Market Outlook cum AlphaBeta Indicator Prototype 5 (contd)

0. Stocks

1. Clusters 2. Cluster Statistics

3. Short/Long Areas of the SOM

4. Growth Type Area

5. Quality (Risk Aversion) Type Area

6. Value Type Area

This week, we need to do some major re-thinking. The S&P500 was up 2.9 % for the week while we had predicted a Bearish market outlook. I realize that I was looking at the trees instead of the forest, though the prediction was wrong mainly because of the effect of Earnings Surprise % during the week . Since we are attempting to predict the direction of the S&P500 Index, we should be concentrating on the area(s) of the SOM occupied by the S&P500 components instead of the area (nodes) occupied by the ValuEngine screens. [BTW, in case all this sounds unintelligible, the explanation on methodology and terminology is at http://www.technifundamentals.com/2010/07/explanation-page.html ]. With that in mind, let's proceed.

Market Outlook

The first point to note is that S1 which is the cluster containing the least number of ValuEngine screened stocks [and therefore contains the most number of S&P500 omponents] has only 58.26 % of our stock population. S2 which contains most of the S stocks is 21.23 % of the stock population, and S3 which contains most of the L stocks is 20.51 %. This means that any prediction of market outlook will be less reliable. Previous week's % of stock population in areas clusters occupied by S or L stocks was under 10 %. I have added a new label 'I' to indicates nodes occupied by S&P500 components, and you can see that although most of the I's in image 1 are in S1, quite a number of 'I's are also in S2 and S3.

The second point to note is that we should not be comparing S2 and S3 length of bars with S1 length of bars when trying to predict the direction of the S&P500. While S1 is the closest approximation of the S&p500, it is NOT the S&P500. The S&P500 on the SOM by definition will have cluster statistics where all the bars of the model variables will be zero standard deviation i.e. length of all bars will be zero. But the best we can do is to use S1 as a proxy, just bearing in mind that the shorter the length of bars in S1, the more accurate the prediction. If we look at image 3 which shows Bullish areas (Red) and Bearish areas (Blue) we can see that in S1, the Bullish area is slightly bigger than the Bearish area. So the outlook for the market is Bullish for next week. But the signal is a weak one. Again, much will depend on Earnings Surprise %

Individual stocks

The L stocks from the ValuEngine screens are mostly in S3 and the S stocks in S2. However, as explained earlier, many non-screened stocks are also in S2 and S3 which makes the prediction for the screened stocks less reliable. This is seen in images 4,5 and 6 which show the areas occupied by the Type V, G and Q stocks. Red means presence of the Type, (value=1) while Blue means absence of the Type (value=0). However, a SOM is capable of interpolation between 1 and zero taking into account the values of the neighborhood of the nodes, and so we have non-Red, non-Blue areas too, e.g. Green, Yellow. The more the Green/Yellow areas, the less clearly defined by Type. Thus we see that V and Q have larger areas of Green/Yellow, while G is more clearly defined. From this, we can generalize that G stocks selected by the ValuEngine G screen will be more reliable to base decisions on. This also indicates that the dominant characteristic of the market next week will be an emphasis on Growth stocks. Which growth stocks? See image 0 (except Whirlpool)

* Note I have put aside Alpha and Beta for the moment while I am working towards a quantitative interpretation for Alpha.

Market Outlook cum AlphaBeta Indicator Prototype [4] (contd)

1. Self-Organized Clusters 2. Cluster Statistics

3. Short/Long Window

4. EPS Surprise % Window

5. Beta Window

6. Last 12-Month Return % Window

7. Market Cap Window

8. Selected Stocks

The S&P was down by about 1 % last week while we had predicted a mild Bullish outlook. Most of the down action happened on Friday (-2.88 %), so our prediction was not such a disaster. This week, we have made some improvements to the model. The attribute windows are now much more defined [more colorful] as we plotted the transforms of the attributes instead of the original values. Defining nominals and using interpolation we can now plot Short/Long as well Type [V,G,Q] on the SOM where 1= the presence of a defined nominal and 0= absence. Just to remind that the explanation of methodology and terminology is at http://www.technifundamentals.com/2010/07/explanation-page.html . And also to remind that the difference between our indicators and ordinary technical analysis indicators is that our technicals are based on fundamentals viz using ValuEngine pre-screened stocks. That's the reason for the name of this Blog (Technifundamentals). Now, being based on fundamentals means that it's not unusual to see selected stocks fall, as short-term noise clouds the fundamentals, only to rise again the week after. This is the case for one of last week's stocks- Gannet GCI-which despite earnings of 61 cents per share , up from last year's 30 cents a share and beating analyst's estimate of 53 cents a share- dived 10.7 %. ValuEngine screening criteria also include a minimum market cap and minimum average daily volume, so that leaves out the smaller and more illiquid stocks and reduces volatility.
Market OutlookIn image 1, the L stocks are mostly in cluster S2 but a number of them are also in S3 which has a majority of S stocks. This 'dilutes' the strength of the signal, and we have a rather ambivalent market outlook next week. This can be seen in image 2 cluster statistics where the bars for Short/Long are negative in S1 [which contains most the S&P500 components] and S3 and only positive in S2. In image 3, Red area is Long while Blue area is short. Red area is slightly bigger than Blue area, but Blue area is more defined, with S stocks bunched in a sub-cluster i.e. nodes of S stocks are closer together. On the other hand besides the sub-cluster of L stocks, the other L stocks are spread sparsely over the Red area. On a SOM it indicates a less homogenous grouping. On balance the outlook is more Bearish than Bullish. Type-wise or Sector-wise there is no clear pattern. V, G and Q stocks as well stocks from Sectors B,C,D, E,F,H,ND,S,T,TP,or U are distributed almost randomly over the topology of the SOM. Also, comparing the length of the bars in S2(L cluster) and S3(S cluster), in general the S3 bars are longer. Since the length of bars denote the standard deviation from the Mean of the entire data set, this implies that Bearish indicators as represented in S3 have a stronger signal strength than the Bullish indicators in S2. S3 Lastly in S1 [Blue cluster] which most approximates the Index, the ratio of L to S stocks is 8/26 or a very low 0.30.

Alpha/Beta

EPS Surprise % will still play a big part in selection of Long stocks. Earnings Surprise % here is of course ex-poste i.e. a stock's past record in Earnings Surprise i.e. deviation from analyst estimates (Please see the EPS Surprise % bars in image 2]. And here we are assuming that a stock that has a good record of springing Earnings Surprise is more likely to do so again. Using the EPS Surprise attribute window (image 4) to manually select the nodes [darkened areas] with the highest EPS Surprise %. we find that the stocks selected have certain attributes: they have a high Beta [image 5], high 12-month Return % (Momentum) [image 6] and smaller market cap [image 7]. We can sum up by saying that it's the mid-caps with high Beta and momentum that will show the greatest Earnings Surprise %. The average Beta of S2 Bullish cluster is 2.13 while the average Beta of the S3 Bearish sector is 1.71. i.e. the risk is greater on the Bullish side. Other than Earnings Surprise % there are no other signifcant factors contributing to a stock's Alpha at the moment. ValuEngine's Senior Analyst Steve Hach showed me an article titled "Has Alpha Turned To Beta?". That about sums up the market at this point of time. If you must play the market go do something safer and less volatile-like buying a reverse ETF of the S&P500.

