I am thinking
whether using an AR model to do the fitting is better than using the weighted average. Using an AR model, we can remove the a’s from the optimization problem (curse of dimensionality)
Instead modeling as Yt = SMA(p1) – SMA(p2), which is essentially a AR(p2) model. Should we just model dYt = AR(p)?
Ultimately, we care only the delta for trading the next day. Also, we reduce from 2 to 1 parameter in the optimization problem.
How do we quantify when this particular model (or in general this kind of indicators) work? That is, to which type of stochastic processes this signal will apply? If we can quantify this, we can decide to turn on/off trading using this signal. Obviously, in a trending regime, this indicator would work. But could we quantify exactly what “trending regime” means?