Guest Author - Guido Deboeck

Six years ago in March of 2000, just a couple of weeks before the market collapsed which brought the S&P500 down by 45% over three years, I was lecturing at my old alma mater, the Catholic University of Leuven or the KUL in Belgium. My presentation at the time was about picking value stocks using self-organizing maps.

Two years earlier I had written a book on self-organizing maps (SOM) in collaboration with Professor T. Kohonen in Finland. The title of my book was Visual Explorations in Finance using self-organizing maps. In it I demonstrated many different examples of applications of SOM, which at the time were greatly facilitated by software provided by and produced by Dr Gerhard Kranner, CEO of Eudaptics in Austria

In case you did not read my book, imagine a computer algorithm that is capable of sorting objects without being provided sorting criteria. Say you have a shoebox in the garage that contains all kinds of screws, nails, bolts, rubber rings, wooden pegs etc. Assume you can describe each object by a dozen attributes. Now if you were to apply this computer algorithm, then the result would be a two-dimensional map on which all objects are grouped according to similarity or dissimilarity.

SOM would not group objects by say usefulness or size because it does not know what are “nails”, “screws” or “bolts”. SOM would use all the attributes that are provided and create a map. Many other examples of applications of SOM can be found in my book, but for now let’s focus on stock picking (just like I did in my lecture of March 2000).

As a starting point I merged the IBD 100, IBD New America and IBD 85-85 stock lists, which you can find on the web if you are a subscriber to IBD. I only used all the numeric columns as attributes (which means that I did not translate the non-numeric ratings of industry group strength, sales-profit-margin, and accumulation/distribution). I applied SOM (actually Viscovery Profiler from Eudaptics) on a table with 225 stocks (duplicate listings were eliminated).

The result was a map that clustered all 225 stocks into 17 clusters. Many of these clusters are very big but not interesting; it is the smaller clusters that contain the interesting stocks e.g. with average percentage off the recent high less than 5%.

The profile of the stocks in that cluster showed average EPS ranking of 90, RS ranking of 90, annual EPS growth of 34%, ROE of 17.5% etc. The cluster contained seven stocks.

From the previous article (Market timing and sector rotations) we know that the market direction (the M in CAN SLIM) turned on August 15th. If on August 15th we invested an equal dollar amount in each of the seven stocks picked by our computer algorithm then the return as of September 25th would be 13.4%.

Here are the 7 stocks that were picked by the computer out of 225 in the original merged IBD lists (here followed by the percentage increase in price since August 15th):
TNL +13.8%, NITE + 14.2%, VOLV +12.6%, IDCC +14.4%, KPN + 5.63%, TPX +18.3% and TAM + 13.3%

While our SOM picked portfolio produced 13.4% the S&P 500 produced in the same period just 3.9% and the NASDAQ 8.2%. Another proof that self-organizing maps, which is really an unsupervised neural network technique, performs a superior job over other techniques.

In my lecture in 2000 in Leuven, I stressed that if you ignore technology, software technology in this case, too much it will haunt you -- just like it haunted some clever traders and investment managers at a large international institution in Washington, but that is another story--.