Instructors

Sophia Kazinnik (Stanford University)

Dates

17-21 August 2026

Hours

15:00 to 18:30 CEST

Format

In person

Practical Classes

A (voluntary) introductory class on Python will be taught by a teaching assistant during the first day of the course.

Intended for

Researchers, central bank economists, and PhD students who want to use large language models (LLMs) in data-driven research. Researchers working at the intersection of monetary and financial economics who want to go beyond basic text analysis and use LLMs for things like better measurement, generating new research ideas or data, and running agent-based simulations.

Prerequisites

A basic familiarity with coding, econometrics, and statistics. The lectures will use Python notebooks prepared in advance. No prior deep learning experience is required. The course is designed for applied researchers who want to build reproducible workflows with AI that work in real-world settings (including working with confidential data) and who are interested in using LLMs for measurement, generation, and agent-based simulation, particularly in the context of monetary and financial economics.

Overview

Monetary and financial economics increasingly hinge on language and communication, whether we are trying to decode central bank signals, measure narratives in markets, or understand how households and firms form expectations. This course focuses on three complementary ways modern AI can be used in research:

• Extracting economic signals from text (and, when relevant, audio/video): turning complex documents and texts into interpretable measures that can enter standard empirical designs (e.g., stance, uncertainty, risk, guidance, narrative shifts, etc.).
• Producing structured research outputs with generative models: using LLMs to create objects economists actually use: forecast variables, scenario consistent summaries, structured extractions, synthetic survey responses, counterfactual message variants, and scalable measurement targets.
• Simulating economic actors and institutions: building agent-based systems (e.g., forecasters, committee members, depositors, etc.) that interact and deliberate under controlled information sets, creating an “in silico” laboratory for counterfactual policy and market experiments.

A core theme throughout is measurement discipline: we treat LLMs as powerful but fallible instruments and emphasize validation, data leakage safeguards, and error correction, especially when models are used to generate labels or behavioral outcomes at scale.

There are three parts to the course:

• Text as Data (Measurement as Data Creation). Practical approaches for extracting economic signals from text: stance, risk, uncertainty, narratives, and regime shifts, paired with transparent evaluation and error analysis.
• Generation as a Research Tool. Using LLMs to generate research-relevant outputs (e.g., forecast objects, structured extractions, synthetic survey responses, and scalable measurement targets) while being able to validate and “explain” these outputs and enforce constraints when needed.
• Agents and Simulations. Building agent-based workflows that combine personas, real-time/vintage data, and interaction protocols to study phenomena such as expectations formation and decision-making, linking micro-level reasoning and communication to macro/market outcomes.

Topics

• AI for central banking and financial texts: stance and intent classification; information extraction; narrative and uncertainty measurement; explainability and systematic evaluation
• LLMs as measurement instruments: human anchoring, bias correction, robustness to prompts/models, and safeguards against data leakage
• Generative AI for research workflows: structured extraction and generation (schemas/JSON); forecast generation; synthetic survey responses and scalable labeling; retrieval augmented generation (RAG); model finetuning
• Agents and multi-agent simulations: personas, controlled information sets, deliberation protocols, and counterfactual experimentation for policy and markets, model steering
• Best research practices: reproducibility, versioning, and deployment constraints in applied policy/finance environments

Sophia Kazinnik is a Research Scientist at Stanford’s Digital Economy Lab (HAI), where she builds generative AI systems to explore how language and behavior shape economic outcomes. Her work turns economic questions into computable experiments, using LLM-powered agents and multi-agent simulations to study financial fragility, policy communication, and market expectations. In some of her recent projects, she has modeled bank runs, simulated FOMC deliberations, and evaluated how today’s AI interprets central bank language.

Before joining Stanford, Sophia spent seven years as a Financial Economist & Quant at the Federal Reserve, where she reviewed stress test models and developed natural language tools for bank supervision. Across her research on AI-augmented surveys, simulated professional forecasters, and the analysis of verbal and nonverbal cues in central bank communication, her goal is to make AI a lab for studying economic behavior, and a tool for designing better policy.