Using pattern recognition techniques to determine profitability on financial markets.

Ionescu, Stefan Alexandru ; Gheorghe, Camelia Monica ; Zaheu, Ioana 等

1. INTRODUCTION

Analysis of the financial performance of companies, both from the accounting and business development from the perspective on the capital market, is for financial investors, but also strategic, a permanent fixture, depending on the accuracy of this analysis, ultimately, the structure of portfolios they hold and the yields obtained.

One of the methods used to formation of diversified portfolios is considering inclusion in the same portfolio of companies coming from various sectors of activity. The logic of this method departs from the recognition that the prices of financial securities on the capital market and yields generated by them are influenced by a number of risk factors, which affect not the same as all companies.

We propose in this work, to make a first step towards the construction of standards sector by seeking to achieve a grouping of companies listed on the Bucharest Stock Exchange, depending on their financial performance. Subsequently, we intend to notice the evolution over time of the companies analysed by reference to such groups made up. Dozen of indicators are used to make predictions, but very little work consistently in time. Investing in companies that have done well in a certain period of time can make a profit higher than the use of indicators classics as S&P500. (Damodaran, 2003)

2. PREVIOUS RESEARCH

Sharpe is the one that opened the way to theoretical developments in the centre who share these risk factors specific to the firm and variable following the formation of portfolios, on the one hand, and systematic, affecting all companies on the same capital market (Sharpe, 1963).The latter are considered to be unvariable nationally, but variable at international level.

King was the first which showed a common movement of stock prices for companies in the same industry, which is stronger than the market effect (King, 1966). Using cluster analysis for grouping companies according to their performance can be found from the work produced by Jensen (1971), Meyers (1973) and Livingston (1977), who tries to identify the natural groupings of U.S. companies along sectors of activity from which they originate. All the three authors mentioned using a relatively small set of companies and industries, and their results are satisfactory (Livingston, 1977)

3. METHODOLOGY OF CALCULATION

Applying this method to the Romanian capital market seems at first glance, a fairly easy task. Nevertheless, reduced the number of companies traded on the Bucharest Stock Exchange (65 currently, those two classes of stock exchange) and relatively low financial information regarding the task they are much more difficult. Too, believe that the distribution companies in terms of trade and often the presence of a very small number of companies in a particular sector raises additional application of such methods.

A method of grouping or classification of similar entities according to a series of features that are considered each entity is represented by cluster analysis. In this type of analysis, which belongs to the group multivariety analysis methods, the problem is that the classification entities in a number of groups, called clusters, where the entities in the clusters are more homogeneous than entities in different clusters. This sort of analysis in the discovery of natural cluster according to a specific criterion internally, without knowing in advance party entities considered to these clusters. Criterion internally that is used in cluster analysis is based on a measure of similarity or approach between the entities concerned.

Applying the cluster analysis raises, in general, two issues: the first is that the methods of measuring the distance between the entities concerned. We note entities by i and j, given the coordinates ([X.sub.1i], [X.sub.2i]), respectively ([X.sub.1j], [X.sub.2j]), depending on the [X.sub.1] and [X.sub.2] considered. A survey i is considered to be closest (more similar) of j remark then remark k if D (i, j) <D (i, k). The variables are standardized before applying cluster analysis. The second problem concerns the methods of formation of clusters, they may be divided into two broad categories, depending on assumptions which are based on the nature and results:

* hierarchical clustering methods, including clustering methods through aggregation and clustering methods by division for each of the procedures being used to prepare clusters, among which are the most popular single linkage method, complete linkage method, or Ward's method.

* iterative methods of clustering also include a series of algorithms, of which the most popular and use the algorithm are the K-means, k-medoids algoithm, CLARA algorithm, or Fuzzy c-means clustering. As iterative method we use nonhierarchical grouping, known mostly as k-means clustering is a priori principle fixing the number of groups to be obtained and grouping objects (cases) so as to steal the biggest differences between media groups (statistically tested on the basis of an ANOVA type procedure).

After finding a corresponding way of separating [ohm] set elements on the prediction classes found, the main task for the discriminante analysis is to decide regarding membership of the n classes of new objects from the multitude [ohm] or to make predictions concerning the affiliation of these objects.

