I have a concern about my use of SMA and EMA, especially regarding potential market timing signals, and some of you might have opinions regarding best practice that would help me settle on an approach. The concern has to do with comparing a current value, or a short average of values to smooth the signal, to a longer period averaged to represent “normal activity.”
I see three basic choices for the longer period to average, without resorting to statistical analysis that could be misleading. The period could be one or more reporting periods, such as quarters or years, especially if dealing with fundamentals. If dealing with what could be cyclical data, the period might be chosen so that over/under conditions are approximately balanced 50% of the time. And the third choice is simply to make use of whatever period gives satisfying results, whether or not there is any underlying meaning. The third choice gives me heartburn!
For example, the 50 vs 200 day and 15 vs 50 day price averages typically used for Death Cross and Golden Cross timing seem for me to fall into the third category and don’t have much meaning other than they tend to give satisfying results. It doesn’t sit well with my desire for everything to be logically based, even though the logic might be hidden.
I have searched for a set of factors that tend to either lead or be in step with overall market direction, without using the market indexes, to be used for market timing. I have a seven factor model using one or two year averaging periods which performs nicely and satisfies a desire for the periods to “make sense.” But I also have a simpler two factor model, using periods initially chosen to smooth out the transitions and roughly balance around 50% of the time which gives nearly identical results to the seven factor model, but there is no true meaning to the periods chosen.
Is it best then to ignore the simpler model because of a higher chance of over optimizing? There are few signaling events involved, so there is a higher risk of a false model. Or am I being too analytical and should just “ride the pony” until it proves unworthy? Thoughts please!
Simpler models lead to LESS over-optimizing or overfitting in things like regressions and machine learning.
Marc and others have a LOT of success with a large number of rational factors in ranking systems and P123 sims so these may be (or even seem to be) different. But Marc (and others) aren’t just rational: they have degrees in this stuff. Marc in particular develops a rational story with features that interact in a positive way.
For things like market timing and moving averages, I think there is the risk of chasing noise with too many factors. And being out of a generally increasing market….well, you should have a reason that you are confident in or risk underperforming the market.
I think we can all agree that longer periods have greater statistical significance.
But we can probably also agree that recent events carry greater weighting.
Therefore, I contend that using the longest possible time-frames and time-varying weightings for the baseline reconciles the best of both possible worlds.
In P123, the basic tools are SMA and EMA. Here are a few other tools:
[font=courier new]ChaikinTrend(bars [, offset, increment, series])[/font] (“a special purpose double smoothed exponential average”) [which] will be both faster moving and less volatile than a basic EMA.
We can also write our own moving average functions which best reflect our feelings on how we (or the market) should weigh precedent events. For my part, I have been able to subvert the [font=courier new]LoopSum()[/font] and [font=courier new]LoopAvg()[/font] functions for my purposes. To make this work for you, I would just remind you that a weighted average is simply defined as a summation series where the weights add up to one (or an average series where the sum of weights add up to the number of terms).
Yes, you are right Jim, and I realize that I didn’t state my primary concern properly with the simpler model. It is that the only way it works is to use averaging periods that don’t have an apparently logical basis for choosing them, other than they give good results. Whereas the seven factor model’s averaging periods seem founded in logic to a certain extent.
A great way to develop an analysis, and from my perspective both of the models I list as examples make use of factors that should pass muster this way.
Truth is I am just talking. Anymore: show me the cross-validation results (positive or negative) and……Well, with the cross-validation results: it doesn’t matter what I say.
Thank you, David, for your thoughts. The logic of longer periods having greater statistical significance helps. I had not thought to try using ChaikinTrend and will now.
That can take the process to a whole different level, if I understand correctly. Not certain that I have the required confidence in keeping my biases out of the analysis!
I find that VMA(15)/VMA(210) works well as a ranking factor, for what it’s worth. In other words, three-week volume-weighted moving average to ten-month volume-weighted moving average.