Machine Learning Methods Economists Should Know About

• Machine-learning methods are increasingly important for economics, particularly in “big data” settings, but economists should preserve the traditional strengths of applied econometrics.These settings may involve many units, many features per unit, or data structures beyond a simple cross-section.Source: §1 · pp. 2–3
• Economics and machine learning emphasize different formal goals: econometrics often studies consistency, Normality, and efficiency, whereas machine learning often provides setting-specific error-rate guarantees.The paper notes that no formal results establish uniform superiority of deep learning or neural networks over regression trees or random forests.Source: §1 · p. 3
• Out-of-sample predictive power can be valuable even without conventional confidence intervals, although valid uncertainty quantification remains important for parameters such as average treatment effects.Researchers should clearly articulate their analysis goals and why particular algorithmic or estimator properties matter.Source: §1 · pp. 3–4
• Off-the-shelf machine-learning algorithms can work well, but economists often need careful tuning and adaptation to exploit causal, endogenous, panel, choice-model, or theory-based structure.Adaptations can change optimization criteria, control for confounders, target treatment-effect heterogeneity, and use sample splitting or orthogonalization.Source: §1 · p. 4
• The review proposes a subjective, introductory list of machine-learning tools for empirical economists and core econometrics courses, beginning with supervised learning for regression and classification and unsupervised learning.The list is intended to convey main ideas and insights rather than provide comprehensive coverage, and it will evolve with the literature.Source: §1 · pp. 4–5
• Bandit approaches are presented as transforming effective experimentation, especially in online settings.The paper discusses machine-learning approaches to experimental design before highlighting bandit methods.Source: §1 · p. 5
• The paper discusses matrix completion for causal panel-data models and consumer choice among discrete products.Source: §1 · p. 5
• The paper concludes its methodological coverage with approaches for analyzing text data.Source: §1 · p. 5
• The discussion is positioned alongside reviews covering important ML methods, supervised learning for regression, causal applications, and text analysis in economics.Varian [2014], Mullainathan and Spiess [2017], Athey [2017] and Athey et al. [2017c], and Gentzkow et al. [2017] are identified as related contributions.Source: §1 · p. 5
• It also points readers to accessible statistics and computer-science textbooks, including Efron and Hastie [2016], Hastie et al. [2009], Burkov [2019], Alpaydin [2009], and Knox [2018].The cited texts differ in accessibility and disciplinary perspective.Source: §1 · p. 5

2 Econometrics and Machine Learning: Goals, Methods, and Settings

• The paper frames the contrast as a difference in goals and concerns that affects how researchers choose among specific econometric and machine-learning methods.Source: §2 · p. 5

2.1 Goals

• Traditional econometrics specifies an estimand as a functional of the joint data distribution, often a parameter in a statistical model [2010], [2008], [2000].Researchers estimate target and nuisance parameters from a random sample using objectives such as least squares or likelihood.Source: §2.1 · p. 6
• Econometric analysis focuses on estimator quality, traditionally measured by large-sample efficiency, and commonly reports confidence intervals, point estimates, and standard errors.Source: §2.1 · p. 6
• Machine learning typically develops algorithms to predict variables from others or classify units using limited information, such as pixel values for handwritten digits [2008].Source: §2.1 · p. 6
• For prediction, the target is the outcome of a new unit given its feature vector, with estimator quality evaluated through the loss associated with that decision.Source: §2.1 · pp. 6–7
• When feature dimension exceeds two, decision theory shows that estimators can achieve lower expected squared error than least squares, which is not admissible.Source: §2.1 · p. 7

2.2 Terminology

• ML replaces familiar econometric labels with terms such as training sample, training, features, and weights.Regression parameters are called weights, while regressors, covariates, or predictors are called features.Source: §2.2 · p. 7
• Supervised learning observes predictors and outcomes, whereas unsupervised learning observes predictors alone and seeks clusters or their joint distribution.Problems with unordered discrete responses are generally called classification problems.Source: §2.2 · p. 7

2.3 Validation and Cross-validation

• Model validation targets predictive power for new units rather than estimation of a particular structural or causal parameter.The relevant setting is prediction for a new unit randomly drawn from the same population as the sample.Source: §2.3 · pp. 7–8
• Out-of-sample cross-validation compares predictive fit without relying on in-sample goodness-of-fit, helping decide whether to retain regressors such as Xi.If β is zero, or sufficiently close to zero, omitting Xi can improve prediction.Source: §2.3 · pp. 7–8

