OLS with robust standard errors
Few controls that enter linearly
With fewer than 10 controls and no sign of nonlinearity, OLS already adjusts for the confounders; report the DML estimate as a robustness check …
Lecture 02: OLS inference →Choose a dataset
Nine real datasets: five prediction problems and four causal questions. Each one leads to a different method.
Upload your data
First choose the aim. Then upload a CSV with one row per observation: the outcome in the first column, then the other variables.
Drop a CSV file here, or
Your file is analysed on your own device and is not uploaded anywhere. Numeric columns only (code categories as 0/1 dummies). A column named year, quarter, month or date marks the data as a time series; for forecasting, enter the predictors already lagged (for example infl_l1). Files above 4,000 rows are subsampled for the forest checks. Format examples: econmath.csv (prediction), k401_irm.csv (causal).
Answer from your own results
Use this if you have already run the checks. If you have not, choose “Use my own data” and they are run and explained for you.
Method guides
Nine routes to a finished paper
Each guide takes one real dataset through tuning, validation, diagnostics and writing, with Python code, real output and sample text for the paper.
Unobserved confounders
Selection on unobservables, no instrument
Machine learning cannot fix confounding by variables you do not observe. Use a research design: difference-in-differences, synthetic control or regression discontinuity …
Open the Causal Inference Guide →How the tool decides. Prediction. (A1) A date column with at least 50 periods marks a time series: each series is tested with ADF and KPSS, I(1) series are differenced, and validation uses rolling-origin splits that never shuffle time. (A2) With fewer than 100 observations only regularised linear models are used. (A3) Otherwise a random forest is compared with LASSO by five-fold cross-validated RMSE; if the forest is at least 5% better, the data are treated as nonlinear. (A4) Linear: if any two predictors are correlated above 0.8, the elastic net; otherwise LASSO if it keeps fewer than half of the predictors at the one-standard-error penalty, and ridge if it keeps more. (A5) Nonlinear: a random forest below 1000 observations, XGBoost (gradient boosting) above. Causal effect. (B1) If confounders are unobserved and there is no instrument, machine learning cannot help. (B2) With fewer than 10 controls that enter linearly (the forest does not beat LASSO by 5% for the outcome or the treatment), OLS is enough. (B3) An instrument leads to DML-IV (the first-stage F should be at least 10); a binary treatment to the interactive model, after checking that the propensity scores lie in [0.05, 0.95]; a continuous treatment to the partially linear model. (B4) DML is estimated with LASSO and random-forest learners, five-fold cross-fitting; if the two differ by more than one standard error, the forest is reported. (B5) With at least 2000 observations and a question about who is affected most, a causal forest. These cut-offs are common rules of thumb, like T < 15 in the Panel Data Guide; report the choice and its robustness.