AI_LOG.md recording how you used AI
assistance. Every number you report must come from code you ran.
15% of the grade. Part A out October 8, Part B out October 15; both due Thursday, October 29. A small research project on Sessions 1–5, in which you are the project manager: one model, a positive question (how much of U.S. business-cycle volatility do TFP shocks explain?) and a normative one (what do aggregate fluctuations cost in welfare?), taken from the charter to the paper.
A one-page project charter; the planner problem, the recursive competitive equilibrium and what a representative agent cannot say; a unit test of your solver against the closed-form growth model; and your own data targets and calibration from FRED, with an AI-assisted provenance table you verify by hand.
Choose the algorithm on the evidence; solve the
calibrated model by a global method, first-order perturbation and Chebyshev collocation
(Session 5 parts after October 22); answer both questions; and deliver a six-page paper, five
slides, a one-command replication package, and an AI_LOG.md with a validation
ledger. Question B2(d) feeds the anonymous class accuracy leaderboard. You also choose
one route to go deeper — R1 the numerical frontier, R2 data and
estimation, or R3 capital-income taxation — and end by pitching your own question on the
forum's Pitches and teams board. Part A includes an
AI red-team item: find and document an error in an AI assistant's derivation.
15%. Part A out October 29, Part B out November 12; due Monday, November 23. Sessions 6–9. A common core — the Aiyagari economy as a nested fixed point, a neural-network solution validated against a classical one, a first reinforcement-learning exercise, and a first text-as-data measurement — plus one route of your choice: aggregate risk and the welfare cost of cycles with heterogeneity, fiscal policy and transitions, AI solution methods, or AI and inequality. You also audit a classmate's Homework 1 replication package, and red-team an AI assistant once more.
40%, in teams of 2–3. The project assembles what the homework built: combine at least one Homework 1 component and one Homework 2 component from your team, validate each against its homework benchmark, and add what your team's question still needs — a model, a data source, or a method — with one classical and one AI/ML method cross-validated on the same object.
| Milestone | Weight | Due |
|---|---|---|
| Pitch posted on the forum (Homework 1, B5) | — | October 29 |
| Teams and topics, on the Pitches and teams board | — | November 5 |
| Two-page proposal — question, components reused, what you add, division of labor, risks | 5% | Monday, November 16 |
| Homework 2 components ready | — | Monday, November 23 |
| Presentation, and serving as discussant for another team | 15% | December 3, Block B |
| Paper (≤ 10 pp.) and replication package | 20% | December 10 |
The replication package must reproduce every number and figure with one command. Projects are graded on framing, integration and correctness, validation, reproducibility and provenance, quality of AI collaboration, communication, and the discussant report.
30%. There is no examination. Exit tickets at the end of each session, engagement in discussion in class and on the forum, the live demos, and your role as discussant at the final-project conference.