Econometrics bridge: Probability, likelihood, regression, latent state, prediction, optimization, and decision theory
Estimated time: 90 min
Lab: Open browser lab
Code: Python - R

This course starts from what you already know

An econometrician approaching modern AI is not starting from zero.

Probability models, conditional expectations, likelihoods, regression, latent variables, time dependence, out-of-sample evaluation, and decision-making under uncertainty already appear in econometrics, statistics, finance, and related quantitative fields. Modern AI reuses many of these objects, but changes what is learned, how flexibly it is parameterized, how much computation is used, how training data are constructed, and whether the fitted system only predicts or also acts.

You can begin by recognizing four kinds of connections:

  1. what is already familiar;
  2. what has been renamed or reframed;
  3. what is only an analogy;
  4. what genuinely new machinery AI introduces.

This module provides that translation map.

What we assume an econometrician may already know

A conventional applied econometrics sequence such as Stock-Watson or Wooldridge typically develops a core progression resembling:

$$ \text{probability/statistics} \rightarrow \text{OLS} \rightarrow \text{inference} \rightarrow \text{nonlinear models} \rightarrow \text{panel/IV} \rightarrow \text{prediction and time series}. $$

The exact sequence varies by textbook and program, but the common foundation usually includes:

  • probability and sampling;
  • linear regression and least squares;
  • sampling distributions, hypothesis tests, and confidence intervals;
  • nonlinear transformations and interactions;
  • heteroskedasticity and specification problems;
  • panel data and fixed effects;
  • instrumental variables and two-stage least squares;
  • binary or limited dependent-variable models;
  • experiments and quasi-experimental reasoning;
  • time-series regression and forecasting.

Modern econometrics texts may also introduce material that is commonly associated with machine learning. Stock-Watson, for example, includes predictive methods such as trees, logistic regression, neural networks, cross-validation, nearest neighbors, clustering, ROC/AUC evaluation, Naive Bayes, TF-IDF, and n-grams.

Different econometricians also arrive with different specializations.

A finance econometrician may already know:

$$ \text{ARIMA},\quad \text{GARCH},\quad \text{forecast evaluation}. $$

A microeconometrician may know:

$$ \text{panel data},\quad \text{IV},\quad \text{DiD},\quad \text{selection models}. $$

Someone trained in structural equation modeling may already be comfortable with latent variables. A Bayesian econometrician may already think in terms of priors, posteriors, and posterior predictive distributions.

But this course does not assume that the learner already knows:

  • gradient descent or modern stochastic optimizers;
  • backpropagation;
  • modern kernel methods;
  • MCMC algorithms;
  • deep neural-network architectures;
  • attention and transformers;
  • Bellman equations;
  • stochastic control;
  • policy gradients;
  • reinforcement-learning algorithms.

The goal of this orientation is to make those starting points explicit—not to assume that every econometrician has followed the same specialization. Numerical and AI-specific techniques will be introduced when they become necessary.

Two traditions, overlapping mathematics

Econometrics and modern AI did not develop as a single unified curriculum. Econometrics, statistics, machine learning, control, operations research, signal processing, and AI developed partly separately, with different research questions, terminology, computational practices, and disciplinary conventions.

As a result, an econometrician can encounter an AI concept that is mathematically familiar but expressed in unfamiliar language.

Sometimes the connection is very close:

Econometrics / statistics AI / machine learning Shared idea
OLS fitted values and residual loss empirical risk minimization choose parameters to reduce an average loss
Logit/probit + MLE classification + cross-entropy estimate conditional outcome probabilities
Forecast MSFE/MSPE out-of-sample predictive loss evaluate performance on unseen observations
Ridge/LASSO regularization / weight decay trade fit against model complexity
Numerical MLE gradient-based optimization optimize a differentiable objective
N-grams language models estimate conditional probabilities over sequences

Sometimes the relationship is a useful structural analogy rather than an equivalence:

Econometrics / statistics AI / machine learning Useful bridge
PCA / factor models representation learning / embeddings compress information into lower-dimensional structure
Dummy / fixed-effect lookup embedding lookup table map discrete identities to learned numerical quantities
AR / ARDL sequence prediction use previous observations to predict later ones
Moving averages / distributed lags convolution combine information over a local window
State-space models recurrent hidden state propagate a state recursively through time
Time-varying weighted averages attention form context-dependent weighted combinations
Finite mixture / regime models mixture of experts combine specialized components through latent or learned routing
Bootstrap bagging / ensembles reuse resampled data to construct multiple fitted models
Dynamic choice / stochastic control MDP / reinforcement learning choose actions while accounting for future consequences

These bridges save time, but they must be used carefully.