Market Outlook cum Alpha/Beta Indicator Prototype (3)[contd]

1. Self-organized clusters formed by ValuEngine model screens
2. Beta window of the Self-Organized map

3. Statistics of the clusters S1, S2 and S3

4. Summary of cluster statistics

5. Selected stocks from the Long cluster S3 The Market Outlook cum Alpha/Beta Indicator [MOCABI]is still undergoing testing although I am slowly nailing down some of the issues. Last week, I predicted that the selling had abated and there would be a mild rebound. Well, maybe I was wrong in using the word 'mild'. The S&P500 increased by 5 % or so for the week. Let's see how we do this week. I'll try to make the analysis as quantitative as possible. The Explanation Page for MOCABI methodology is at http://www.technifundamentals.com/2010/07/explanation-page.html.

1. S1 holds the S&P500 stocks, S2 most of the Short (S) stocks and S3 most of the Long (L) stocks. This is the same order as last week. The average Beta of S2 stocks was 1.60 last week and is now 1.50. The average Beta of S3 stocks was 1.98 and is now 2.62. There is more 'sensitivity' to the Index on the upside than on the down side.

2. Last week the ratio of L stocks/ S stocks that are in the S1 cluster which holds the S&P 500 stocks was 28/21=1.33. This week it is 42/29= 1.44. This is an indication of increased Bullishness for next week.

3. S3, the cluster which has Bullish characteristics holds only 6.14 % of our stock Universe for this SOM while S2 the cluster with the Bearish characteristics holds 17.33 % . This point somewhat tempers the Bullishness of (2) above.

4. In image 3 , the length of the bars denoting our Model variables above measures the deviation of the cluster Mean from the Mean of the entire data set. Thus, the longer the bars, the more those stocks in the cluster with those bars will differ from the performance of the Index as represented by S1 bars. The longest bars in S3 with the L stocks are Beta (first Green bar) Momentum [12-m return%] fourth bar Mauve color and Volatility (Red bar). In a Bullish market, high momentum, high Beta and high volatility means a sharp run-up. Also take a look at the Purple bar next to the Red Volatility bar, which measures EPS surprise % in the stock's history . Note its length in S3 as well as S2. EPS surprise will be an important factor in a stock's movement for the coming week. Thus the sharp run-up is based on fundamentals, particularly EPS surprise.

4. On the whole, it can be summarized thus: The market outlook is decidedly more Bullish this week as compared to last week. But only a few stocks will have move significantly. The stocks will be those with good fundamentals and the run-up will be speedy. The L stocks from the ValuEngine L screens will move up more than the S stocks from the ValuEngine S screens will move down.

5. Image 5 shows the selected stocks from S3 cluster. Those that are listed more than once are the output of more than one screen. Type V, G, and Q represents Valuation, Growth and Quality. The multi-listed stocks without V, G or Q are also components of the S&P500 Index. Note that most of the screened stocks come from the Q model and not the V or G model. So although the market will run on momentum, volatility and beta, it will be the Quality stocks [i.e. those with good fundamentals and risk/reward ratio] that run. This is confirmed by the statistics bars of S3 cluster for P/E, M/B (market/book) and P/S (price/sales) which are all significantly different (in a good way) from S1 the reference cluster and S2 the S cluster i.e. P/E, M/B,P/S and Valuation are all lower in value.

Market Outlook-cum-AlphaBeta Indicator Prototype(contd)

1. This week's clusters

2. Last week's clusters

3. Beta window
4. Cluster statistics

5. Statistics summary

6. Stock list

Last week, our Market Outlook-cum-AlphaBeta Indicator [MOCABI] had a pessimistic view of the market, and this proved to be true. This week, we continue to explore MOCABI]. I have put all the explanation of methodology and terminology on a stand-alone page here: http://www.technifundamentals.com/2010/07/explanation-page.html.

There have been some changes in MOCABI and it will continue to have changes as it is being developed. This week's MOCABI is not a creation of a new model using this week's data. Instead, this week's data is applied by last week's model which has been refined by Dr. Gerhard Kranner of Viscovery. So we are monitoring the results. Also, at the suggestion of Dr. Kranner, Alpha should be defined and calculated in the conventional way. My 'unconventional' definition of Alpha [see explanation] I will call by another letter of the Greek alphabet, maybe Zeta. We shall then post stock lists as selected by high Alpha and high Zeta and compare their subsequent performance. We are also debating the interpretation of the SOM output and will refine the model accordingly. In the meantime, it's on with the show:

This week's market outlook

In times of low confidence and jittery nerves in the market, the average Beta of stocks rise. Thus, no matter how good a stock picker you are, for the short term, you are at the mercy of the market, and perhaps 80 % of a stock's performance is determined by its Beta. Comparing between image 1 and image 2, shows that there is not much change in the situation as evidenced by the three clusters, and their not-much-changed sizes. Most of the Long [L] stocks are in S3 but S3 only comprises 7.52 % of our stock Universe, while S2 which contains the Short stocks [S] comprises 19.63 %. The average Beta of S3 is 1.985 which is higher than the average Beta of S2 which is 1.604 so the L stocks in S3 are more sensitive to the overall market. S1 with a average Beta of 1.189 confirms that our SOM is well constructed- S1 which comprises 72.84 % our stock Universe, is approximately the S&P500 and of course the Beta of the Index is defined as 1.0. The high Beta of selected S3 L stocks is shown by the stock list in image 6. In such a wishy-washy situation, I would base my prognosis of the market outlook for this week, by looking at the number of L and S stocks in S1. There are 28 L stocks and 21 S stocks in S1. Thus the optimistic/pessimistic ratio is positive at 28/21= 1.33. Iwould venture to say that the selling has abated and this week will see a mild rebound. Is there any point in going on to select stocks even if they are high Alpha and low Beta? I do not think so. And there's no point picking stocks to Short too, as the Risk/Reward ratio for going Short does not justify it. Let's wait till next week when we see if the average Beta level has come down.

