The real difference: what question are you asking?
A common question is whether econometrics or machine learning is becoming obsolete. Eight years on, the answer is clear: neither is. The useful distinction is not between old and new tools but between two kinds of question. A prediction question asks what the value of an outcome will be for a new case. An estimation question asks how an outcome would change if one variable changed, holding others fixed. Mullainathan and Spiess (2017) put this as the difference between predicting y and estimating beta.
Econometrics developed to estimate beta credibly from observational data. Its central concern is identification: the conditions under which a coefficient can be read as a causal effect rather than a correlation (Angrist & Pischke, 2009; Wooldridge, 2010). Machine learning developed to predict y accurately for cases the model has not seen. Its central concern is generalisation: avoiding a model that fits the training data closely but fails on new data. Methods such as random forests trade interpretability of individual coefficients for predictive accuracy (Breiman, 2001).

How the two traditions compare
| Feature | Econometrics | Machine learning |
|---|---|---|
| Main target | A parameter, often a causal effect | Out-of-sample prediction of an outcome |
| Role of theory | Guides specification and identification | Guides features and evaluation; model form is learned from data |
| Typical models | Linear and nonlinear regression, instrumental variables, panel and time-series models | Regularised regression, tree ensembles, boosting, neural networks |
| How success is judged | Unbiased or consistent estimates with valid standard errors | Error on held-out data, such as mean squared error or AUC |
| Main risk | Omitted variables, reverse causality, weak instruments | Overfitting, leakage between training and test data |
| Interpretation | Coefficients have a stated meaning | Parts of the model may be hard to interpret |
The table simplifies. Econometricians forecast, and time-series econometrics is a forecasting discipline with its own evaluation methods (Hyndman & Athanasopoulos, 2021). Machine learning researchers increasingly care about interpretation. But the difference in targets explains most disagreements between the two communities.
Why prediction alone cannot answer causal questions
A model that predicts house prices well may use the number of bathrooms, the floor area and the postcode. It does not follow that adding a bathroom raises the price by the amount the model implies. Predictors are often correlated, so many different models predict equally well while assigning very different weights to each variable (Mullainathan & Spiess, 2017). That is harmless for prediction and fatal for estimation.
The same caution applies to the explanation tools now common in machine learning. Shapley-value methods describe how much each feature contributed to a model's prediction (Lundberg & Lee, 2017), but those contributions describe the model, not the economy. For high-stakes decisions, some researchers argue for models that are interpretable by design rather than black boxes explained after the fact (Rudin, 2019). Causal claims still need a research design that rules out alternative explanations (Antonakis et al., 2010).
Econometrics has its own history of fragile results. Leamer (1983) showed how conclusions could change with small, undisclosed choices of specification, which helped push the field towards credible research designs and robustness checks.
Combining both: machine learning inside econometrics
Since 2017 the most productive work has used machine learning for the parts of a model where prediction is what is needed. The best-known example is double/debiased machine learning (Chernozhukov et al., 2018). Suppose you want the effect of a treatment D on an outcome Y, controlling for many variables X. The method uses machine learning to predict Y from X and D from X, takes the residuals of each, and regresses one set of residuals on the other. Two features make it work: the estimating equation is constructed so that small errors in the machine learning predictions have little effect on the treatment estimate (Neyman orthogonality), and the predictions for each observation are made on data that did not include it (cross-fitting). The result is an estimate of the treatment effect with valid confidence intervals, provided the usual identifying assumption holds: that X captures all confounders.
Machine learning can also help econometric work in other ways: selecting among many candidate controls, building a first stage for instrumental variables, estimating how effects vary across groups, and creating new variables from text, images or satellite data (Mullainathan & Spiess, 2017). In the other direction, econometric thinking about identification and inference makes machine learning research more credible. Packages implementing double machine learning are available in R and Python.
