Ep. 365: Stefan Jansen on Agentic AI, ML Workflows, and the Evolution of Machine Learning for Trading

Stefan Jansen, founder and CEO of Applied AI, returned to the podcast to discuss his completely rewritten book, Machine Learning for Trading, and the state of AI in finance. The conversation quickly settled on a sobering reality: the most hyped uses of AI in trading—letting a model make money on its own—are the least likely to work. The real value, Jansen argued, comes from using AI to support the human-led workflow of research, testing, and execution.

Why Trading Is Harder Than Everything Else

Jansen explained that applying machine learning to financial markets is fundamentally harder than in most business contexts. The core problem is a "double whammy": it is difficult to predict returns, and even harder to profit from those predictions. Unlike customer churn or ad clicks, financial markets are a "perfect competition" environment where information is rapidly incorporated and any edge decays quickly. The data itself is a constraint. "You have shorter time series," Jansen noted. "One trading session after another adds new data, it doesn't get any faster than that." This limited effective sample size makes overfitting a constant threat. A model that looks predictive in a backtest often fails because the signal was simply noise tortured into confession. Even when a signal is real, it can be eaten up by market impact and illiquidity. "You have this double whammy, right? Ha