Explanation Page

Explanation of MOCAB Indicator prototype.
Introduction
Stock markets are Complex Adaptive Systems [CAS] with all the unique properties of such systems. Properties of CAS such as feedback loops, emergence, self-organization, self-similarity, co-evolution and distributed connectivity require that a meaningful analysis of CAS be done with tools that take into account such complexities. To deal with the complexities of CAS requires a paradigm shift from hard computing to soft computing, from linear parametric modelling to non-linear non-parametric modelling and from exact solution to approximate solution.While most of the world’s phenomena (natural or man-made) are CAS, it is only in the last decade or so that with the exponential leap in the power of computers and software, we are able to use tools more suited to the analysis of CAS. Such tools include neural networks, fuzzy logic, evolutionary algorithms, wavelets,swarm intelligence and, in particular, Self-Organizing Maps Click www.viscovery.net/self-organizing maps for a short summary).Neural Networks are a class of Artificial Intelligence and a SOM is a class of Neural Network that mimic the biology of the human brain. Neural Networks are capable of associative memory recall, pattern recognition, classification, forecasting, optimization and noise filtering - which are all forms of generalization. In other words, neural networks learn from specific situations and apply the learning to new, hitherto unencountered situations. Most Neural Networks are supervised networks, i.e. they are ‘taught’ the relationship between input data and a target variable.On the other hand, a SOM is one of the few classes of unsupervised Neural Networks. A SOM does its work without having to be ‘taught’. The SOM algorithm self-organizes data of similar characteristics into clusters, very much like the brain does. The most useful feature of a SOM is that it can be used for the exploration, classification, analysis and visualization of large sets of multi-dimensional data. We can illustrate this point by taking an example from the stock market data of a Company. If data is available it is possible to plot the relationship between a stock’s price and its PE ratio as a two-dimensional relationship which can easily be visualized. A three-dimensional visualization is also possible if we have a third variable e.g the stock’s Price/Book Value ratio and three axes x,y,z to construct a 3D chart. But for complex models such as the ValuEngine models where close to 30 variables are involved, it is an impossible task to graphically depict the inter-relationships between all the variables. Only a SOM can do that. A SOM represents a perceptual space where data objects have been ordered in a “landscape” with respect to their overall similarity. SOMs have a wide variety of application in various fields from predictive analytics for marketing, to classification of wines, detection of credit card fraud, optical character recognition and medical imaging.
The MOCAB Indicator.
The MOCAB Indicator that is compiled at the end of each trading week, contains information to gauge the market outlook for the coming week based on the situation as at Close on Friday. (if Friday is not a trading holiday) and to select stocks with high Alpha and the appropriate Beta. The input of the MOCAB Indicator are the time-tested stock analysis Models of ValuEngine Inc of Princeton, NJ and the output and analysis is done using the SOM-based Data Mining software of Viscovery Software GmbH in Vienna, Austria.
MOCAB Indicator Information Content
Market Mode: The screening output of the three ValuEngine models (Valuation, Growth and Quality) and how they are positioned on the SOM will be an indicator of the strength of each mode.
Long/Short: Each of the three ValuEngine screens has Long/Short versions, and how the L and S stocks are clustered and positioned on the SOM will be an indicator of the market strength.
Sector Information: Each of the screened output is also labelled to indicate the sector they belong to, and how they are clustered and positioned on the SOM may yield information on sector strength .
High Alpha stocks: Alpha is the return in excess of the return of some market Index. In other words, the non-Beta part of a stock's movement or to put it even more clearly the return on a stock if the market return were zero. In the context of this Blog and its SOM technology, Alpha is defined as the degree of dis-similarity of a stock with a market Index and is represented by the cluster that has the most difference with the cluster approximating the Index in terms of exhibiting desirable values of the ValuEngine model variables. And within this cluster, the individual stocks that most represents the properties of the cluster are the high Alpha stocks. On a SOM, overall degree of similarity/dis-similarity is measured within a cluster by the Euclidean [mathematical space] distance between nodes with the greater the distance indicating the greater the dis-similarity. Among clusters on a SOM, the degree of dis-similarity of a cluster with a market Index is measured by the deviation of the Mean of the cluster from the Mean of the entire data set or a cluster approximating the market Index. * for justification of my definition see section on heuristic approach and holistic perspective below.
Beta: Beta is a measure of a stock’s performance relative to the general market. It is calculated by doing a regression analysis of a stock and the market’s price movement over a period of time. The market as represented by an Index is designated a Beta of 1, and so a stock with a Beta of 1.5 will theoretically move 50 % more than the market when it is going up and also when it is going down. Our objective is to look for high Alpha high Beta stocks when the market is trending up and high Alpha low Beta stocks when the market is trending down. On a SOM, high Beta or low Beta stocks can be identified using the Beta attribute map which plots the model variable Beta as a ‘heat map’ with color intensity scale tending towards Red for high Beta and towards Blue for low Beta.
Heuristic approach based on domain expertise.
1. It is my belief that beyond a point, it is necessary for quantitative analysis of financial markets to incorporate heuristics based on domain expertise and experience. It is also my belief that a heuristic approach brings with it a more holistic perspective which is essential for the soft sciences like Economics and Finance that have to contend with human behaviour. Shu-Heng Chan and Paul P. Wang as editors of Computational Intelligence in Economics and Finance [Springer-Verag 2004] have also mentioned the need for a heuristic approach based on domain expertise because of the special issues that economics and finance modelling involve viz extremely noisy data, behavioural changes and non-linear relationships. To which I might add the issues of long fat-tailed probability distributions, Black Swan events, missing data, heteroskedasticity, autocorrelation and frequent regime switches. To quote from their book: “In the domain of highly complex problems, precision is neither possible nor often desirable. Heuristics or approximate algorithms become the only acceptable tools” A sentiment also shared by Prof. Lotfi Zadeh, the father of Fuzzy Logic. A heuristic approach is also an inherently more robust approach with more room for accommodating higher degrees of uncertainty a point not to be dismissed considering the characteristics of modern financial markets.
2.The market can be described by three modes which can be characterized as: Valuation, Growth and Quality. Valuation mode is characterized by investors’ emphasis on fundamentals with the accompanying technical characteristics of Oversold/Overbought and reversion to the Mean . Growth mode emphasises the future, places less weighting on present fundamentals and is accompanied by the technical characteristics of Momentum and Trend. Quality mode is concerned with volatility and stocks are selected based on their risk/reward ratio as represented by a metric such as the Sharpe Ratio. At any point in time, the market is a combination of various degrees of Valuation, Growth and Quality.
3. The definition of Alpha using SOM is a ‘purer’ and more holistic definition of Alpha. The traditional Alpha which can be depicted mathematically as the point where the Beta line intersects the Y axis is an ex-post statistic calculated from the Beta and dependent on (arbitrary) choice of time period for its calculation and has little predictive value. It is a market dynamic statistic like a technical analysis indicator. The definition of Alpha that is used here measures dis-similarity based on all the variables of the SOM model which are in turn derived from the fundamentals-based ValuEngine models and therefore have predictive value since fundamentals such as earnings, sales, book value, cash flow, market cap, earnings surprise, yield on long term treasuries, etc have been proven to have predictive value for medium and longer term investment time frame.
4. Alpha values alone is not sufficient for stock selection. Alpha and Beta must be used together. On a SOM we can use the ‘heat map’ of the model variable Beta to pinpoint the high Beta or low Beta stocks among the stocks which in the previous step had been selected for high Alpha. Our methodology selects the best of the best in the sense that the selected stocks could be from any of the three ValuEngine screens based on the three ValuEngine models Valuation, Growth and Quality. Our method also does not limit us to a fixed number of selected stocks. After the high Alpha stocks have been selected we discard those with undesirable Beta values. In a up-trending market, stocks with high Alpha and high Beta are selected to take advantage of the upward move. But high Beta is a double-edged sword, and in a down-trending market stocks with high Beta will also move down more than the market. Therefore in a down-trending market, we should select stocks with high Alpha and low Beta. In a sideways trendless market, it is safer to stick to high Alpha and low Beta stocks.
Methodology
1.A SOM of the component stocks of the S&P500 is created. This SOM represents the basic topology of the market.
2. Six sets of 20 (22 stocks for Growth model) stock portfolios based on Long and Short versions of the three ValuEngine screens [Valuation, Growth, Quality] are created. (1). Valuation Long. (2).Valuation Short (3). Growth Long (4).Growth Short (5). Quality Long (6). Quality Short. * Note ValuEngine uses a different name for their screens. For our purpose, ValuEngine Standard= Valuation; ValuEngine Forecast= Growth, and ValuEngine Star= Quality. Note #2: Growth model has 22 stocks instead of 20 because it’s set-up of 2 stocks per sector out of the 11 sectors S&P500 is fixed.
3. The screened stocks are marked such that the screen from which they originated, the sector they belong to, and whether they are Long or Short positions are all included in the selection. In addition, some of the stocks are labeled in the map. The position of a label is approximately the position of the node on the SOM that the stock occupies. The S&P500 component stocks are not labelled and the empty spaces represent the nodes on the SOM that they occupy.
4.The acronyms used in the labelling are:V=Valuation; G=Growth; Q=Quality; L=Long; S=ShortB= Basic Industries C= Capital Goods D= Consumer Durables E= EnergyF= Finance H= Healthcare ND= Consumer Non-Durables S= Consumer Services T= Technology TP= Transportation U= Public Utilities* S&P500 stocks are not labeled and thus occupy the 'empty' spaces on the SOM. Thus GLT is a growth model Long stock from the Technology sector and VSS is a valuation model Short stock from the consumer services sector.