2.4 Over-fitting, Regularization, and Tuning Parameters

• ML emphasizes avoiding over-fitting by selecting flexible models that fit well without compromising out-of-sample prediction.Methods are compared on specific data sets to assess what works well, with less emphasis on asymptotic superiority results.Source: §2.4 · p. 8
• Regularization adds a complexity penalty to an objective function, with model capacity measurable by parameters or the Vapnik–Chervonenkis (VC) dimension.The penalty can modify objectives such as residual sum of squares or the log likelihood.Source: §2.4 · p. 8
• Modern regularization determines the penalty from out-of-sample predictive performance, unlike approaches based on subjectively chosen priors.In an ML approach, λ is selected through out-of-sample cross-validation; this is closer to an Empirical Bayes approach where data estimate the prior distribution (Morris [1983]).Source: §2.4 · pp. 8–9

• ML applications often involve many features, but researchers may not know which can be removed without substantially reducing predictive power.Source: §2.5 · p. 9
• The sparsity principle supports ℓ1-based recovery when the true signal is sparse, but performance deteriorates when the underlying truth is not sparse in the chosen bases.If the sparsity assumption is wrong, no method can perform well relative to the Bayes error (Hastie et al. [2015]).Source: §2.5 · pp. 9–10
• Approximate sparsity can suffice when most explanatory variables have limited power and only a few features are substantially important (Belloni et al. [2014]).Source: §2.5 · p. 10
• Allowing data-dependent variable selection improves on traditional hand-limited specification, although approximate sparsity remains a strong assumption and post-selection inference is challenging.Source: §2.5 · p. 10

Computational Issues and Scalability

• ML favors computationally scalable methods for large datasets, even when alternatives have more attractive statistical-efficiency properties.This emphasis distinguishes ML from traditional statistics and econometrics.Source: §2.6 · p. 10
• LASSO has effective methods for calculating estimates with millions of regressors, whereas best subset selection is NP-hard and was thought feasible only with regressors in the 30s until recently.Source: §2.6 · pp. 10–11
• In low-signal-to-noise-ratio settings common in social science, LASSO may perform better, although the relative merits of LASSO and subset selection remain unresolved.Best subset selection is also feasible for many social-science problem sizes, so computational considerations may be less important than substantive ones.Source: §2.6 · p. 11
• SGD reduces optimization cost by taking many noisy, small steps using unbiased minibatch gradient estimates instead of accurately computing each full gradient.With an appropriately decreasing learning rate, SGD converges almost surely to a global minimum for convex or pseudoconvex objectives and to a local minimum otherwise.Source: §2.6 · p. 11
• The SGD approach can be pushed further when ∇Qi(θ) is itself an expectation, by considering Monte Carlo integration to evaluate ∇Qi.Source: §2.6 · p. 12

Empirical example

• Ensemble weights, or votes, are chosen by optimizing out-of-sample performance, and Netflix Prize leaders combined many often-averaged models (Bennett et al. [2007]; Bell and Koren [2007]).The competition illustrates the predictive advantage of combining potentially quite different models.Source: §2.7 · p. 12

Relation to econometrics

• Unlike Bayesian and mixture-model averaging, modern ensembles can average substantially different model specifications rather than similar models differing mainly in parameter values.Traditional approaches implicitly average over posterior parameter values or combine parameter values within mixture models.Source: §2.7 · p. 12

Core motivation

• Combining diverse models can produce strictly better out-of-sample predictions than any single model because their strengths differ across irrelevant features, nonlinearities, and interactions.Random forests, neural nets, and LASSO address these features differently, motivating model averaging.Source: §2.7 · p. 13

Panel data application

• In panel data, Athey et al. [2019] combine synthetic-control and matrix-completion methods and find that the combinations outperform the individual methods.This extends ensemble methods beyond standard predictive-model combinations.Source: §2.7 · p. 13

