Thursday, 3 November 2016

Factor multi-temporality and the layer model is born

Another element of the approach I read about when I'm reading around equity factors which I don't like much is the implict attempt to provide a single time granularity acoss all factors. 

In their steady march to generalisation, modellers have arrived at fundamental and/or economic factor models which all sit on underying observation data which occurs at the same time frequency - often end of day.  In the same way as they assume or manufacture more-or-less homogeneous equity atoms, with a singular universal semantics, so too do they expect their operation to work at the same speed.  But some equity factors could be slow moving, some could be much more rapidly moving.  They hope that the combination of steady time evolution plus a rolling window of history over which parameter updates are performed will provide enough dynamism to allow fast moving and in-play factors to come to the fore.  I'm thinking here on the fundamental factor side of momentum based factors, and on the economic factor side, of major economic announcements.  Major economic announcements, scenduled and unscheduled, can move some assets, indeed some stocks, more than others.

If Ihad additional buckets to push stocks, then this opens up the possibility of various factor models being optimised to different speeds of the market.

A factor, after all, in its most general sense, is just a series of observables which may change the expected return of a stock.    I am all for making factors be as widely scoped as possible.  If a factor exists out there which isn't commonly recognised as such but which is predictive on the expected return of a stock, then it ought to be considered.  Time does not move homogeneously for all factors, nor does a factor have to look like a factor to the factor community.

As well as looking at sector ETFs for a leg up into the world of factor modelling, I'll also be looking at multi-factor ETFs to see what's being implemented out there, as a way of getting straight to the kinds of factors which the community keep coming back to.  I expect to see some momentum factors in there, and some value/steady growth ones.  I expect there to be a preponderance of fundamental factors too - i.e. based on the idea that equity factor modelling is at heart the replacement equity analysts.  But to me non-fundamental factors will be the heart of my approach.

This regime switching idea I have can be described as follows.  For certain periods in the evolving life of a publicly listed company, the dominant factor is the beta to the market.  But at other times the regime switches and the dominant attractor factor is in-play m&a, or, stock being shorted in the lead up and launch of a new convertible issue, or stock being bought up by virtue of its inclusion (or exclusion) in a major market index.  Or that the stock has entered the 'week before FOMC announcement', it being particulaly interest rate sensitive, during earnings season, around dividend ex date periods.

In keeping with my humility approach, I set the hurdle high for regimes other than the beta regime - since that's the least damaging position to adopt.

Of course driving this all would be an asset allocation model, which again defaulted in moments of ignorance to the set of parameters which are generally considered a good mix.  This would give your stock allocation some fixed amount within which to play.

The sector/geography/ETF context would be the main habitat of a stock and it only gets stolen away by other pseudo sectors for a set of specific reasons.  To repeat, the alternative is to fully include all the relevant factors on an equal footing and to let the rolling window of calibrating market parameters to drive weight to the in play factors.  I think this is going to expose the model to sparse data problems but in some primitie sense they can be considered compatible.  In one you get the benefit of a single driving equation but weakened use of limited data below.

It is my view that one must be prepared to keep a really good model on ice for potentially years until the triggering signal is stong enough to activate it.  I shall refer to these as episodic factor sets.  Having them 'always on' and bleeding s-called explanatory power to them each day seems wrong to me.

So my model shapes up as follows:  there's an asset allocation driving layer. Within that, for each asset, there's a layer which sets a target long/short ratio.  (These two together represent your setting of the leverage).  When you have your asset size and long/short ratio, you set about finding the set of stocks in your sectors/regimes which ought to contribute positively to alpha.  FInally, especially for short term economic or fundamental factors, your way of expressing your equity exposure can be done through long or short the stock, but also through more or less complex options strategies.