Market Outlook-cum-AlphaBeta Indicator (Prototype)

1. Beta Attribute Map
2. Self-Organized Clusters

3. Cluster Statistics
4. Cluster Statistics Summary
5. Long Stocks
6. Short Stocks

This is a lengthy post explaining my methodology for a prototype Market Outlook-cum-Alpha/Beta Indicator (MOCAB). For the benefit of those not interested in explanations I'll begin with the market analysis and stock selection for this week based on MOCAB and then follow up with the explanation on methodology and labelling. But first I need to mention that this is not some kind of technical analysis indicator based on just Price and Volume data. This is an indicator based on the clustering of fundamental data and on the screens of ValuEngine models. Without the data and software provided by the good people at ValuEngine, my work would not have been possible. As mentioned, MOCAB Indicator is currently a prototype and as the weeks go by, it will be fine-tuned; and for this I would like to express my thanks for the mentoring and assistance given to me by Dr. Gerhard Kranner, CEO of Viscovery.
Market outlook for coming week
This is going to be another nail-biting week, with a bias towards 'down'. Cluster S1 holds most of the S&P500 component stocks, but many of the Short stocks (with 'S' as the second letter of their label) in the ValuEngine screens are also in S1, and since S1 is approximately the Index, the implication is that the Index will be bearish. Another bearish sign: Most of the other Short stocks are in cluster S3, while the Long stocks (second letter of their label is 'L') are in S2. S3 is a more cohesive cluster (smaller area) than S2 and the stocks in it are closer together i.e. more homogenous. Which means the bearish signal (as represented by the S stocks) is stronger than the bullish signal (as represented by the L stocks). Yet another sign: Image 3 the cluster statistics shows that in general S3 bars are much longer than S2 bars. Since the length of the bars represents deviations of the cluster from the entire data set, and S1 is the closest approximation of the entire data set, the much longer bars of S3 as compared to S2 means that the bearish (Short) stocks in S3 have a greater deviation from the market than the bullish (Long) stocks in S2. So there you are, thats the general outlook for the market
High Alpha, Low Beta stocks
It has been confirmed by many studies that about 70 % of the gain (or loss) from a stock is due to the market. The tide lifts all boats up and down. With such a market outlook for the coming week, we should be looking at high Alpha AND low Beta stocks. High Alpha stocks with high Beta are not going to fare very well. With that in mind, I used the Beta attribute window (image 1) to look at those areas on the map with low Beta. (see sliding scale at bottom of map). Then in those areas I selected the high Alpha stocks (see below for how high Alpha stocks are selected).
The pickings as shown in image 5 are slim. And these stocks while having the lowest Beta among the ValuEngine screened Long stocks still have a relatively high Beta value of 1.73, 1.77 and 2.13. The small number of suitable stocks and the high Beta of these 'suitable' stocks, is an indication that theres not much scope for playing the market this coming week with even the high Alpha stocks swinging up and down with it. Also, there is not a significant degree of dis-similarity between the S1 control cluster and S2 the Long stocks cluster as seen by the difference in the length of bars in the Statistics bar graph. Which is an indication that the Alpha strength is not high and movements in the general market are more likely to have an impact even on the selected stocks. Caveat: The market outlook is based on all known information as at end of Friday close, and not taking into account any external shocks such as economic and political developments and natural disasters.
BTW, Image 6 shows the Short picks. There seems to be a strong sell sign for Cemex and Dania Holding Corp.
Now comes the explanation of what MOCAB Indicator is all about.
Introduction