Inference

• Machine learning has prioritized out-of-sample performance over the ability to conduct inference and construct valid confidence intervals, as Efron and Hastie [2016] note.The passage characterizes predictive algorithmic development as outpacing its inferential justification.Source: §2.8 · p. 13
• Recent methods enable inference for low-dimensional functionals in specific settings, including random forests and neural networks, but many methods still lack asymptotically valid confidence intervals.Examples include Wager and Athey [2017] for random forests and Farrell et al. [2018] for neural networks.Source: §2.8 · p. 13
• For some decision problems, prediction may be primary and inference secondary, so the importance of confidence intervals depends on the decision context.The passage presents this as a question about whether traditional econometric emphasis on inference is universally appropriate.Source: §2.8 · p. 13
• Inference requirements can reduce predictive performance: in kernel regression, valid intervals require a smaller-than-optimal bandwidth that removes asymptotic bias while increasing variance.The bandwidth minimizing expected squared error leaves asymptotic bias that invalidates standard confidence intervals; undersmoothing fixes bias at the cost of higher variance.Source: §2.8 · p. 13

3 Supervised Learning for Regression Problems

• The section contrasts machine learning regression with econometrics by emphasizing prediction accuracy over coefficient interpretation, allowing many covariates, nonparametric conditional distributions, and nonmonotone relationships.Higher-order interactions may matter substantially, although economic applications often make monotonicity plausible for important predictors such as education and prior earnings.Source: §3 · p. 14
• Random forests can outperform kernel regression in sparse, high-dimensional settings by ignoring irrelevant features and detecting severe nonlinearities and high-order interactions.Traditional kernel methods waste degrees of freedom on irrelevant covariates and are generally limited to low-dimensional applications.Source: §3 · pp. 14–15

3.1 Regularized Linear Regression: Lasso, Ridge, and Elastic Nets

• For penalized linear prediction, q = 1 corresponds to LASSO, q = 2 to ridge regression, and q → 0 to best subset regression, which penalizes nonzero covariates.Source: §3.1 · p. 16
• Least squares can have poor predictive and sampling properties when K is large relative to N, while best subset regression is computationally hard when N and K are large.The text also notes that regularization can help even when K is modest, and that sparse variable selection is not always substantively important.Source: §3.1 · p. 16

3.2 Regression Trees and Forests

• Regression trees and random forests flexibly estimate regression functions for settings prioritizing out-of-sample prediction, with strong out-of-the-box performance and limited tuning.Trees recursively split observations on one covariate at a time at threshold c and estimate the regression function within each resulting subsample by its average outcome.Source: §3.2 · p. 17
• A regression tree recursively chooses covariate-threshold splits that minimize average squared error, then uses regularization or cross-validation to prevent overfitting from excessive splitting.A common implementation grows a deep tree and prunes it to a cross-validated depth.Source: §3.2 · pp. 17–18
• Because tree leaves are selected using the training data, their sample means can be too extreme for new test sets; sample splitting makes leaf means unbiased for the corresponding true leaf means.Trees are interpretable, but split variables should not be treated as definitive because correlated covariates can substitute for one another.Source: §3.2 · p. 18
• Random forests average randomized trees to smooth discontinuous predictions and achieve better predictive power than a single tree, with little tuning relative to deep neural networks.Trees use bootstrap samples or subsamples and random covariate subsets at each split; averaging creates a smoother, though still discontinuous, predictor.Source: §3.2 · pp. 19–20
• Random forests remain strong with many irrelevant features because their splits largely ignore covariates unrelated to the outcome, unlike kernel regression, whose performance degrades when irrelevant covariates are added.This makes forests particularly effective in sparse, high-dimensional settings.Source: §3.2 · p. 20
• Wager and Athey [2017] show that an honest subsampled forest yields an asymptotically normal estimate centered on the true regression function and supports variance estimates and centered confidence intervals.Each tree uses disjoint subsamples to define its structure and estimate leaf means, and this honesty is crucial for the asymptotic analysis.Source: §3.2 · p. 20
• Generalized random forests construct forest-based neighborhoods, estimate a GMM model at each test point, and deliver asymptotically normal parameter estimates under appropriate honest estimation.Nearby observations are weighted more heavily when they frequently occur in the same leaf as the test point.Source: §3.2 · pp. 21–22
• Local linear forests can match a simple polynomial function even in relatively small data sets, whereas traditional forests tend to have bias, particularly near boundaries.Figure 1 compares predictions from random forests and local linear forests on 600 test points.Source: §3.2 · pp. 22–23, Figure 1 · p. 23

