AI for Economic Research: Dynamic Models, Language, and Agents

Zhigang Feng

E-mail: z.feng2@gmail.com

Webpage: https://sites.google.com/site/zfeng202/

Format · 課程結構

Six sessions of about three hours each, 32 decks in all. Each session pairs its decks with a hands-on lab. Taught as an intensive it fits into two weeks.

Level: graduate. Every lab runs offline on a laptop CPU with fixed seeds — no API key, no GPU.

Class Meetings — Fall 2026 · 上課安排

Date	Time	Location	Session
Tue, October 13	Morning 9:30–12:30	TBA · 待定	Session 1 — Orientation
Tue, October 13	Afternoon 14:30–17:30	TBA · 待定	Session 2 — Deep learning
Thu, October 15	Morning 9:30–12:30	TBA · 待定	Session 3 — Reinforcement learning
Fri, October 16	Morning 9:30–12:30	TBA · 待定	Session 4 — Heterogeneous agents
Tue, October 20	Morning 9:30–12:30	TBA · 待定	Session 5 — Text and LLMs
Thu, October 22	Morning 9:30–12:30	TBA · 待定	Session 6 — RAG and agents

Course Ecosystem

This course sits at the frontier of a larger body of teaching and research. The complete foundation in classical computational methods — value function iteration, perturbation, projection, EGM, and high-performance computing — is the companion course, Quantitative Macroeconomics with AI and Machine Learning, which has its own slides, labs, and complete video series (free, online) from 2026 and 2025; the same ground is covered in print by the textbook 机器学习与数量宏观经济学 by Zhigang Feng. Together they form a three-stage arc:

Classical Methods → ML & Deep Learning → Agentic AI Research

Students can enter at any stage. Those new to computational economics are encouraged to explore the video series alongside this course; those with classical training will find the bridge to modern AI tools immediately actionable.

Course Description

Economics is being transformed. AI tools that once belonged to computer science labs are now reshaping how economists model complex behavior, extract signal from vast text corpora, and automate the entire research pipeline. This course equips students — at the upper-undergraduate through graduate level — to work at that frontier.

The course runs in six sessions: an orientation to AI as a tool, an economic object and an agent, with the workbench every lab runs on; deep learning, where functions become trainable and equations become losses; reinforcement learning, where decisions are learned from interaction; heterogeneous-agent models, where distributions become state variables; text and large language models as data, measurement and generation; and retrieval and agents in one research workflow. The organizing premise is that as AI absorbs more of the implementation, the economist's edge shifts to designing algorithms and validating results.

Applications span macroeconomics, finance, and economic policy — the text work ranges from FOMC minutes to congressional records, and the agentic tools cover every stage of the research workflow from literature review to replication. The specific topics will evolve as the frontier moves.

The course builds from foundations in dynamic programming and quantitative methods toward the cutting edge of AI research. Students arrive with economic intuition; they leave with the technical fluency and critical judgment to use AI as a genuine research partner.

What You Will Build: Four Integrated Competencies

Theoretical & Algorithmic Mastery — Master economic theory and numerical methods on paper before touching code
Technical Fluency — Working knowledge of Python, PyTorch, and the modern computational ecosystem
AI-Augmented Implementation — Specification-driven development through detailed prompts to AI coding assistants
Critical Validation — Shift from writing code to becoming a rigorous validator of technical correctness and economic plausibility

Prerequisites: Intermediate-level microeconomics and macroeconomics, and comfort with dynamic macroeconomic models. Familiarity with dynamic programming or recursive methods is a plus but not required — the course builds up what it needs. Python experience helps but is not required — Session 1's last deck, The Workbench, teaches the Python the labs use and sets up the tools, and the companion course teaches it from scratch in Topic 3, Programming Basics for Economists. No prior ML experience required.

Course Outline

Six sessions of about three hours each. Every session is a set of decks and ends in a lab; the labs are the part that makes the material stick. Decks are on the slides page, one at a time or a session in one PDF; notebooks on the labs page.

