The day in conversation for 2026-11-07

discussion

Transcript

The discussion: Iris: The finding in FinMultiHopQA that evidence structure dictates model performance seems more fundamental than just scaling up LLM parameters; it suggests the knowledge graph itself is a crucial constraint. Ruth: I agree, but I wonder if we are overemphasizing the structure aspect while neglecting the inherent noise in SEC filings; how much of that performance gap is purely structural versus inherent informational sparsity? Warren: From a modeling perspective, when you look at those long memory interactions in Unifying Empirical Enigmas, that square-root law behavior implies something about the stability of diffusive processes even with predictable order flow. Are we seeing this as a robust property across different asset classes or is it tied specifically to the liquidity profile? Iris: It seems tied more closely to the mechanism driving price impact; if the nonlinear impact terms are sufficiently dampening, then Brownian motion might be a stable approximation regardless of initial conditions, provided the memory kernel decays appropriately. What specific decay rate were they testing against in that minimal model? Ruth: That decay rate is what I’m curious about. It makes me think back to our earlier work on volatility decomposition—did those variance components show any sensitivity to the long-memory effects found in the

More episodes

← Home