CAS BMAI FS26 · Week 1 · coding exercise · un games

The bait, the crash, the patch, the reveal, the trust.

You work for the UN. Your brief: predict GDP per capita. So you pull everything you can lay hands on — rule of law, life expectancy, but also UNESCO sites, vowel counts, the Scrabble score of each country’s name. 493 columns in total. For this tutorial we’ve quietly tagged every column as causal, spurious, or incidental. In real life you wouldn’t know. Five chapters. One investigation. The data fights back.

causal — a plausible economic mechanism spurious — correlation without cause incidental — real statistic, not a root cause

Five chapters

What predicts GDP?

You start where any honest analyst starts: you plot. You scatter. You rank every feature by its correlation with GDP. Rule of law scores high. So does the Scrabble value of the country’s name. You can’t tell which is which by looking — and that’s the bait.

02 · THE CRASH

The hockey stick of doom

Emboldened, you hand all 493 features to OLS. Train R² climbs toward 1: you’re a hero. Then you check the test set and the number is negative. Slide the feature count across the cliff at p = n and watch the wheels come off in real time.

03 · THE PATCH

Regularisation rescues prediction

Ridge, Lasso, Elastic Net. One extra knob — α — drags the test score out of the crater. Prediction is saved. Then you look at which features Lasso kept, and the relief doesn’t last.

04 · THE REVEAL

The model kept the wrong features

Lasso’s top 20 includes features the codebook labels spurious. Random Forest’s top 20 is a different set. Both models predict well. They can’t both be right about what matters — and you have no way to decide.

05 · THE TRUST

One split lied; what does cross-validation say?

Every number so far came from one random train/test split. Resample a hundred times and a wide distribution appears where you thought you had a single answer. You end the story chastened. Also: better at your job.

Data source

Everything on this site is computed live in your browser from gdp_spurious_regression_dataset.csv (254 rows × 495 columns) and codebook.csv, both lifted directly from the eth-bmai-fs26/coding-exercises Colab notebooks. No pre-baked charts, no cherry-picked numbers. Pick a feature, toggle a scale, and watch the story rearrange itself.