Session 1: Orientation

≈3 hours — AI as tool, as object, and as an agent with incentives. 5 decks, 140 pages, and a lab Lab

1.1 AI in Science: What Has Already Happened — A frontier proof, research effort that scales, a bottleneck that moves, and agents that act on their own.
1.2 The Economics of Producing Intelligence — From the facts of Deck 1.1 to research questions: the factor, its unit, its measurement, and its risks.
1.3 What Is AI? Foundations — A working definition, seventy-five years of history, why now, and what an agent is.
1.4 Workhorse Macroeconomic Models — The models this course solves, the object each one asks for, and where grids stop working.
1.5 The Workbench: Python, VS Code, Git and GitHub — From a clean laptop to Python code that runs again, on any machine.
Lab — Two Machines, One Environment: a neural network learns the growth model's saving policy and is checked against the closed form; a small transformer trained on FOMC statements writes central-bank prose at three temperatures. Its first cells check your workbench. Released before the session.

Session 2: Deep learning

≈3 hours — Functions become trainable; equations become losses. 6 decks, 172 pages, and a lab Lab

2.1 The Learning Problem: a Handwritten Digit, and the Neuron — What it means to learn a function from examples, and the smallest unit that does it.
2.2 Nonlinearity: the Activation, XOR, and What a Network Can Approximate — Why a layer without buys nothing, and which functions the family with it can reach.
2.3 Forward Propagation, Backpropagation, and Training — How the parameters of a network are chosen, one step computed by hand, and when to stop.
2.4 PyTorch, Curve Fitting, and Yield Forecasts — The training loop in code, a curve fitted and broken, and an honest one-month-ahead forecast.
2.5 Solving the Growth Model as Supervised Learning on the Euler Equation — The network is the policy, sampled states are the inputs, and the Euler equation supplies the labels.
2.6 Architecture as a Constraint on W, and What More Parameters Buy — Restrictions that encode what we know about the input, and the opposite move: more parameters than observations.
Lab — From One Neuron to a Solved Growth Model: tensors and autograd with backpropagation checked by hand; a curve fitted and broken outside its data; next month's 10-year Treasury yield against a random walk; the growth model solved by supervised learning on its Euler equation. Released before the session.

Session 3: Reinforcement learning

≈3 hours — Sequential decisions are learned from interaction. 6 decks, 162 pages, and a lab Lab

3.1 The Growth Model, With One Thing Hidden — Dynamic programming works, and says exactly what it needs; this deck is what happens when it cannot have it.
3.2 The Simplest Actor–Critic: How It Works, Why It Works, and the Wall It Meets — Nine numbers on the continuous growth model, with the shock probabilities hidden.
3.3 Monte Carlo: Evaluation, Control, and the Data — Average simulated returns to value a rule; then find out what improving a rule from its own data costs.
3.4 Temporal Difference: TD(0), SARSA, Q-Learning — Update after every step by leaning on the estimate you already have: when that is safe, and what it costs.
3.5 From Tables to Networks: Function Approximation, Policy Gradient, Actor–Critic — Parameters replace the table; then a gradient step replaces the max.
3.6 The Growth Model Across Algorithms — The full table Deck 3.2 opened; then actor–critic on the continuous model, and the bridge to Session 4.
Lab — One Growth Model, Every Reinforcement-Learning Method: an environment solved exactly, then Monte Carlo, TD, SARSA, Q-learning, linear TD and a small DQN, policy gradient and actor–critic — every learner graded against the exact dynamic-programming solution. Released before the session.

Session 4: Heterogeneous agents

≈3 hours — Distributions and equilibrium become state variables. 4 decks, 168 pages, and a lab Lab

4.1 Aiyagari, Solved Twice: A Fixed Point, Then a Law of Motion — The classical solve returns one stationary distribution. An actor–critic solve with the distribution as its state returns how every distribution evolves.
4.2 Krusell–Smith, Solved Three Ways — Generalised moments, a histogram, a history of shocks: three ways to put a moving distribution into a network, and what each can and cannot see.
4.3 Cohorts Under Aggregate Risk — From Krusell–Smith to overlapping generations: a finite state, a network that clears a market, and why idiosyncratic risk sends us to histories.
4.4 Optimal Taxation Without Commitment — Put a government into Deck 4.1's economy: a tax rule on the whole distribution, a Markov-perfect fixed point in rules, and the algorithm that finds it.
Decks 4.1, 4.2 and 4.4 include pages marked Preliminary, based on research in progress by Feng, Han, Sargent and Zhu (September 2026): read them for the course and do not circulate them further.
Lab — Distributions as State, Solved Two Ways: classical answer keys (Aiyagari by the endogenous grid method, Krusell–Smith, a transition path, a life-cycle economy), then PyTorch networks that carry a distribution: Deep Equilibrium Nets, the sequence space, and an OLG economy. Released before the session.

