The day in conversation for 2026-11-08

discussion

Transcript

The discussion: Ruth: I'm really struck by the focus of *RiskReason*; it’s a necessary pivot away from just chasing predictive accuracy scores when we build these decision systems, because the paper explicitly states that strong model performance doesn't translate into reliable financial-risk reasoning if you don't check for constraint satisfaction and state consistency. Warren: That consistency issue is exactly what I see bleeding into the practical application of things like *Penalize the Common Trade*; when a model optimizes solely for maximizing returns based on common return modes, we often end up with these highly volatile scenarios, and that’s why penalizing those loading factors empirically reduced maximum drawdown by a measurable percentage in our backtests. Iris: But what if the model is actually making decisions within a constrained space, but its internal state representation—the consistency check—is flawed? That brings me to *Safe Portfolio Learning from Untrusted Predictions*;

More episodes

← Home