The day in conversation for 2026-11-10

discussion

Transcript

The discussion: Ruth: That e-backtesting work seems to be pushing us toward a very specific kind of risk formulation under those new Basel Accords, but what I’m curious about is how robust those e-values are when you compare them against historical tail events versus purely simulated outcomes. Warren: The simulation studies mentioned in that paper suggest the optimal process construction depends heavily on the assumed underlying distribution; if we force it into a specific class of processes, does that constraint limit our ability to capture genuine model risk inherent in the market data? Iris: It sounds like you're focusing on the predictive accuracy of those forecasts, but I wonder if the advantage isn't just in the forecast itself, but in defining what constitutes 'optimal' construction—is it minimizing expected shortfall or maximizing a certain stability metric across different stress scenarios? Ruth: If we lean too heavily on stability metrics, are we implicitly ignoring potential alpha generation opportunities that might arise from higher volatility regimes which those new rules might be trying to mitigate? Warren: That brings up the CosmosAlpha paper, which uses graph structures to find factors; if the e-backtesting procedure is constrained by a specific risk measure, does factor discovery become biased toward factors that correlate strongly with that constrained risk variable? Iris: The CosmosAlpha framework relies on relational evidence propagation; how much weight are you giving to direct price correlations versus the indirect paths through the LLM structure when mining those factors? That determines whether we are finding true economic drivers or just noise amplification.

More episodes

← Home