An RNN is not simply an AR model. Attention is not merely a moving weighted average. An embedding is not simply a fixed effect. A neural network is not merely nonlinear regression with more coefficients.

The analogy identifies the familiar mathematical skeleton. The learner must still ask what has changed.

The translation rule used throughout the course

For every AI concept, ask five questions:

  1. What is the familiar statistical object?
    A conditional distribution, latent state, likelihood, weighted average, dynamic recursion, loss function, or decision rule?

  2. What became learnable?
    Coefficients, features, latent state, representations, routing weights, value functions, or the policy itself?

  3. What objective is optimized?
    Likelihood, squared error, classification loss, reconstruction error, expected return, or a preference objective?

  4. What generates the evidence?
    A fixed observational dataset, an experiment, a self-supervised transformation, retrieved documents, or trajectories generated by the current policy?

  5. Where does the analogy break?
    Which familiar coefficient, factor, state, or causal interpretation no longer carries its old meaning?

This translation rule is more important than memorizing one-to-one terminology. Many of the most useful bridges in this course are deliberately partial.

A central shift: more of the representation becomes learnable

Consider a familiar regression:

$$ y_i=\beta_0+\beta_1x_{i1}+\beta_2x_{i2}+\epsilon_i. $$

The researcher normally decides what \(x_{i1}\) and \(x_{i2}\) mean before estimation. The model learns the coefficients.

Modern representation learning pushes much more of that construction inside the fitted model. Instead of receiving only researcher-defined features, the model may learn an internal transformation

$$ h_\theta(x) $$

and then use that learned representation for prediction:

$$ \hat y = g_\theta(h_\theta(x)). $$

This is one reason neural networks, embeddings, CNNs, recurrent networks, attention, and transformers can look conceptually different from classical econometrics even when familiar loss functions and probability models remain underneath.

The key change is not merely “linear versus nonlinear.” It is that features and representations that were previously supplied by the researcher can themselves become objects of estimation.

Econometric questions and predictive questions are often different

Econometrics and AI overlap heavily, but much of modern applied econometrics emphasizes estimation and causal identification, whereas much of modern machine learning emphasizes out-of-sample prediction and decision performance.

Causal or structural question

An econometrician may ask:

What is the causal effect of \(X\) on \(Y\)?

In a simple regression,

$$ Y=\beta X+\epsilon, $$

the key issue is not only whether the fitted model predicts \(Y\). The interpretation of \(\beta\) depends on assumptions such as

$$ E[\epsilon\mid X]=0, $$

or on another identification strategy.

This leads to concerns such as:

  • omitted variables;
  • simultaneity;
  • selection;
  • instrumental variables;
  • experiments;
  • difference-in-differences;
  • regression discontinuity;
  • other identification assumptions.

Predictive question

A machine-learning system often asks:

For a new observation with \(X=x\), how accurately can I predict \(Y\)?

The target is closer to expected predictive risk:

$$ R(f) = E_{(X,Y)\sim P_{\text{future}}} \left[ L(Y,f(X)) \right]. $$

Now the central concerns become:

  • generalization;
  • train/validation/test separation;
  • overfitting;
  • regularization;
  • cross-validation;
  • hyperparameter selection;
  • distribution shift.

A model can predict well without identifying a causal effect. Conversely, a credible causal design may answer an economically important question even when it is not the best pure prediction system.

Keep four tasks separate:

Task Central question
Parameter inference What does a specified statistical model imply about an estimand and its uncertainty?
Causal identification Under what assumptions can an intervention or counterfactual effect be learned?
Prediction How well does a fitted rule generalize to new observations?
Sequential decision Which action maximizes expected utility when actions affect future observations?

A flexible AI model does not eliminate endogeneity or identification problems when the question is causal.

Computation becomes part of the model-design problem

In much traditional applied econometrics, computational details can often remain in the background. A researcher specifies a regression, calls an estimator, and focuses on identification, inference, or interpretation.

Modern AI frequently makes computation itself part of the modeling problem.

Practical choices may depend on:

  • parameter count;
  • accelerator memory;
  • batch size;
  • training time;
  • inference latency;
  • context length;
  • communication cost;
  • numerical precision;
  • quantization;
  • parallelism.

A model that is statistically attractive but impossible to train, store, or serve within available resources is not operationally useful.

A useful high-level view is therefore:

$$ \boxed{ \text{statistical quality} + \text{computational feasibility} + \text{data availability} } $$

This perspective will later help explain why AI research cares so much about architecture, sparsity, attention cost, parameter-efficient fine-tuning, mixture-of-experts routing, quantization, and hardware-aware design.