Stock markets are Complex Adaptive Systems [CAS] with all the unique properties of such systems. Properties of CAS such as feedback loops, emergence, self-organization, self-similarity, co-evolution and distributed connectivity require that a meaningful analysis of CAS be done with tools that take into account such complexities. To deal with the complexities of CAS requires a paradigm shift from hard computing to soft computing, from linear parametric modelling to non-linear non-parametric modelling and from exact solution to approximate solution.While most of the world’s phenomena (natural or man-made) are CAS, it is only in the last decade or so that with the exponential leap in the power of computers and software, we are able to use tools more suited to the analysis of CAS. Such tools include neural networks, fuzzy logic, evolutionary algorithms, wavelets,swarm intelligence and, in particular, Self-Organizing Maps Click www.viscovery.net/self-organizing maps for a short summary).
Neural Networks are a class of Artificial Intelligence and a SOM is a class of Neural Network that mimic the biology of the human brain. Neural Networks are capable of associative memory recall, pattern recognition, classification, forecasting, optimization and noise filtering - which are all forms of generalization. In other words, neural networks learn from specific situations and apply the learning to new, hitherto unencountered situations. Most Neural Networks are supervised networks, i.e. they are ‘taught’ the relationship between input data and a target variable.
On the other hand, a SOM is one of the few classes of unsupervised Neural Networks. A SOM does its work without having to be ‘taught’. The SOM algorithm self-organizes data of similar characteristics into clusters, very much like the brain does. The most useful feature of a SOM is that it can be used for the exploration, classification, analysis and visualization of large sets of multi-dimensional data. We can illustrate this point by taking an example from the stock market data of a Company. If data is available it is possible to plot the relationship between a stock’s price and its PE ratio as a two-dimensional relationship which can easily be visualized. A three-dimensional visualization is also possible if we have a third variable e.g the stock’s Price/Book Value ratio and three axes x,y,z to construct a 3D chart. But for complex models such as the ValuEngine models where close to 30 variables are involved, it is an impossible task to graphically depict the inter-relationships between all the variables. Only a SOM can do that. A SOM represents a perceptual space where data objects have been ordered in a “landscape” with respect to their overall similarity. SOMs have a wide variety of application in various fields from predictive analytics for marketing, to classification of wines, detection of credit card fraud, optical character recognition and medical imaging.The MOCAB Indicator.The MOCAB Indicator that is compiled at the end of each trading week, contains information to gauge the market outlook for the coming week based on the situation as at Close on Friday. (if Friday is not a trading holiday) and to select stocks with high Alpha and the appropriate Beta. The input of the MOCAB Indicator are the time-tested stock analysis Models of ValuEngine Inc of Princeton, NJ and the output and analysis is done using the SOM-based Data Mining software of Viscovery Software GmbH in Vienna, Austria.

MOCAB Indicator Information Content

Market Mode: The screening output of the three ValuEngine models (Valuation, Growth and Quality) and how they are positioned on the SOM will be an indicator of the strength of each mode.

Long/Short: Each of the three ValuEngine screens has Long/Short versions, and how the L and S stocks are clustered and positioned on the SOM will be an indicator of the market strength.

Sector Information: Each of the screened output is also labelled to indicate the sector they belong to, and how they are clustered and positioned on the SOM may yield information on sector strength .

High Alpha stocks: Alpha is the return in excess of the return of some market Index. In other words, the non-Beta part of a stock's movement or to put it even more clearly the return on a stock if the market return were zero. In the context of this Blog and its SOM technology, Alpha is defined as the degree of dis-similarity of a stock with a market Index and is represented by the cluster that has the most difference with the cluster approximating the Index in terms of exhibiting desirable values of the ValuEngine model variables. And within this cluster, the individual stocks that most represents the properties of the cluster are the high Alpha stocks. On a SOM, overall degree of similarity/dis-similarity is measured within a cluster by the Euclidean [mathematical space] distance between nodes with the greater the distance indicating the greater the dis-similarity. Among clusters on a SOM, the degree of dis-similarity of a cluster with a market Index is measured by the deviation of the Mean of the cluster from the Mean of the entire data set or a cluster approximating the market Index. * for justification of my definition see section on heuristic approach and holistic perspective below
Beta: Beta is a measure of a stock’s performance relative to the general market. It is calculated by doing a regression analysis of a stock and the market’s price movement over a period of time. The market as represented by an Index is designated a Beta of 1, and so a stock with a Beta of 1.5 will theoretically move 50 % more than the market when it is going up and also when it is going down. Our objective is to look for high Alpha high Beta stocks when the market is trending up and high Alpha low Beta stocks when the market is trending down. On a SOM, high Beta or low Beta stocks can be identified using the Beta attribute map which plots the model variable Beta as a ‘heat map’ with color intensity scale tending towards Red for high Beta and towards Blue for low Beta.

Heuristic approach based on domain expertise.

1. It is my belief that beyond a point, it is necessary for quantitative analysis of financial markets to incorporate heuristics based on domain expertise and experience. It is also my belief that a heuristic approach brings with it a more holistic perspective which is essential for the soft sciences like Economics and Finance that have to contend with human behaviour. Shu-Heng Chan and Paul P. Wang as editors of Computational Intelligence in Economics and Finance [Springer-Verag 2004] have also mentioned the need for a heuristic approach based on domain expertise because of the special issues that economics and finance modelling involve viz extremely noisy data, behavioural changes and non-linear relationships. To which I might add the issues of long fat-tailed probability distributions, Black Swan events, missing data, heteroskedasticity, autocorrelation and frequent regime switches. To quote from their book: “In the domain of highly complex problems, precision is neither possible nor often desirable. Heuristics or approximate algorithms become the only acceptable tools” A sentiment also shared by Prof. Lotfi Zadeh, the father of Fuzzy Logic. A heuristic approach is also an inherently more robust approach with more room for accommodating higher degrees of uncertainty a point not to be dismissed considering the characteristics of modern financial markets.

2.The market can be described by three modes which can be characterized as: Valuation, Growth and Quality. Valuation mode is characterized by investors’ emphasis on fundamentals with the accompanying technical characteristics of Oversold/Overbought and reversion to the Mean . Growth mode emphasises the future, places less weighting on present fundamentals and is accompanied by the technical characteristics of Momentum and Trend. Quality mode is concerned with volatility and stocks are selected based on their risk/reward ratio as represented by a metric such as the Sharpe Ratio. At any point in time, the market is a combination of various degrees of Valuation, Growth and Quality.

3. The definition of Alpha using SOM is a ‘purer’ and more holistic definition of Alpha. The traditional Alpha which can be depicted mathematically as the point where the Beta line intersects the Y axis is an ex-post statistic calculated from the Beta and dependent on (arbitrary) choice of time period for its calculation and has little predictive value. It is a market dynamic statistic like a technical analysis indicator. The definition of Alpha that is used here measures dis-similarity based on all the variables of the SOM model which are in turn derived from the fundamentals-based ValuEngine models and therefore have predictive value since fundamentals such as earnings, sales, book value, cash flow, market cap, earnings surprise, yield on long term treasuries, etc have been proven to have predictive value for medium and longer term investment time frame.

