Blog – Joao Jeronimo

2026

LASSO Cannot Be Trusted for Variable Selection. Here Is the Geometry That Explains Why.

causal inference

machine learning

regression

L1 regularisation is excellent at predicting y-hat in sparse, high-dimensional problems, but it will not select the same predictors twice across partitions of the same dataset. This post reconstructs the Mullainathan and Spiess (2017) demonstration and makes the geometry of that instability visible in the hyperparameter space.

2025

How to Enhance Forecasts of the Scottish Economy with New Hire Wages and Machine Learning

forecasting

labour economics

machine learning

The Scottish Fiscal Commission forecasts private-sector real wages with an error correction model built on average wages. I show how two changes - separating new hire from incumbent wages, and swapping ARIMA for machine learning on short panels - measurably improve forecast accuracy, with simulated evidence in Python.

No matching items