AI also changes how training data are constructed

Econometric data are usually generated by households, firms, markets, governments, experiments, or other economic processes. The researcher observes a limited sample from that data-generating process and generally cannot manufacture arbitrary new independent observations.

Modern AI also depends on externally generated data, but digitization produces enormous quantities of text, images, audio, video, code, sensor records, and interaction logs.

The more important difference is that self-supervision can turn raw digital artifacts into training targets.

For example, the text

The central bank raised rates today.

can create multiple next-token prediction problems:

$$ P(\text{central}\mid \text{The}), $$
$$ P(\text{bank}\mid\text{The central}), $$
$$ P(\text{raised}\mid\text{The central bank}), $$

and so on.

The world supplied the text, but the training procedure automatically constructed many input-target pairs from it.

This does not mean useful AI training data are infinite. Data quality, duplication, domain coverage, privacy, licensing, contamination, and computational cost remain important constraints.

The key lesson is that advances in representation learning, self-supervision, storage, and computation can make previously unusable digital information economically and statistically useful as training evidence.

Prediction is not the end: AI can also make decisions

A predictive model estimates something about the world. A decision system uses information to choose an action.

A simple decision problem can be written as:

$$ a^* = \arg\max_a E[U(a,Y)]. $$

This should already feel familiar to learners exposed to optimization, decision analysis, portfolio choice, operations research, or linear programming.

Sequential decision-making adds an important complication: today’s action may change tomorrow’s state and therefore tomorrow’s data.

A Markov decision process represents this using

$$ p(s_{t+1}\mid s_t,a_t) $$

and a reward

$$ r_t=r(s_t,a_t,s_{t+1}). $$

The goal becomes finding a policy

$$ \pi^* = \arg\max_\pi E_\pi \left[ \sum_{t=0}^{T}\gamma^t r_t \right]. $$

Reinforcement learning asks how values, models, or policies can be learned when the sequential decision problem is not completely solved in advance.

This connects statistics and learning directly to action:

$$ \text{estimate or predict} \rightarrow \text{evaluate consequences} \rightarrow \text{choose action} \rightarrow \text{observe new state} \rightarrow \text{learn again}. $$

This feedback loop becomes especially important in reinforcement learning and agentic AI.

What these connections prepare you to understand

The comparisons above are not an end in themselves. They prepare you to answer increasingly demanding questions as you move through the course:

  • From probability to evolving states: How do conditional independence and latent-state models help reason about systems that change over time?
  • From regression to learned representations: What changes when the model learns intermediate features rather than only coefficients on researcher-specified variables?
  • From forecast evaluation to training systems: How are loss functions, gradient-based optimization, self-supervision, and evaluation combined at scale?
  • From distributed lags to sequence architectures: Which pieces of recurrence, local filtering, and weighted aggregation survive in CNNs, RNNs, attention, and transformers?
  • From prediction to action: How do decisions, rewards, and state transitions transform a predictive model into an interactive policy?
  • From statistical models to AI systems: How do LLMs, retrieval, post-training, and tool use combine these ideas, and where can their outputs fail despite an apparently capable model?

You do not need to master all these topics in Module 0. You need to recognize the questions and see how your existing knowledge provides useful entry points. That gives the course map below a purpose: it shows which mathematical path develops each connection.

The mathematical path through the course

The route uses two recurring mathematical spines as navigation aids—not as a taxonomy of mutually exclusive AI methods. In particular, learning, inference, optimization, and decision-making frequently occur together.

Spine A: probabilistic structure and sequential decisions

Start with a joint distribution and conditional independence:

$$ p(x_1,\ldots,x_d) = \prod_i p(x_i\mid \operatorname{pa}(x_i)). $$

This leads to Bayesian networks, Markov blankets, exact and approximate inference, Markov chains, hidden-state models, and dynamic Bayesian networks.

When actions and rewards are added, the state-transition structure becomes a Markov decision process:

$$ p(s_{t+1}\mid s_t,a_t), \qquad r_t=r(s_t,a_t,s_{t+1}). $$

A simplified course path is:

$$ \text{probability} \rightarrow \text{conditional independence} \rightarrow \text{latent/temporal state} \rightarrow \text{Markov processes} \rightarrow \text{MDPs} \rightarrow \text{reinforcement learning}. $$

Spine B: function approximation and learned representations

Start from estimation through a loss or likelihood:

$$ \hat\theta = \arg\min_\theta \frac{1}{n} \sum_i L(y_i,f_\theta(x_i)). $$

OLS, logit, and other familiar models restrict the form of \(f_\theta\). Modern neural networks make the fitted function far more flexible and increasingly learn the representation used internally.

