Methodological Insights

Technical breakdowns of model failure, stochastic stress-testing, and the application of statistical physics to industrial data problems.

Case Study Feb 2026

Why 90% Accuracy is Often a Symptom of Model Failure

High AUC and accuracy metrics can mask structural flaws like temporal myopia or severe class imbalance. This breakdown explores how to use non-Gaussian null hypotheses to identify when a model is simply "memorising" noise rather than extracting signal.

Coming soon on Medium
Methodology Drafting

Causal Discovery in Multi-Scale Time Series

Standard transfer entropy often fails on complex industrial sensor data. By using symbolic calculus and ordinal patterns, we can assess the directionality of information flow across multiple time scales simultaneously.

Expected publication: March 2026
Research Scheduled

The Physics of Pathological Residuals

When residuals exhibit power laws or Lévy flights, Gaussian-based stress testing produces systematically wrong answers. A look into why "fat tails" are the primary source of predictive failure in high-stakes environments.