Discuss how machine-learning techniques differ from classical econometrics in both philosophy and practice.
The idea
Classical econometrics starts from a theory, specifies a model such as a linear regression, and uses a modest number of variables to estimate parameters and test hypotheses about them. Success is judged by interpretable coefficients, standard errors and statistical significance, and by unbiased estimation. Machine learning starts from the data. It lets flexible algorithms find the structure, often with far more features than the researcher could specify by hand, and it is judged by out-of-sample predictive accuracy rather than by t-statistics. ML therefore accepts some bias, uses regularisation and holdout samples, and often sacrifices interpretability for forecasting power. Econometrics answers why and what if. ML answers what will happen. The exam point is that in-sample fit proves little for ML, and out-of-sample performance is the judge.
A practice question
Which statement best describes the philosophical difference between classical econometrics and machine learning?
- Econometrics begins with a theory-based model and tests hypotheses about its parameters, while machine learning lets the data determine structure with the goal of prediction. - correct
- Econometrics is purely data-driven and atheoretical, while machine learning requires a specified economic theory before estimation.Why not: Reverses the two approaches: it is econometrics that starts from theory.
- Both approaches test hypotheses about individual coefficients; they differ only in the software used.Why not: Ignores that ML generally targets prediction rather than parameter inference.
- Machine learning aims at unbiased estimation of causal parameters, while econometrics aims at forecasting accuracy only.Why not: Confuses the goals: causal inference and unbiasedness are the econometric focus.
Econometrics is theory-led and focused on inference about parameters; machine learning is data-led and focused on predictive performance on unseen data.
A harder one
A risk analyst has three tasks. Task 1: test whether a rise in the policy rate lowers bank loan growth, controlling for GDP growth. Task 2: forecast next-month default for 40,000 borrowers using 600 behavioural features. Task 3: estimate the CAPM beta of one stock. Which assignment of method is most appropriate?
- Task 1 suits machine learning; Tasks 2 and 3 suit classical econometrics.Why not: Uses ML for a hypothesis test about a single parameter, where interpretable inference is needed.
- All three tasks suit machine learning, because it always outperforms regression.Why not: Overlooks that ML does not provide the inference and parsimony these theory-based tasks require.
- Tasks 1 and 3 suit classical econometrics; Task 2 suits machine learning. - correct
- Task 2 suits classical econometrics, because a large number of features requires hypothesis testing.Why not: Misjudges 600 features: a hand-specified regression cannot cope well, while regularised ML can.
Tasks 1 and 3 ask about a specific parameter with theory behind it, which is econometric inference. Task 2 is a prediction problem with many features, the natural setting for ML.
The hardest tier
An analyst has 800 monthly observations of a fund's return and 1,200 candidate explanatory variables. Why can a machine-learning method with regularisation still be used here while ordinary least squares cannot?
- OLS has no unique solution when the number of variables exceeds the number of observations, whereas a penalty such as LASSO shrinks or drops coefficients and yields a usable model. - correct
- OLS requires normally distributed returns, which a penalty removes.Why not: OLS can be computed without normality; the obstacle is that there are more variables than observations.
- OLS can only handle up to about 100 variables regardless of sample size.Why not: Invents a fixed variable limit; the limit is the sample size relative to the variable count.
- Regularisation increases the sample size by generating synthetic observations.Why not: A penalty restricts coefficients; it does not create data.
With more features than observations the OLS normal equations are singular, so coefficients are not uniquely determined. Regularisation constrains the fit and selects or shrinks variables.
In the app: 6 practice questions on this outcome, spaced review that brings it back before you forget it.
Start free →More in Machine-Learning Methods
- Compare the two methods used to rescale variables when preparing data, and apply them.
- Explain what the training, validation and test sub-samples are, and the distinct purpose each serves.
- Examine how underfitting and overfitting differ and what follows from each, and outline possible remedies for both.
- Use principal components analysis to compress a set of features into fewer dimensions, and explain why that works.
- Outline how K-means assigns the observations in a sample to separate clusters.
- Describe how natural language processing works and what it is used for.
- Distinguish unsupervised learning models from supervised ones and from reinforcement learning models.
- Describe the operation of reinforcement learning and calculate the Q-values used when making decisions.