4. Alpha values alone is not sufficient for stock selection. Alpha and Beta must be used together. On a SOM we can use the ‘heat map’ of the model variable Beta to pinpoint the high Beta or low Beta stocks among the stocks which in the previous step had been selected for high Alpha. Our methodology selects the best of the best in the sense that the selected stocks could be from any of the three ValuEngine screens based on the three ValuEngine models Valuation, Growth and Quality. Our method also does not limit us to a fixed number of selected stocks. After the high Alpha stocks have been selected we discard those with undesirable Beta values. In a up-trending market, stocks with high Alpha and high Beta are selected to take advantage of the upward move. But high Beta is a double-edged sword, and in a down-trending market stocks with high Beta will also move down more than the market. Therefore in a down-trending market, we should select stocks with high Alpha and low Beta. In a sideways trendless market, it is safer to stick to high Alpha and low Beta stocks.
Methodology
1.A SOM of the component stocks of the S&P500 is created. This SOM represents the basic topology of the market.
2. Six sets of 20 (22 stocks for Growth model) stock portfolios based on Long and Short versions of the three ValuEngine screens [Valuation, Growth, Quality] are created. (1). Valuation Long. (2).Valuation Short (3). Growth Long (4).Growth Short (5). Quality Long (6). Quality Short. * Note ValuEngine uses a different name for their screens. For our purpose, ValuEngine Standard= Valuation; ValuEngine Forecast= Growth, and ValuEngine Star= Quality. Note #2: Growth model has 22 stocks instead of 20 because it’s set-up of 2 stocks per sector out of the 11 sectors S&P500 is fixed.
3. The screened stocks are marked such that the screen from which they originated, the sector they belong to, and whether they are Long or Short positions are all included in the selection. In addition, some of the stocks are labeled in the map. The position of a label is approximately the position of the node on the SOM that the stock occupies. The S&P500 component stocks are not labelled and the empty spaces represent the nodes on the SOM that they occupy.
4.The acronyms used in the labelling are:V=Valuation; G=Growth; Q=Quality; L=Long; S=ShortB= Basic Industries C= Capital Goods D= Consumer Durables E= EnergyF= Finance H= Healthcare ND= Consumer Non-Durables S= Consumer Services T= Technology TP= Transportation U= Public Utilities* S&P500 stocks are not labeled and thus occupy the 'empty' spaces on the SOM. Thus GLT is a growth model Long stock from the Technology sector and VSS is a valuation model Short stock from the consumer services sector.

Back To The Drawing Board

1. The portfolio after two weeks
2. A Market Direction Gauge-cum- Stock Picker

3. Cluster Statistics

4. The High-Alpha Stocks

After two weeks, our monthly re-balance market-neutral portfolio is down 1.42 % while the S&P500 is up 5.18 %. That's because of our Short position via the ProShares S&P500 ETF/ (Ticker symbol SH). If not for SH, our tech-dominated portfolio would be up 2.32 %. This is still underperforming the Index but we still have two weeks to see whether the tech stocks will power on. Still, the lesson to be learned from this is that we should not ignore the basic fact that the natural tendency of stocks is to have an upward bias. Thus giving a 50% weight to the Short position is a mistake. An algorithm must be devised for the weighting of the Short component of a MNS strategy.

This week we attempt to (1)create a market direction gauge with SOM and (2) refine our technique for using SOM to pick high Alpha stocks.

Approach is based on my following beliefs that:
1.The market can be completely described by three modes: Valuation, Growth and Quality. Valuation mode is characterized by investors’ emphasis on fundamentals with the accompanying technical characteristics of Oversold/Overbought and reversion to the Mean . Growth mode emphasises the future, places less weighting on present fundamentals and is accompanied by the technical characteristics of Momentum and Trend. Quality mode is concerned with volatility and stocks are selected based on their risk/reward ratio as represented by a metric such as the Sharpe Ratio. At any point in time, the market is a combination of various degrees of Valuation, Growth and Quality.
2.Alpha which is the risk-free return in excess of some market Index is defined as the degree of dis-similarity of a stock with the market index. In other words, the non-Beta part of a stock's movement. The higher the degree of dis-similarity, the higher the Alpha. In the context of this Blog and its SOM technology, Alpha is represented by the cluster that has the most difference with the cluster approximating the Index in terms of exhibiting desirable values of the ValuEngine model variables. And within this cluster, the individual stocks that most represents the properties of the cluster are the high Alpha stocks. On a SOM, overall degree of similarity/dis-similarity is measured in the cluster by the Euclidean [mathematical space] distance between nodes, and between clusters by the deviation of the cluster Mean from the Mean of the entire data set. Picking high Alpha stocks means picking stocks which outperform the market index when the market is rising and under-perform the Index when the market is falling . . i.e. falling at a slower rate than the Index.

3. Our analysis is based on the assumption that the output of each ValuEngine screen accurately reflect their respective model selection criteria of valuation,growth and quality.

4. This approach is also based on my philosphy of incorporating heuristics based on experience into quantitative analysis..
Methodology
1.A SOM of the component stocks of the S&P500 is constructed. This SOM represents the basic topology of the market.
2. Six sets of 20-stock portfolios based on Long and Short versions of the three ValuEngine screens [Valuation, Growth, Quality] are created. (1). Valuation Long. (2).Valuation Short (3). Growth Long (4).Growth Short (5). Quality Long (6). Quality Short. * Note ValuEngine uses a different name for their screens. For our purpose, ValuEngine Standard= Valuation; ValuEngine Forecast= Growth, and ValuEngine Star= Quality.
3.Prior to overlaying on the SOM, each of the screened stocks are labeled such that the screen from which they originated, the sector they belong to, and whether they are Long or Short positions are all denoted by the label. The position of a label is approximately the position of the node on the SOM that the stock occupies. The S&P500 component stocks are not labeled and the empty spaces represent the nodes on the SOM that they occupy.
4.The acronyms used in the labelling are:V=Valuation; G=Growth; Q=Quality; L=Long; S=Short
B= Basic Industries C= Capital Goods D= Consumer Durables E= Energy
F= Finance H= Healthcare ND= Consumer Non-Durables S= Consumer Services T= Technology TP= Transportation U= Public Utilities
* S&P500 stocks are not labeled and thus occupy the 'empty' spaces on the SOM. Thus GLT is a growth model Long stock from the Technology sector and VSS is a valuation model Short stock from the consumer services sector.
Market Outlook

Image 2 shows that the SOM is composed of three natural clusters: S1 the biggest cluster contains 66.33 % of our stock population. With a lot of space unoccupied by the ValuEngine screened stocks, S1 is where most of the S&P500 stocks are located. S2 contains 21.90 % of the stock population and is mostly occupied by the Short (S) stocks. S2 contains 17.77 % of the stock population and is mostly occupied by the Long stocks (L). To have a sense of the market direction for next week, we first look at the integrity of S2 and S3 clusters and their degree of dis-similarity with the market as represented by the S&P500 (the data set from which our SOM is constructed). Image 3 is an indicator of the degree of dis-similarity of S2 and S3 with the market. The degree of dis-similarity is measured by the length of the bars on the bar chart which measures the deviation of the Mean of the cluster from the Mean of the entire data set (the market). S2 and S2 have relatively long bars as compared with S1 which is the nearest approximator of the market since it contains most of the S&P500 component stocks. However, S2 which contains the S stocks is not a 'tight' homogenous cluster. There are big empty spaces in between. On a SOM, this is an indication that the clustering is not 'strong' and thus S2 cannot be a good indcator of market bearishness. S3 which contains the L stocks is also not a distinctive, cohesive cluster. Other than an area at the top, S3 also has many empty spaces.

The overall outlook of the market for next week is thus ambivalent. It could be up or it could be down. But judging from the slightly better clustering characteristics of S3 as compared to S2, market outlook is slightly positive.

