QuoteOriginally posted by: JohnnyQuoteOriginally posted by: NeroTulipOf course there's a lot of truth in this statement, both TA and statarb are an inference game. But it's a bit like comparing astrophysics and astrology... The difference is that one is scientific and the other is not.Then maybe I have mis-understood the point of the "Inference Challenge". Even in the artificial circumstance where you know the underlying model, you still can't fit it reliably to the data. Is stat arb astrophysics or astrology? Sounds like the latter to me ...I think the link between the validity of stat arb strategies and the "Inference Challenge" is extremely tenuous. From my perspective, the model building process should start with what you know about the modeled phenomenon (i.e., participants, common hedging practices, common institutional restrictions, etc...). With the inference challenge we did not know anything about the structure (i.e., we knew the equations, but we did not have relationships or knowledge that we could use to customize our approach by reducing the number of variables we need to estimate in some clever way). That is part of the reason I did not bother with it. Converting knowledge about participants, institutions, and pure arbitrage or relative value relationships into a model is what these models are about. If you start the process with nothing you end the process with nothing.One needs to contemplate the underlying process and why it should/does have structure to be uncovered. The ideal combination is an arbitrage/relative value framework and a statistical model to capture the frictions that account for the basis between the liquid markets and the less liquid markets. To me stat arb is primarily a classification problem. It is about deconstructing set of instrument time series into homogeneous classes. Instruments within a class move together for a reason and stop moving together for a reason. Stat arb models are supposed to uncover structure where we know there is structure, but much of the game - from my perspective - is not really about building better and better inference tools. It is about building better and better classification algorithms.That is my 2 cents worth.Regards,