Worked example (illustrative)
A doctoral researcher in development economics has administrative data on 20,000 small firms, including whether each received a government training grant, its revenue growth and 150 firm characteristics. She has two questions.
| Question | Type | Suitable approach | What to report |
|---|---|---|---|
| Which firms are most likely to fail within two years? | Prediction | Gradient boosting or regularised logistic regression with cross-validation | Held-out performance, calibration, comparison with a simple baseline |
| Did the grant raise revenue growth? | Causal estimation | Double machine learning with the 150 characteristics as controls; or a regression discontinuity if eligibility used a score cut-off | Effect estimate and confidence interval, identifying assumption, sensitivity checks |
For the first question she reports which model predicts best and how much better it is than a simple rule. She does not interpret its feature importances as causes. For the second she states the assumption that the controls capture everything that drove both grant receipt and growth, and she discusses what could violate it. If the grant was allocated by a score with a cut-off, a regression discontinuity design gives a more credible estimate for firms near the cut-off, and she can use the double machine learning estimate as a comparison.
Evaluating prediction models honestly
Prediction claims fail in a predictable way. Kapoor and Narayanan (2023) documented data leakage, in which information from the test data reaches the model during training, across many fields that use machine learning, producing inflated performance claims. Economic and financial data are especially exposed because they are ordered in time. Random cross-validation can leak future information into the past, so time-series work needs evaluation designs that respect order (Bergmeir & Benítez, 2012; Hyndman & Athanasopoulos, 2021).
Always compare a machine learning model with a sensible econometric benchmark. In forecasting South African inflation, for example, Botha et al. (2023) compared machine learning methods applied to a large set of predictors with conventional benchmark models, which is the comparison an examiner will expect.
Choosing for your thesis, and what to report
Start from the sentence that states your research question, not from the method. If the question contains words such as "effect of", "impact of" or "does X raise Y", you are estimating a parameter, and the first chapter of your methodology should be the identification strategy: the source of variation in X that you treat as good as random, and the assumptions that make it so. Machine learning may then help with controls, heterogeneity or measurement, but it does not replace that argument. If the question asks "which cases", "how many" or "what will happen next", you are predicting, and the methodology should centre on the data split, the benchmark and the loss function that reflects the real cost of errors.
Many economics and finance theses contain both kinds of question, as in the worked example. Keep them in separate sections with separate evaluation criteria, so that an examiner never has to wonder whether a coefficient is being read as a cause or a forecast is being judged by its standard errors. Whatever you choose, report the following.
- For estimation: the estimand, the identifying assumption, the estimator, standard errors and how they were clustered, robustness checks and the main threats that remain.
- For prediction: the outcome and prediction horizon, how data were split, the benchmark, the tuning procedure, held-out performance and any checks for leakage.
- For combined methods: which parts of the model use machine learning, which learners and settings were used, how cross-fitting was done, and how sensitive the estimate is to the choice of learner.
- For both: software, versions, random seeds and code, so that the results can be reproduced.
Related pages: the steps of an empirical econometric study, how to choose a machine learning algorithm, what big data and machine learning can and cannot show in empirical finance and reporting regression analyses.
Common mistakes
- Reading coefficients or feature importances from a prediction model as causal effects.
- Using random cross-validation on time-ordered data, which leaks future information.
- Reporting only the best of many models without saying how many were tried.
- Comparing a tuned machine learning model with an untuned or weak benchmark.
- Assuming that using machine learning for controls removes the need for an identification strategy.
- Omitting the code and settings needed for another researcher to reproduce the results.
Key takeaways
- Econometrics and machine learning answer different questions: estimating parameters versus predicting outcomes (Mullainathan & Spiess, 2017).
- Good prediction does not identify causal effects; that still needs a credible research design (Angrist & Pischke, 2009; Antonakis et al., 2010).
- Double/debiased machine learning uses prediction for nuisance parts of a model to support valid causal estimates (Chernozhukov et al., 2018).
- Prediction claims must be tested on held-out data without leakage, respecting time order (Kapoor & Narayanan, 2023; Bergmeir & Benítez, 2012).
- Explanation tools describe the model, not the world (Lundberg & Lee, 2017; Rudin, 2019).