High Alpha stocks

To pick high Alpha stocks as defined above, we need to identify the area of S3 which is most densely populated with the ValuEngine screened stocks. This area would contain stocks which are very similar to each other in having the characteristics of S3 cluster. The darkened area is the selected area and the stocks occupying the nodes of the darkened area is in image 4. Note that some stocks e.g. EK appear in more than one ValuEngine screen. Our methodology selects the best of the best in the sense that the selected stocks could be from any of the ValuEngine screens. Our method also does not limit us to a fixed number of selected stocks. The sectors the high Alpha stocks belong to are: Finance, Consumer_Non_Durables, Transportation, Basic Industries, Technology and Consumer Services. The absent sectors are Consumer_Durables, Energy, Capital Goods and Public Utilities and Healthcare.

Quantifying The Elusive Alpha

2. Mean Statistics of the clusters
3. Clusters: ValuEngine Screens Against The Backdrop Of The S&P500

4. Cluster Statistics (Standard Deviations From Mean of Data Set)

5. Performance of Last Week's Market Neutral Strategy With Short S&P500 ETF [SH]

6. Performance of Last Week's Market Neutral Strategy With UltraShort S&P500 ETF [SDS]

First, to get it out of the way. The performance of last week's Market Neutral Strategy is shown in images 5 and 6 above. The strong performance of Technology sector stocks helped lift the portfolios up. If not for the Short on the S&P500, we would be in positive territory. But never, never regret it. The Short is a hedge and it should stay. Remember, this is a portfolio for Joe Smith and he cannot withstand stomach-churning volatility and steep drawdowns on his investment capital. Anyway, it's early days yet, as this is a monthly re-balance type of portfolio.

Now, for the serious business. Fund managers are always seeking Alpha or professing to have Alpha, but has anybody been able to measure Alpha? Alpha can be defined as the return in excess of some market index. Mathematically, if Beta measures the volatility of a stock against the market, and is measured by the slope of a line in a chart showing returns of a stock verus the market's return, then Alpha is where this line intercepts the Y axis of the chart. But although Beta can be meaningfully expressed in a quantitative way (e.g a stock with Beta of 1.5 will go up [or down] 1.5 times as much as the Index), Alpha cannot be similarly expressed. And thus it remains a nebulous concept for fund managers to foist on you.

In this post, I use a Self-Organizing Map to define Alpha as the ' degree of dis-similarity of a stock with the Index'. The implication of this is that the stock will outperform the Index when the Index is rising. Note that a stock can be dis-similar to the Index in a bad way too- which means that it will 'outperform' the Index negatively when the market is falling. However, Alpha is only concerned with 'good' dis-similarity.

This is a valid and practical definition and a SOM is a good way to measure dis-similarity in a holistic way. SOMs measure dis-similarity taking into account all the attributes [variables] of the stock against all the attributes of all other stocks at a certain point of time t. Degree of dis-similarity/similarity is demarcated by the boundaries of each of the self-organized clusters and a metric like Euclidean distance measures degree of similarity between nodes in the SOM's topology, thus making our Alpha measurable.

So much for the theory. Now for the real-life example. As in last week's post, ValuEngine's three screens: 20 stocks Standard (Valuation), 22 stocks Forecast (Growth), 20 stocks Star (Quality) are plotted against the backdrop of the S&P500 component stocks. The terminology: V=Valuation, G=Growth, Q=Quality. Sectors: B=Basic Industries, C=Capital Goods, D=Consumer Durables, ND=ConsumerNonDurables, S=Consumer Services, E=Energy, F=Finance, H=HealthCare, T=Technology, TP=Transportation, U=Public Utilities. Thus VU is a Valuation stock in Public Utilities sector. GE is a Growth stock in Energy etc.

Our objective is to meet two conditions (1) first to pinpoint the cluster(s) which is most dis-similar to the market. (2) and is the most densely populated. (3) Finally, we pick the stocks from the three ValuEngine screens which inhabit the densely populated zone of the cluster which is the most dis-similar to the Index. In image 3, the space on the SOM that is not labeled represents the SP500 stocks which are not output by the ValuEngine screens. This includes most of the area of cluster S1, S2 and parts of S3. In image 4, the length of the bars represents the degree of difference [standard deviation] from the Mean of the entire data set. Thus S5 cluster is greatly different from S1 and S2 clusters and in turn, S1 and S2 by their short bars are not so different from the Mean of the entire data set, which in our case essentially means the market as defined by the S&P500 . S5 is different from S1 in a good way too: The different colored arrows show that for example, S5 is more undervalued, yet has greater momentum and has a much higher 1-month forecast return %. Other traits of S5 include smaller market cap, higher volatility and higher Beta. These last three traits may or may not be good depending on the mode and mood of the market. If confidence returns [for the coming week at least], and risk appetite increases, being a smaller market cap stock with higher volatilty could be a plus in terms of outperforming the market. Neverthess, we are assured by the fact that our crop of stocks is a good mix of Valuation, Growth and Quality (see the V, G, and Q prefix in the image 1 stock list]. A good mix of sectors is also present: see the suffix of B, E, F, ND etc]. And don't forget that we can have a Short position in our portfolio via the SP500 reverse ETF SH or SDS. All in all, a good portfolio for poor Joe Smith. We won't track this portfolio as last week's portfolio one is only one week old and not due to be re-balanced for another three weeks. But at least, we have defined Alpha in a quantitative way.

Designing An Investment Strategy For Joe Smith The Babyboomer

1. Clusters
2. Cluster Means

3. Cluster Statistics

4. The Market Neutral Portfolio

The financial objectives of Joe Smith Babyboomer will be quite different from a younger person. Joe is still shell-shocked from the damage to his 401K that was caused by the financial crisis. Joe doesn't trust Wall Street anymore. All he wants is a stable return that is significantly more than the pittance that banks currently pay for his deposit. But Joe will not do risky investments that may endanger whatever sum of money he still has. Still, Joe thinks it would be nice if his portfolio is able to generate a little extra pocket money for that fly-fishing trip, a trip to Shanghai with the Missus, or that Gibson archtop guitar in the shop window that's staring at him and making him drool. Let's try to design an investment strategy for Joe that is simple and doesn't involve having Joe's blood pressure rocketing as volatility causes huge drawdown on his investment capital. This will be an experiment- an ongoing exercise- with records updated every week, so that Joe can see how well his investment is doing. First, the perimeters and parameters of this strategy:

1. It will be for an investment capital of $30000 of spare cash that Joe has.

2. Since it is impossible to predict the market, and Joe can't buy-and-hold for as long as Warren Buffet, it will be a market neutral strategy [MNS], with monthly re-balancing of portfolio. Another reason why Joe has lost faith in Buy and Hold is that he knows everything in the world changes faster than it used to, and today's Blue chip could be tomorrow's has-been.

