Two of them — reinforcement learning and large language models — grew into a course of their own. The other two are simply out of scope for now, though their sources are still in this repository and their decks can be rebuilt.
AI for Economic Research: Dynamic Models, Language, and Agents — ten lectures on deep learning and RL for high-dimensional models, language models as economic measurement, retrieval-augmented generation, and agentic research pipelines. It is the sister course to this one, and it picks up exactly where Topic 12 leaves off.
Stochastic discount factors, the consumption-CAPM and the Lucas tree, the equity premium puzzle; then the empirical pivot from Fama-French to the "factor zoo" of 300+ predictors, and the Gu, Kelly and Xiu (2020) framework treating return prediction as a non-parametric problem where neural networks beat OLS and Lasso. The hands-on session replicated GKX with a rolling training window and a long-short decile portfolio.
Where it went: nowhere — this one is simply
out of scope. The sources remain at source/13_Asset_Pricing/ and the notebook is in
the repository history if you need it. The finance side sits somewhat apart from the rest of the
arc, which is why it was the first to go.
MDPs and Bellman optimality; Monte Carlo and temporal-difference learning; exploration and the on-policy/off-policy distinction; then function approximation, DQN, policy gradients, actor-critic and PPO, applied to optimal taxation, monetary policy and portfolio choice.
Where it went: it became Lecture 5 of the sister course ("Reinforcement Learning in a Nutshell"), expanded into three decks and given a lab of its own. A shorter treatment survives here inside Topic 6, where actor-critic appears as a way of solving macro models, and in Topic 11 for heterogeneous agents.
The mean-field approximation as an answer to the curse of dimensionality; the coupled HJB and Kolmogorov forward equations, solved backward and forward respectively; the upwind finite-difference scheme; stationary equilibrium by fixed-point iteration and transition dynamics by backward-forward shooting.
Where it went: out of scope here, but not gone.
The sources are at source/15-MFG/, and the continuous-time material is touched on in
the sister course's heterogeneous-agent lecture. The discrete-time treatment of the same
economics — distributions as state variables — is
Topic 11 here.
Why text is economic data; tokenization and embeddings from TF-IDF through Word2Vec and GloVe; the transformer and self-attention; pre-training, fine-tuning and prompting; then applications — central bank communication, patents as innovation measurement, text-based forecasting. The hands-on session fine-tuned BERT on FOMC statements.
Where it went: it became two lectures of the sister course — Lecture 7 ("Large Language Models & Text as Economic Data", five decks from measurement through attention and the transformer) and Lecture 8 ("Retrieval-Augmented Generation"), with labs on text-as-data, a mini-GPT, and RAG. Considerably more than the single topic it was here.
Sixteen topics was more than any one course could carry, and the last four pulled in a different direction: the first twelve are about computing equilibrium models, while reinforcement learning and language models are about what AI can do for research. Splitting along that seam gave both halves room. This course now runs from floating point to heterogeneous-agent OLG models without rushing; the AI frontier gets ten lectures of its own next door.
source/13_Asset_Pricing/, source/14_RL/,
source/15-MFG/, source/16_LLM/