Jonathanerrils/robust-ml-vs-stats-ghana

# Robust Machine Learning versus Statistical Models for Commodity and Macroeconomic Risk Signals Does machine learning actually beat classical econometrics for financial risk prediction, or does it just look that way because most comparisons never check if the "winning" model's edge is real or just noise? That's the question this project is built around, using Ghana's cocoa, gold, oil and cedi exposure as the test case. Full manuscript: `paper/paper.tex` (PDF), aimed at the *Journal of Risk and Financial Management*. Sister project: tail-dependence-ghana, which looks at how the same commodities co-move in the tails rather than how well they can be predicted. ## What we found Eleven models go head to head: four statistical (a base rate, GARCH-*t*, a vector autoregression, and a two-state Markov-switching model) against seven machine-learning classifiers (logistic, LASSO, elastic net, random forest, XGBoost, SVM, and a small neural net kept deliberately weak). Three daily risk signals for Ghana — cocoa VaR breaches, cedi depreciation episodes, high-volatility days — tested with nested time-series cross-validation on nine years of real data (2015–2026), with the actual 2024 cocoa supply shock held out as a genuine stress test rather than something simulated. The short version: **every serious model clearly beats a naive baseline, but once you correct for the 31 comparisons this study actually runs, nothing beats anything else** for cocoa risk or high-volatility days. Random forest does hold up against the weakest models for cedi depreciation — that one result survives even the strictest correction — but not against the other serious contenders like XGBoost, GARCH, VAR, or Markov-switching. | | | |---|---| | | | We read this as a finding about model risk, not a verdict on which paradigm wins — the same conclusion the companion tail-dependence project reached from a completely different angle. A recent, almost identically designed study on the CAD/USD exchange …