To Optim, and to ALL--
I found this thread a few days ago and have now read through all the nearly 800 posts since its inception. This is an awesome thread, and I want to now thank all of you who have contributed. I have gained a wealth of information that I have not been able to find anywhere else. I really appreciate everyone's comments and dedication to openly sharing their ideas and work.
I was originally looking only for a comparison between NS Trader and Profit, but along the way I discovered many other topics that also greatly interest me. One of those is the MESA algorithm, made popular to the trading community by John Ehlers (and others) and now included in various sw packs such as Metastock. Daniel mentioned his use of MESA a long while back, and I would like to share a few thoughts about it.
Since ANN (for finance) do best with detrended data, one is always trying to find a good way to accomplish this. I have found the MESA Sinewave tool in Metastock to be excellent for this task. It detrends the data automatically and generates a sinewave that attempts to capture whatever cyclical component may be present within the data. As such, the indicator transforms and normalizes data to an oscillator that stays within the range +1 to -1.
I studied John Ehlers work in great detail and also bought his stand-alone MESA program in order to really delve into this tool. I discovered that properly choosing the lookback period (one of the MESA parameters, and the only parameter in the Metastock tool) was critical to getting good results. However, I also discovered that using a fixed lookback period, say of 5 days, produced only marginal results: some cycles were caught, but many others were not. Choosing another lookback period would generate different results, i.e. different cycles.
After much experimenting I finally devised a way to use the MESA indicator (in Metastock) in a way that was totally independent of the lookback period. Instead of choosing just one lookback period, I instead plotted a whole family of MESA sinewaves *on top of one another*, each having a different lookback period--for instance, 2,3,4,5,6,7,and 8 days. Each plot can be color-coded to distinguish it from the others, but superimposing all of them on the same graph is the key.
What results from this is a sort of "quick and dirty" decomposition of the data into individual frequency components, or waves, which is very much like Fourier analysis. But, unlike Fourier analysis, no assumptions about the data are needed, and the mechanism *dynamically* adjusts itself to the data *automatically*, and with no parameters to be adjusted. I believe it is very similar to what wavelets are trying to accomplish.
When the MESA plots are all superimposed on one another, (hopefully each with a different color), you then look for the phase relationships among the individual plots. This can show many things about the data, such as whether the data is moving in cycles, trending, or just going sideways. For example, those places in time where all the MESA waves are relatively in-phase and all peaking (or bottoming) together often denote turning points in the underlying data (local tops and bottoms).
Also, different time frames can be isolated. To depict the short-term cycles, use plots with a lookback period of, say, 2-8 days. To model intermediate-term components, use, say 8-24 day windows, etc. This allows one to decompose the data signal into whatever frequency band one wishes to examine. The spectral components of that frequency band will then appear as quasi-sinewaves.
Additionally, the sinewaves can be projected out into the future by extrapolation techniques, thus producing a leading indicator. Neural nets may be able to produce even better projections of these waves. But even without NN, I was able to write a Metastock code that did an extrapolation based on the last measured frequency of a given MESA wave. Because the sinewaves are relatively smooth and range bounded, they make excellent candidates for projection. Daniel, as you know, this may be why your own Nets liked the MESA indicator so much. It's relatively well-behaved and predictable.
In real life use, this superposition technique does NOT always give useful and accurate information. But many times it does. Mostly, though, it is fairly robust (no optimizing needed), and it is a quick and dirty way to do rudimentary spectral analysis of non-stationary, non-periodic data--financial data! Also, it may approximate wavelet techniques for those of you who don't have wavelet software.
I have more ideas about this topic if anyone is interested, but that's enough for now. This was just a way of giving back a little to this thread, since I have received so much. I look forward to futher collaborations here. Thanks All.
P.S. Has it dawned on anyone that this thread itself is a neural network, and that each of us and our posts are the inputs?! We have here all the requisite feedback loops, iterative learning, training, etc. We probably have several fitness functions running simultaneously, but a main one seems to be to come up with nets that produce good profit in the markets.
Good Learning to All of you...
Kenxx |