Backpropagation makes large compositions trainable. Self-supervised objectives turn raw data into training signals. Attention and transformers allow context-dependent interaction among learned representations and support modern foundation models.

A simplified course path is:

$$ \text{loss / likelihood} \rightarrow \text{function approximation} \rightarrow \text{learned representations} \rightarrow \text{deep neural networks} \rightarrow \text{attention / transformers} \rightarrow \text{foundation models}. $$

Where the spines meet

Deep reinforcement learning combines sequential decision structure with neural function approximation.

LLM post-training combines a pretrained language policy with preference data, rewards, value estimates, reference policies, or direct preference objectives.

Agentic systems add retrieval, tools, memory, planning, and repeated interaction with an external environment.

The course therefore treats modern AI as the repeated interaction of:

$$ \boxed{ \text{probabilistic reasoning} + \text{learned representations} + \text{optimization} + \text{decision-making} } $$

rather than as a memorized sequence of architectures.

The course-wide rules for validation, chronology, causal claims, and deployment evidence are collected on the course page. Here, the important connection is that what counts as evidence depends on the question being asked.

Interactive browser lab

Use the orientation lab to classify familiar econometric and statistical objects using the five-question translation rule:

  1. the familiar statistical object;
  2. what becomes learnable;
  3. what objective is optimized;
  4. what generates the evidence;
  5. where the analogy breaks.

Some objects belong mainly to one mathematical spine. Others are cross-cutting methods and should not be forced into a single category.

The goal is not to memorize labels. First make your own prediction about the five answers, then reveal the worked translation. Compare a close mathematical relationship with a structural analogy and explain why the second requires more caution.

Python and R lab

Run the offline translation examples, then extend their small reference table with one concept from your own econometric background. The starting examples include:

  • OLS;
  • logistic regression;
  • Bayesian networks;
  • hidden-state models;
  • neural networks;
  • transformers;
  • MDPs;
  • PPO.

For each object, record:

  • what is modeled;
  • what is learned;
  • what objective is optimized;
  • where the data come from;
  • which interpretation should not automatically be carried over from the econometric analogue.

Classify the additional bridge as primarily:

  • probabilistic / inferential;
  • function approximation / representation learning;
  • sequential decision-making;
  • or cross-cutting.

Practice:

  1. Explain why a Bayesian network and a neural network solve different problems even though both can contain many parameters.
  2. Place MCMC, a transformer, an MDP, and PPO on the course map. Explain why some of them touch more than one mathematical spine.
  3. Give one example where prediction is useful even though causal identification is unavailable.
  4. Give one example where a deployed policy changes the future data distribution.
  5. Choose a familiar econometric model. Identify which parts are specified by the researcher and which are estimated. Then explain how this division changes in a neural network.
  6. Explain why the statement “an RNN is an AR model” is too strong even though AR models provide a useful bridge for understanding recurrent state.
  7. Explain why a machine-learning researcher may care about GPU memory or inference latency even when two models have similar predictive accuracy.
  8. Give one example of how self-supervision can create training targets from raw digital data without requiring a human-provided label for every observation.

Learner output

Create an econometrics-to-AI translation map that:

  1. identifies at least six familiar econometric or statistical concepts;
  2. connects each to a modern AI concept;
  3. distinguishes direct mathematical relationships from loose analogies;
  4. identifies what becomes newly learnable;
  5. separates causal, predictive, and sequential-decision questions;
  6. explains where computational constraints and self-supervised data change the classical workflow.

Conclude by explaining how reinforcement learning and LLM post-training connect probabilistic structure, learned representations, optimization, and decision-making.

Sources and further reading

Sources informing this orientation

  • Stock, James H., and Mark W. Watson. Introduction to Econometrics, 4th edition. The introductory econometrics sequence and Chapter 14 (Big Data and Machine Learning) provide the starting comparison. Author-maintained student resources.
  • Wooldridge, Jeffrey M. Introductory Econometrics: A Modern Approach, 7th edition. The regression, inference, time-series, panel, IV, and limited dependent-variable chapters illustrate a second common preparation route. Publisher’s contents.
  • Russell, Stuart, and Peter Norvig. Artificial Intelligence: A Modern Approach, 4th edition. See especially chapters 12–14 (uncertainty), 16–17 (decisions), 19–22 (learning and RL), and 23–24 (NLP). Book website and contents.

Optional further reading

You do not need to read all of these before beginning Module 1. Use the econometrics references to recognize familiar material and the AI chapters to locate the newer objects on the map.