The PDF is authoritative. What follows is the same content in brief, for quick reference.
Artificial intelligence enters this course twice. As a tool, it is an extremely powerful instrument for macroeconomic research: neural networks and reinforcement learning solve models that grids cannot, language models turn text into data, and AI assistants and agents write code, search the literature and run analyses. As an object, AI is a transformative economic force that deserves serious macroeconomic research — a new kind of capital, a shock to which tasks people do and what they earn, and a question for policy. The models and methods of this course are the ones needed to study who gains and who loses from it.
When AI can write most of the code, the skills that matter move up a level: which question to ask, which model answers it, which method is accurate enough, which data make it relevant, and whether the answer is right. The course trains exactly these.
| Instructor | Zhigang Feng |
|---|---|
| Meets | Thursdays 9:00–11:50, Sproul Hall 4128 |
| Blocks | A 9:00–10:15 · break · B 10:35–11:50 |
| Term | Ten sessions, September 24 – December 3. No class November 26. |
| Discussion | The course forum (UCR e-mail sign-in): post under your name or anonymously. Anonymous posts are anonymous to classmates; instructors can see who wrote them, except in the anonymous-feedback category. E-mail is for private matters. |
The first-year Ph.D. macroeconomics sequence (or equivalent), including dynamic optimization. Programming experience helps but is not required: Python is taught from the ground up in Session 3, and a setup lab is posted before then. Students from other fields are welcome with the instructor's consent.
| Session | Block A | Block B | Due / out |
|---|---|---|---|
| 1 Sep 24 | Introduction. Who I am and why this course; the macro research frontier; AI as a tool and as an object of macro research. | The optimal growth model. From Solow to the planner; Euler equations and steady states; recursive formulation and dynamic programming. | |
| 2 Oct 1 | Incomplete markets. Bewley and Huggett; Aiyagari; stationary recursive competitive equilibrium; the stationary distribution. | Aggregate risk, finite lives, policy. Krusell–Smith; overlapping generations; Ramsey taxation, time consistency, Markov-perfect policy. | |
| 3 Oct 8 | Introduction to computation. What it means to solve a model; value function iteration on a grid; Euler-equation methods and time iteration. | Programming basics. Python, NumPy and PyTorch from the ground up; vectorization; automatic differentiation; coding with AI assistants. | HW 1 Part A out |
| 4 Oct 15 | Numerical methods I. Root-finding and fixed points; optimization in one and many dimensions; numerical and automatic differentiation. | Numerical methods I (cont.). Approximation and interpolation; quadrature; discretizing AR(1) processes; a fast and accurate stochastic growth model. | HW 1 Part B out |
| 5 Oct 22 | Numerical methods II: perturbation. First- and second-order perturbation; Blanchard–Kahn; linear RE systems; pruning. | Numerical methods II: projection. Chebyshev polynomials, Smolyak grids, finite elements; collocation vs. Galerkin; the same model solved four ways. | |
| 6 Oct 29 | Heterogeneous agents I: stationary equilibria and aggregate risk. Solving Aiyagari as a nested fixed point; the stationary distribution; Krusell–Smith; OLG and transition dynamics; optimal policy with heterogeneous agents, and why grids and moments break down. | Heterogeneous agents II: when heterogeneity aggregates. Complete markets and Gorman aggregation; the Negishi method as the representative-agent construction; transitional dynamics by shooting, and what is lost when markets are incomplete. | HW 1 due; HW 2 Part A out |
| 7 Nov 5 | Machine learning essentials. Learning as function approximation; neural networks; training, backpropagation, generalization. | Deep learning for dynamic models. Euler-residual minimization; deep VFI and actor–critic; structured networks and validation. | Project teams and topics |
| 8 Nov 12 | Reinforcement learning. RL as dynamic programming without the model; Monte Carlo, TD, Q-learning; policy gradients. | Heterogeneous agents at the frontier. The distribution as a network input; deep equilibrium nets and DeepHAM; optimal taxation without commitment. | HW 2 Part B out |
| 9 Nov 19 | Language models and text as data. From models to data; text as measurement; words to vectors; attention and the Transformer. | LLMs as research instruments. Limits and validation; measuring the economic footprint of AI; retrieval-augmented generation. | Proposal due Mon Nov 16; |
| Monday, November 23 — HW 2 due. November 26 — no class (Thanksgiving). | |||
| 10 Dec 3 | Agentic AI and the future of research. How AI agents work; building a research workflow; verification, governance, and what remains human. | Final project presentations (mini-conference with discussants). | |
| Dec 10 | Final paper and replication package due (finals week). | ||
| Component | Weight | Timing |
|---|---|---|
| Homework 1 — from a question to a quantitative answer | 15% | out Oct 8 and Oct 15; due Oct 29 |
| Homework 2 — heterogeneity and learning | 15% | out Oct 29 and Nov 12; due Mon Nov 23 |
| Final project (teams of 2–3) | 40% | |
| Proposal (two pages) | 5% | due Monday, Nov 16 |
| Presentation and discussion | 15% | Dec 3, Block B |
| Paper (≤ 10 pp.) and replication package | 20% | due Dec 10 |
| Participation | 30% | throughout — exit tickets, discussion in class and on the forum, discussant role |
AI_LOG.md with each assignment recording how you
used it. The validation is the part you cannot delegate — and it is what the course
is actually about.
Every computational result in this course is reported with: the method; the grid (points, spacing, bounds); the chain and its implied persistence and standard deviation; the tolerance; the dtype; the maximum and mean Euler residuals on points not used to solve; and the wall time with the hardware. A number without these is not yet a result.