathematical calculations to a verified, external computational engine. The system generates reports where every numerical claim is a direct reflection of an executed model rather than a probabilistic guess by the language model.

Improvement: Transitioning from Vector-based RAG to a Hybrid Semantic-Mathematical Knowledge Graph

What the improved AI system can do: Instead of retrieving isolated text chunks, the system performs structural reasoning. It can navigate complex relationships between economic shocks, variable concepts, model equations, and temporal validity windows. This allows the AI to understand not just what was said in a document, but how a specific shock (e.g., a change in interest rates) propagates through a specific mathematical framework (e.g., a DSGE model).

Improvement: Implementation of a Post-Execution Write-Back Graph Architecture

What the improved AI system can do: It creates a bidirectional, traceable audit trail. Once an external model is executed, its results (output points, stress metrics, forecast paths) are written back into the Knowledge Graph as new nodes. The AI can then generate a report that is traceable by design, allowing a human auditor to click any statement in a narrative and follow a graph path back to the exact model specification, the specific execution run, and the raw mathematical output.

Improvement: Formalized Model-Selection via Structured Implementation Mapping

What the improved AI system can do: It ensures methodological compliance by preventing the agent from selecting models based on mere linguistic similarity. By using a structured request protocol—mapping a retrieved model specification to an approved implementation function—the system ensures that the AI only uses mathematically sound and organizationally approved computational routines to answer complex analytical queries.

Abstract

We propose an AI economist agent for economic and financial scenario analysis. Scenario design often requires analysts to assess emerging risks with limited historical precedent, combine information from many sources, and translate qualitative mechanisms into internally consistent quantitative paths. Large language models (LLMs) can search and synthesize this information, but fluent narratives alone do not establish the model-based calculations needed for economic conclusions. Our framework uses LLM agents to plan the analysis, retrieve relevant evidence, and organize economic mechanisms, while registered quantitative models generate numerical outcomes and predefined tests determine whether intermediate results can be used in the final report. We apply the framework to European macro-financial stress scenarios and bank capital analysis. The empirical analysis evaluates retrieval of economic mechanisms, scenario construction, model execution, and report generation under a historical information cutoff. The results show how the AI economist agent can combine flexible evidence retrieval and scenario construction while keeping the resulting analysis linked to identifiable sources and explicit model calculations.