Topic · Machine Learning & AI
← All topicsML for Risk & Portfolio
26 articles · 4 checkpoints · 16 deeper reads · 6 reference notes
A standalone topic: it is on no roadmap, so read it on its own terms.
Every article, in reading order
plant a flag as you finish eachRead these first
Generative models can produce thousands of plausible-looking market scenarios beyond the handful of crises actually on record, useful for stress testing, but only if 'plausible-looking' is checked against the market's real tail behaviour rather than taken on faith.
Turning down risk before a crisis and turning it back up after sounds simple, but a regime classifier used for risk scaling is only useful if it identifies bad markets fast enough to act on, and the honest evaluation of 'fast enough' is much harder than backtesting the classifier's raw accuracy.
The traditional pipeline predicts returns, then hands those predictions to a separate optimiser to build a portfolio, end-to-end learning collapses the two steps, training the prediction model directly against the portfolio's actual performance instead of against prediction accuracy.
The plain sample covariance matrix is a noisy, overconfident estimate once you have more assets than history to measure them with, machine-learning approaches trade a little bias for a lot less noise, and often win on out-of-sample risk.
Then the rest