4. DATA USED 2002-2006

Analysis of this work is carried out for the period 2002-2006, for a number of 45 companies listed on the Bucharest Stock Exchange (BSE). Although the number of companies listed on the BSE varied during the aforesaid from 65 in 2002 to 58 in 2006, we selected only those who were listed permanently, without being withdrawn from trading in that period. In each case we used eight variables, defined as: liquidity ratios (current ratio); solvability ratios (debt-to-equity ratio, total debt ratio); efficiency ratios (total assets turnover); profitability ratios (return on assets, return on equity, net profit margin); indicators of market value earnings per share)

5. RESULTS OBTAINED

Applying first time the simple aggregation method, we see that the number of clusters can be identified for all those five years review is low, ranging between two and four. For all the years, it worth to underline the grouping of the "Financial Investment Companies" (FIC) in a separate cluster, which highlights the characteristics of this type of investment on the market. In this division cluster of companies in those four years, the identification of sectorial groups is not possible, in the same cluster can be found companies from different sectors: machine tools, chemical industry, textiles, etc.

Applying the Ward's amalgamation method leads to the differentiation of the same five-FIC, again they form a cluster for each of the years taken into account. Since the method uses standard deviation as an instrument for measuring the distance between the entities, the new clusters are more numerous and include a fewer number of companies. Nevertheless, can not speak about the formation of cluster at sectorial level, and even less than their stability over time, companies "moving" from a cluster to another over the period studied. This suggests that the Romanian capital market performance and financial are fluctuant, and, of course, we can not talk about the existence of sectorial performance features companies. Figure 1 (processed in Statistica 7.0) illustrates the dendrogram for the year 2006, the other may not be submitted to the lack of space available.

To confirm the results obtained by applying hierarchical amalgamation methods, we also used the k-means algorithm, while each of the years considered a number of three clusters, which included companies with different performance characteristics. The three clusters formed for those four years shows different attributes, which leads us back to conclusion that the Romanian market can not talk about stability in time of financial performance of companies Thus, for the year 2006 (see Figure 2--processed in Statistica 7.0) we see the formation of a cluster characterized by low solvency, low profitability and low yields on the BSE, another one characterized by the average performance for all aspects of activity and another one with high profitability, accompanied by low efficiency. In the last cluster we find again only FICs.

[FIGURE 1 OMITTED]

[FIGURE 2 OMITTED]

6. CONCLUSIONS

Pseudo-industry groups formed from cluster analysis do not match the performance of industry classifications on an out-of-sample basis (Chan et. al, 2007). The K-means algorithm has provided us an interesting and correct classification of the 45 firms, without being able to establish ranks according to the industry. The only cluster that has gathered companies in the same sector was one that contained FIC. This was somewhat predictable because these companies are investing similar, in the same areas or even in the same company. Because BSE is very young compared with markets in America and Western Europe, these differences appear more overtoned in Romania. With discriminant analysis we calculated the classification of functions that may help in future predictions.

7. WHAT TO DO

The research presented will be continued by including other amalgamation methods, on the one hand, and using discriminant analysis in order to better distinguish between such groups formed. At the same time, identifying industry clusters of companies in the Romanian capital market represents a first step towards building a system of performance rating.

Extending the base of indicators and determining the cluster for each year in part would be useful for an analysis of evolution in the dynamics, an observation of the clusters modification from one year to another and for explanations of the migration of objects from one cluster to another.

8. REFERENCES

Chan, L.K.C.; Lakonishok, J. & Swaminathan B. (2007) "Industry classifications and the comovement of stock returns" Financial Analysts Journal, vol.63, no.6, November/December, 2007, pp56-70, ISSN : 0015-198X

Damodaran, Aswath. (2003) "Investment Philosophies: Successful Strategies and the Investors Who Made Them Work", John Wiley and Sons, ISBN : 0-471-34503-2, NY

King, B.F. (1966) "Market and Industry Factors in Stock Price Behavior", Journal of Business, vol. 39, no.1, January, 1966, pp139-190, ISSN: 0021-9398

Livingston, M. (1977): "Industry Movements of Common Stocks", Journal of Finance, vol.32, no.3, June 1977, pp861-874, , ISSN: 0022-1082

Sharpe, W. F. (1963) "A Simplified Model for Portfolio Analysis" Management Science, vol 9, no.1, January, 1963, pp277-293, ISSN : 0025-1909