Session 5: Text and LLMs

≈3 hours — Language becomes data, measurement, and generation. 6 decks, 208 pages, and a lab Lab

5.1 Text as Economic Data: From Documents to Counts — Clean, tokenize, count: how a document becomes a vector a regression can use.
5.2 Meaning in Vectors: Embeddings, Attention, the Transformer — Embed, then contextualize: from one vector per word to one vector per word in its sentence.
5.3 Pretrained Language Models: BERT, GPT, Alignment, and Five Limits — What pretraining buys, what alignment changes, and the five limits that follow from the architecture.
5.4 The LLM as a Measurement Instrument — What it can label, what the labels are worth, and how to correct what they get wrong.
5.5 Simulated Agents, Prompts as Design, and Reproducible LLM Research — What a model can stand in for, why the prompt is a confounder, and what has to be reported.
5.6 Text Signals and the Bond-Yield Forecast — Four signals, one frozen baseline, one scoring of the test window, and the result as it came out.
Lab — Words as Data, and What a Language Model Measures: a vocabulary built twice; word vectors and a hawk–dove axis that fails validation; attention by hand; a recorded language-model instrument read for agreement and bias; and a text-signal forecast of the 10-year yield. Released before the session.

Session 6: RAG and agents

≈3 hours — Retrieval, tools, traces, and governance in one workflow. 5 decks, 176 pages, and a lab Lab

6.1 From Chat Window to Agentic AI: What an Agent Is, and Where It Pays — Part one of five: chatbot, copilot, agent, agent team — what makes AI agentic, and one year of an agent on a cluster, in the instructor's own projects.
6.2 How Agentic AI Works: A Loop Built From Scratch, the Harness, the Session, and the Team — Part two of five: the loop in thirty lines you could write yourself, each of the four marks in code, the production harness that runs the same loop, what a session carries and costs, and how one agent becomes a team.
6.3 How to Build With Agentic AI: The System, the Homotopy Method, a Research Case, and the Checks — Part three of five: set up a research system of agents you can trust — the repository, the team, the gates — then grow a project with it one validated margin at a time, as a months-long research project did.
6.4 Build Your Own RAG: Give Your Agent a Library — Part four of five: why an agent cannot read your literature, what naive use costs, and a six-step recipe from a folder of PDFs to a cited answer — tied to Session 5 at each step.
6.5 GRAM: A Retrieval System Built for One Literature –- and What Agents Do to Research — Part five of five: rung three, how it was built and how the method transfers; then the close: the two dangers answered, the evidence on the researcher, and the economics of an agent.
Lab — An Agent You Can Audit, and the Retrieval It Reads From: one bounded question on 245 FOMC statements and a yield panel, answered with a citation — the agent loop read line by line, the retrieval it reads from, and the governance around it. Released before the session.

Supplementary Resources

Foundations: Classical Methods — Companion Course, Video Series & Textbook

This course is part of an integrated curriculum. Three foundational resources cover the classical side of the arc, all freely available:

Companion course — Quantitative Macroeconomics with AI and Machine Learning, the sister site to this one. Twelve topics with slides and labs: numerical methods, dynamic programming and accuracy assessment, perturbation and projection, high-performance computing, and machine learning for heterogeneous-agent and OLG models. Start here if you want the classical foundation this course assumes:

https://vonzhg.github.io/Quant_Macro/

Video Lecture Series — the same course taught on camera: value function iteration, time iteration, endogenous grid method (EGM), perturbation methods, projection methods (Chebyshev), numerical optimization and interpolation, and high-performance computing (MPI, GPU). Nine lectures, about nine and a half hours:

https://space.bilibili.com/2142649036/lists/7180709

Textbook — 机器学习与数量宏观经济学 by Zhigang Feng bridges classical numerical methods and modern deep learning for dynamic economic models. It is the primary reference for Sessions 1–4 and connects to the AI frontier covered in Sessions 5–6:

https://book.douban.com/subject/37885381/

References

Judd, Kenneth. Numerical Methods in Economics.
Ljungqvist, Lars, and Thomas J. Sargent. Recursive Macroeconomic Theory.
Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. Deep Learning.
Bertsekas, Dimitri. Reinforcement Learning and Optimal Control.
Feng, Zhigang. 机器学习与数量宏观经济学.
Santos, Manuel S. Numerical Solution of Dynamic Economic Models.
QuantEcon Lectures