3.3 Deep Learning and Neural Nets

• For this setting, methods require substantial tuning to work well; shallow models overfit at around 20 million parameters, while deep models can benefit from having over 60 million.Source: §3.3 · pp. 23, 25
• Neural networks provide flexible regression models whose nonlinear hidden-node transformations can approximate arbitrarily well a rich set of smooth functions, with depth substantially increasing practical flexibility.Their predictions can be attractive even when parameter estimates are poorly behaved, so prediction properties—not parameter properties—are the focus.Source: §3.3 · pp. 23–24
• The deep-learning subsection references a computationally efficient method described in Section 2.6, but the supplied passage does not identify the method’s target or details.The excerpt is incomplete and ends after stating that the method was described as computationally efficient.Source: §3.3 · p. 26

3.4 Boosting

• Boosting repeatedly fits a simple base learner to updated residuals, combining the successive predictions into a stronger predictor.The example uses a two-split regression tree and can repeat the procedure many times; other easily repeated base learners can also be used.Source: §3.4 · p. 26
• With regression trees having L splits, boosting can approximate regression functions formed by sums of functions involving L original features at a time.For L = 1 this includes additive functions, while L = 2 includes additive functions with general second-order effects.Source: §3.4 · p. 26

4 Supervised Learning for Classification Problems

• Classification uses features X_i to assign new units to one of several discrete labels, paralleling discrete-choice analysis but not necessarily requiring estimated label probabilities.Classification methods often target a single predicted label, and applications may assume covariates permit nearly certain assignment rather than accepting high unavoidable error rates.Source: §4 · p. 27
• In digit recognition, SVMs greatly outperformed other methods in the 1990s, while deep convolutional neural networks later reduced error rates further.The task classifies pixel-coded images into the ten digits from 0 to 9.Source: §4 · p. 27

4.1 Classification Trees and Forests

• Classification trees choose covariate thresholds by minimizing impurity, commonly using Gini impurity, penalizing the number of leaves, and extending naturally to random forests.Gini impurity is minimized by pure leaves and maximized when the M label shares are all 1/M.Source: §4.1 · pp. 27–28

4.2 Support Vector Machines and Kernels

• SVMs select a separating hyperplane by maximizing its margin, use bounded constraints when perfect separation fails, and can yield more robust estimates than logistic regression.In nonlinear settings, kernels avoid explicitly specifying transformed features; kernel smoothing parameters are typically selected by cross-validation.Source: §4.2 · p. 28

5 Unsupervised Learning

• Unsupervised learning analyzes observations without labels, either partitioning samples into clusters or estimating the joint distribution of covariates.Source: §5 · p. 30

5.1 K-means Clustering

• K-means clustering partitions feature space into K subspaces, assigning units to nearby centroids and updating centroids as cluster averages (Hartigan and Wong [1979], Alpaydin [2009]).Clusters can create new features, support parsimonious within-cluster models, or organize units by potentially different treatment exposures; choosing K lacks direct cross-validation and often relies on substantive grounds.Source: §5.1 · p. 30

5.2 Generative Adverserial Networks

• Generative adversarial networks (GANs) aim to generate data that look like an observed sample and may also provide an effective way to choose simulation designs.For a successful algorithm, a classifier should not be able to tell generated data from original data.Source: §5.2 · p. 31

6.1 Average Treatment Effects

• Under unconfoundedness, estimating average treatment effects is a canonical problem: potential outcomes are independent of treatment conditional on covariates, and the ATE is identified as a functional of their joint distribution.Source: §6.1 · p. 32
• Belloni et al. [2014] show that selecting covariates predictive of both outcomes and treatment substantially improves the corresponding estimator for τ.Outcome-prediction-focused selection can omit treatment-predictive covariates and create substantial bias, even when their outcome correlation is modest.Source: §6.1 · p. 33
• Directly optimizing balancing weights enables efficient ATE estimation when the propensity score is too complex to estimate well and is effective with many weak confounders.Balancing weights target equal covariate means in treatment and control groups, avoiding instability from dividing by an estimated propensity score.Source: §6.1 · p. 33

6.2 Orthogonalization and Cross-Fitting

• Nuisance components can converge at rates close to N^-1/4 while the ATE estimator remains efficient, because orthogonalization makes estimation robust to nuisance-estimation errors.These guarantees become especially important with many covariates, where accurate nuisance-parameter estimation is harder.Source: §6.2 · pp. 34–35
• Cross-fitting and related sample-splitting methods estimate each observation’s nuisance parameters without using its outcome data, while random forests can use out-of-bag predictions.These methods address overfitting and the influence of an observation on flexible-model predictions.Source: §6.2 · pp. 34–35

