Machine Learning for Trading in the Age of AI Agents — Workshop Review
Gifted review ticket from Packt Publishing — honest, unpaid opinion below.
On Saturday, August 15, 2026, I spent three hours, effectively, in Stefan Jansen's (Packt) live workshop, "Machine Learning for Trading in the Age of AI Agents" — two 90-minute parts with a short break.
hort verdict: the substance was excellent; less of the three hours went into the pipeline itself than the "hands-on" billing suggested.
What impressed me
What impressed me: this is one of the more statistically honest ML-in-finance sessions I've sat through. Stefan built a monthly top-10 LightGBM strategy on free ETF data, walk-forward cross-validated — then did something rare for a public workshop: showed how the signal partly dissolves once you correct properly for autocorrelation (HAC), and how a cost-aware backtest quietly erodes what's left of the edge. That's exactly the kind of self-critical rigor I wish more "AI + trading" content had.
The AI-agent segment
To me, the AI-agent segment was the most distinctive part: a real, recorded trace of three research agents independently estimating the probability of a 2026 Fed hike, an aggregation step that flagged how correlated those "three independent opinions" actually were, three rounds of adversarial debate that never converged, and a supervisor committing to a final call that landed well below where the market was pricing it — with an explicit slide on what a single run does and doesn't prove. Refreshingly free of AI-agent hype.
My feedback
⚠My one real piece of feedback isn't about the notebooks arriving pre-written and pre-executed — with three hours on the clock, that's the right call, not a shortcut. It's that too little of those three hours then went into them: the session ran as guided checkpoints, and my perception was that slides, although deeply insightful, took up half the time or more, when the real hands-on work lives inside the pipeline itself. I'd have liked more time sitting with it instead of moving quickly past each checkpoint back to the next slide.
Q&A
Q&A stayed open throughout, and Stefan answered close to a dozen substantive and distinctive questions live. Demanding ones, like a hybrid GARCH-XGBoost volatility question, often deserved more than the few minutes a session already running long could give them — but it's worth stressing that Stefan did his best to satisfy each questioner.
Net verdict
Net: definitely worth the three hours, and worth going back through the repository slowly afterward — budget real time for that, since three hours left too little room to dwell on the pipeline itself in the room, let alone practice it live. What Stefan hands you isn't something you get to exercise deeply during the event itself, but it is a genuinely solid pipeline sketch to adapt and develop further on your own afterward. Recommended if you want an unusually honest look at where ML trading signals actually survive scrutiny.
Disclosure: I received a complimentary ticket to this event from Packt Publishing (via Anjitha M Nair). The views above are my own.