3. For ease and practicality the Short side of the portfolio will be the ProShares Short S&P500 ETF. [ticker symbol SH] or the ProShares UltraShort S&P500 ETF [ticker symbol SDS]. *SDS goes up/down twice as much for every point of change in the S&P500.

4. The MNS portfolio will have equal $ weighting. 50 % Long and 50% Short and equally spread among the Longs.

5. Transaction cost will be $15 per trade. But for ease of calculation, we'll leave it out of our figures and you can include it in your own calculations.

6. The maximum number of Long stocks will be 10, but the actual number may be less if there are no stocks which meet the selection criteria.

7. The methodology for selection is to plot all 3 ValuEngine screens [Standard (for our purpose denoted as valuation mode)], Forecast [growth mode] and Star [Quality mode] on a Self-Organisng Map [SOM] and select those stocks which have good clustering properties whether they be from the Value, Growth or Quality screens. *clustering properties and SOM technology is discussed elsewhere throughout this Blog. So here we go:

1. In image 1, the stocks from ValuEngine's Value [V], Growth [G] and Quality [Q] screens are plotted over a SOM that has been constructed from the 500 stocks of the S&P500.

2. In image 2 and 3 we select cluster 2 [Red] to be the cluster with the most attractive clustering characteristics. The V, G and Q labels denoting nodes that are occupied by the screened stocks are packed densely together unlike S 1 and S3. Cluster statistics also show Undervaluation but with Momentum [12-m return %] The current market situation is risk-averse and places emphasis on valuation and quality.

3. Image 4 shows the final list of stocks selected, the number of shares of each, and their last Closing price. Our total investment is Long 14864 and Short 14988 if SH is used and 14979 if SDS is used. We will monitor the results to see if SH or SDS is a better choice for the Short side.
Lastly, please remember that this is an experiment, the results of which will be used to tune the SOM. Ultimately it is hoped that a good newsletter can be written that meets the financial objectives we discussed.

A Market Neutral Strategy Using Self-Organizing Map

1. Self-Organizing Map of ValuEngine Engine-Ratings [Composite Ranking] of SP500 Component Stocks.

2. Average Composite Ranking Score of Clusters

3. Self-Organized Clusters of SP500 Component Stocks Labeled WIth Their ValuEngine Engine Ratings Composite Ranking Score.

4. SOM Top 10 Overall Best Stocks

5. ValuEngine Top 10 Overall Best Stocks.

6. SOM Bottom 10 Overall Worst Stocks

7. ValuEngine Bottom 10 Overall Worst Stocks

The aim of a market neutral strategy is to go Long on the best stocks, and Short on the worst stocks. It is a good strategy when markets are very unpredictable, as the gains come not from market direction, but the 'spread' between best and worst. In theory, the best stocks will outperform the market positively when the market rises, and outperform the market negatively when the market falls. But for a quantitative strategy, 'Best' and 'Worst" has to be quantitatively defined. ValuEngine's approach is to give stocks their unique Engine Ratings;1-Engine to 5-Engine with 5 being the most desirable. "Engine Ratings: ValuEngine's proprietary Engine Rating system provides an overall assessment of a stock's attractiveness. It combines valuation, risk-return tradeoff, momentum, market capitalization and forecasted future return data into an easy to understand and actionable format." And as can be seen from the top image, Engine Ratings have proven to be quite a good indicator of stock performance. The Engine Ratings in turn are based on a Composite Ranking score system. Using ValuEngine's Engine Rating system, we could have a Long portfolio comprising, say, the x number best stocks and a Short portfolio of the y number worst stocks, where x=y.

In this post, we use a SOM to find the 10 best stocks and the 10 worst stocks, and we compare the list of stocks that the SOM identifies, with ValuEngine's list. A SOM offers a more holistic way of looking at 'best' and 'worst'. It simultaneously takes into account all the relevant model variables and their non-linear inter-relationship with each other. Using the self-clustering properties of the SOM, and Euclidean distance on the SOM to measure degrees of similarity, it was hypothetised that the SOM's list of 'best' and worst' would be different from ValuEngine's list.

Some of the assumptions we used in this exercise are:

1.Our Universe of stocks comprises only the component stocks of the SP500 Index because (a) we want stocks with good liquidity and reasonable size [market cap].(b) Having only Index component stocks will also lower the volatility of the portfolio, making our strategy more suitable for ordinary folks who don't indulge in high frequency algorithmic trading.

2. An arbitrary 10 stocks per portfolio is used, because a 20-stock portfolio is manageable and affordable for the average investor with an investment capital of $20000 and yet offers suffiicent diversification.

3. The portfolios are equal weighted in terms of $, as I feel that trying to give weightings for each stock is a futile exercise in optimization and is a case of false precision. Optimization can be done, but I prefer the use of Genetic Algorithms rather than raditional statistical methods for optimization.

A SOM was constructed and the results are summarized below:

1. Image 1 shows the ValuEngine composite ranking scores of SP500 stocks as a heat map where areas of higher scores tend towards the color Red. [see scale below map].

2. Image 2 shows the SOM has divided the stocks into seven distinct clusters based on their degree of similarity in overall characteristics, with algotithmic thresholds to determine cluster boundaries.

3. Image 3 shows that cluster S4 [Green] has the highest average composite score of 713 and cluster S6 [Pink] has the lowest average composite score of 525. [Cluster S7 actually has the highest score of 884 but is a one-stock cluster [Eastman Kodak] and has very volatile characteristics. Image 4 also shows that the composite scores in each cluster do not run'consecutively' i.e. you may find a score of 656 in one cluster, while the next score of 657 may be in another cluster. This is the main difference between ValuEngine and the SOM.

Image 4 and 5 compare the SOMs 10 best stocks with ValuEngine's 10 best stocks. Note the differences and how the SOM puts more emphasis on the big cap technology sector e.g. Intel, Apple, Cisco and on raw materials e.g. Noble, Corning, Freeport-Mcmoran.

Image 6 and 7 compare the SOM's 10 worst stocks with ValuEngines 10 worst stocks. Goodyear, Office Depot, SprintNextel and Tesoro appear in both list but other stocks are not in both list. The 'worst' list seems to be more stock-specific and less sector-specific.

Last thoughts: Each week, or even each day, a stock's Engine rating may change. Those near the boundary between discrete Engine ratings may bounce back and forth between two ratings. The SOM has tools for displaying the finer details of the properties of a cluster and the stocks within it e.g it's centroid, its neighborhood, slope of the cluster boundaries, contour, and U-matrix. using these tools will further enhance knowledge of a cluster's characteristics and enable us to better select the stocks for the Long/Short portfolios. But for that, I will need to consult with Dr. Kranner, the inventor of Viscovery and my mentor.

This MNS strategy will not yield as good a return as ValuEngine's current market neutral strategy using the forecast model [as in their MNS newsletter by Steve Hach]. But it may be a more suitable strategy for conservative investors with steady longer term returns in mind, and less frequency in portfolio re-balance.