6.3 Heterogenous Treatment Effects

• The section targets low-dimensional heterogeneity summaries, flexible estimates of τ(x), and treatment policies that assign treatment or control using covariates.The CATE is τ(x) = E[τ_i|X_i = x] and is identified under unconfoundedness, but individual treatment effects are never observed.Source: §6.3 · p. 35
• Causal model selection is harder than predictive modeling because observational-study CATE MSE is infeasible and lacks a simple model-free unbiased estimate.Regularization and cross-validation therefore require criteria adapted to causal parameters; Athey and Imbens [2016] propose relative-MSE estimators using transformed outcomes, though propensity-score estimation introduces modeling choices.Source: §6.3 · p. 36
• Athey and Imbens [2016] develop causal trees that split covariate space for treatment-effect heterogeneity, use sample splitting, and provide reliable confidence intervals.Their splits optimize heterogeneity rather than prediction, while leaf effects in randomized experiments are sample average treatment effects.Source: §6.3 · pp. 36–37
• Causal trees are interpretable but their partitions can be unstable, whereas smooth individual-level estimates may be preferable when nearby covariate values should receive similar treatment decisions.Traditional kernels and matching perform poorly with many covariates, motivating forests and other flexible approaches.Source: §6.3 · p. 37
• Wager and Athey [2017] introduce causal forests, averaging subsampled causal trees to select covariate dimensions and establish asymptotic normality with confidence intervals for honest trees.Forests improve flexibility in many covariates but produce complex CATE functions that are difficult to describe.Source: §6.3 · p. 37
• Other approaches address structured heterogeneity, including methods for testing ranked-group effects, heterogeneous instrumental-variables parameters, local-linear forests, neural nets, LASSO, meta-learners, and BART.These alternatives differ in available theoretical guarantees; for example, the cited neural-net estimator lacks distributional theory.Source: §6.3 · pp. 37–38
• Optimal policies assign treatment using estimated CATEs, but the resulting rules may be complex and nonsmooth; adding treatment or control costs can make this optimal under some conditions.Hirano and Porter [2009] establish optimality under some conditions.Source: §6.3 · p. 38
• Kitagawa and Tetenov [2015] estimate limited-complexity policies by minimizing regret, while accommodating observational data with known propensity scores.Athey and Wager [2017] additionally accommodate constraints such as treatment budgets and propose an algorithm for estimating optimal policies.Source: §6.3 · p. 38
• Athey and Wager [2017] derive tighter performance bounds using semiparametric efficiency theory, narrowing the algorithms that might achieve the regret bound.Their analysis covers unknown propensity scores under confoundedness and extends to settings without unconfoundedness, including instrumental variables.Source: §6.3 · pp. 38–39
• The optimal-policy optimization can be reframed as a classification problem and solved with off-the-shelf classification tools.For unconfoundedness, the recommended procedure uses cross-fitting and orthogonalization.Source: §6.3 · p. 39

7.1 A/B Testing versus Multi-Armed Bandits

• Thompson sampling and UCB provide alternative adaptive rules: Thompson sampling is stochastic, whereas UCB assigns units to the arm with the highest confidence bound.UCB can be problematic in batched settings when two arms have similar upper confidence bounds, while Thompson sampling supports randomization inference.Source: §7 · p. 39, §7.1 · p. 41
• Multi-armed bandits adapt assignments using accumulated outcomes, reducing exposure to inferior treatments and eventually assigning nearly all new units to the optimal treatment.Thompson sampling assigns each treatment with probability equal to its estimated probability of being optimal.Source: §7.1 · p. 40

7.2 Contextual Bandits

• Contextual bandits exploit treatment-effect heterogeneity by assigning units based on observed features, but misspecified models can prevent learning the optimal policy in unobserved regions.Random forests can reduce functional-form assumptions, while adaptive assignments require bias control in reward-function estimation.Source: §7.2 · pp. 41–42
• Propensity weighting improves contextual-bandit performance on real-world datasets, while doubly robust estimation provides performance bounds.Dimakopoulou et al. [2018] studies these bounds and demonstrates the performance improvement.Source: §7.2 · p. 42
• Contextual bandit applications may require flexible outcome models, for which random forests can be a good alternative.Source: §7.2 · p. 42

