AI for Econometricians

The Translation Rule

Start from the statistical object you already know.

For each AI concept, first identify the familiar econometric object. Then identify what becomes learned, nonlinear, sequential, conditional, or decision-dependent.

This course is for learners with an econometrics or closely related quantitative background who want to understand and investigate modern AI. You need not discard what you know about likelihood, regression, latent variables, forecasting, or identification. Much of that mathematics reappears in AI, sometimes under different terminology and sometimes in a significantly different computational form.

The aim is translation with precision: use familiar statistical reasoning where it still applies, recognize where an analogy is only partial, and learn the new machinery needed to understand contemporary AI systems.

Who this course is for

The starting point is typical applied-econometrics training: probability and statistical inference, OLS and nonlinear regression, specification and model evaluation, and some exposure to panel data, instrumental variables, causal reasoning, or time series. Your specialization may instead emphasize financial forecasting, microeconometrics, latent-variable modeling, or Bayesian analysis.

You do not need prior training in backpropagation, modern stochastic optimizers, MCMC algorithms, deep learning, transformers, Bellman equations, or policy-gradient methods. These are developed when needed. Module 0: Orientation reviews the assumed starting knowledge in more detail.

What you will learn

By the end of the course, you should be able to:

• identify the familiar probabilistic or statistical structure inside an AI method and explain where the analogy breaks;
• distinguish parameter inference, causal identification, prediction, and sequential decision-making, and select evidence appropriate to the question;
• explain how learning extends from coefficients to representations, states, embeddings, values, and policies;
• follow the mathematical and computational logic of neural networks, backpropagation, sequence models, attention, and transformers;
• explain how self-supervision, data construction, scale, and hardware constraints affect modern model design;
• connect Bayesian and Markov models to decision processes, reinforcement learning, LLM post-training, and agentic systems;
• use small, reproducible experiments to check mechanisms and evaluate limitations rather than relying on architectural labels.

The goal is conceptual and implementation literacy, not a survey of every AI architecture or a collection of API tutorials.

How the course works

Each module begins with an econometric or statistical bridge, introduces the corresponding AI object and its mathematical core, identifies what is learned and optimized, and tests the mechanism in a small offline browser lab and matched Python/R examples. Every module ends in an explanation, diagnostic, or empirical check.

The five-question translation rule is defined in Module 0. That is the canonical version used by the rest of the course; this page does not establish a competing checklist.

The learning route

The 15 modules (numbered 0–14) form a connected path, not a claim that each subject belongs to only one branch of mathematics.

• Orient and build probabilistic foundations (Modules 0–1): translate familiar econometric objects; then examine Markov structure, hidden states, and MCMC.
• Learn representation and prediction machinery (Modules 2–7): n-gram language models, embeddings, neural networks, backpropagation, supervised learning, and unsupervised learning.
• Learn modern sequence and language architectures (Modules 8–11): recurrent/sequence models, attention and transformers, foundation models, and retrieval-augmented generation.
• Connect learning to decisions and interaction (Modules 12–14): reinforcement learning, LLM post-training, and reasoning systems that use tools, memory, and feedback.

Two recurring mathematical paths help you navigate these topics: probabilistic structure and sequential decisions and function approximation and learned representations. They overlap extensively; optimization, data, and evaluation connect them. Module 0 introduces the map after explaining why these connections matter.

Evidence and evaluation conventions

A statistical claim, a predictive claim, a causal claim, and a policy claim require different evidence. Throughout the course:

• Training data are used to fit parameters; validation data support model, hyperparameter, and threshold choices; test data evaluate performance after these choices are fixed.
• For time-ordered data, preprocessing, fitting, and tuning must respect chronology, so future information does not leak backward.
• A low predictive loss does not establish causal identification. Structural and causal interpretations need defensible assumptions or research designs.
• Deployment may differ from development data. In decision systems, the chosen policy can also change which observations are collected next.
• Results should be interpreted alongside computation, memory, latency, data quality, and robustness constraints.

The module exercises make these distinctions concrete rather than treating evaluation as a single accuracy score.

Curricular anchors

The econometrics starting point is illustrated by Stock and Watson’s Introduction to Econometrics and Wooldridge’s Introductory Econometrics: A Modern Approach. The broad AI backbone is Russell and Norvig’s Artificial Intelligence: A Modern Approach (4th edition). These provide reference points, not required cover-to-cover reading.

Module pages cite the portions relevant to their own material and distinguish sources that inform the lesson from optional further reading. Begin with Module 0 for the precise translation framework and orientation readings.