8 Matrix Completion and Recommender Systems

• Matrix completion addresses longitudinal or panel-data settings by recovering missing entries in structured outcome matrices.The section presents this as a canonical panel-data problem alongside methods developed for recommender systems.Source: §8 · p. 43

8.1 The Netflix Problem

• The Netflix Prize accelerated matrix-completion methods, with competitive solutions relying heavily on model averaging, matrix factorization, and nearest-neighbor methods.The competition asked researchers to predict unobserved movie ratings from training data containing movie and individual characteristics, and offered a $1,000,000 prize (Bennett et al. [2007]).Source: §8.1 · p. 43
• Binary-treatment panel data can be represented as two incomplete potential-outcome matrices, making average-treatment-effect estimation a matrix-completion problem.Causal-effect estimation becomes imputing missing treated and control outcomes; machine-learning methods target settings with large N and T and many missing observations.Source: §8.1 · pp. 43–44

8.2 Matrix Completion Methods for Panel Data

• Matrix-completion literature uses low-rank representations for the expected-value matrix L, expresses L as USV⊤ through singular-value decomposition, and applies nuclear-norm regularization.Source: §8.2 · p. 44

8.4 Demand Estimation in Panel Data

• Consumer-demand panel data can be studied with matrix completion by learning from similar purchase patterns across consumers and products.Traditional demand models analyze discrete choices among a small number of products and often estimate cross-price elasticities for merger or price-change counterfactuals; the matrix-completion framing draws on cross-consumer and cross-product purchase similarities.Source: §8.4 · p. 46
• Models that combine consumer-choice structure with matrix completion tend to outperform models that omit problem structure, while exploiting cross-category and cross-product information in panel data.Economic functional forms can improve estimation efficiency when they approximate reality, whereas matrix completion methods use information that traditional economic models often leave unused.Source: §8.4 · p. 47
• Latent factorization represents consumer preferences and product characteristics with latent variables, but the nonlinear choice probabilities make computation challenging.Semenova et al. [2018] provide a complementary approach using observational high-dimensional product attributes such as text descriptions and images.Source: §8.4 · pp. 47–48
• Ruiz et al. [2017] combine stochastic gradient descent, variational inference, and approximations to scale latent-choice models to thousands of consumers, items, and shopping trips.Their model learns product substitution or complementarity from data rather than using predefined product categories.Source: §8.4 · p. 48
• Athey et al. [2017b] impose category information, one-product-per-category purchasing, and a nested-logit structure, while Wan et al. [2017] model category, item, and quantity choices sequentially.These approaches incorporate price variation and use latent factorization to represent correlated consumer choices within categories.Source: §8.4 · p. 48
• Economic utility maximization preserves analyses such as the impact of price changes on consumer welfare, which standard product-prediction applications do not target.Jacobs et al. [2014] instead predict which new products customers will buy and do not model price responses or substitution between similar products.Source: §8.4 · pp. 46–48

• Text can be represented by a document-by-word indicator matrix, but unigram representations lose word order, motivating bigrams and longer sequences.Source: §9 · pp. 48–49
• Unsupervised learning seeks a low-rank factor representation with latent document characteristics and topic weights, evaluated by predicting randomly held-out matrix entries.This formulation turns text analysis into a matrix-completion problem, allowing the matrix-completion methods described earlier to be applied.Source: §9 · p. 49
• Topic models, including Latent Dirichlet Allocation, estimate latent topics, topic-specific word distributions, and article-level topic weights through a generative model [2009].The supplied passage identifies Blei and Lafferty [2009] as a review of topic models.Source: §9 · p. 49
• More complex language models have emerged, and word embedding methods incorporate problem-specific structure in state-of-the-art machine-learning applications while attempting to capture latent semantic structure in language.Source: §9 · p. 49
• Supervised learning predicts labeled text characteristics such as review favorability, legislative polarization, or tweet sentiment, using regularized methods because T is usually much greater than N.Researchers may apply dimension reduction first or use supervised topic models that predict characteristics in unlabeled test documents.Source: §9 · p. 50

• Machine learning has much to offer empirical researchers in economics.Source: §10 · p. 50
• The review identifies ML methods useful for economists and important for core graduate econometrics sequences.Source: §10 · p. 50
• Familiarity with these methods enables more sophisticated empirical work and more effective communication with researchers in other fields.Source: §10 · p. 50