CLASSIFY STOCK USING KOHONEN SELF ORGANIZING MAP (SOM)

Size: px

Start display at page:

Download "CLASSIFY STOCK USING KOHONEN SELF ORGANIZING MAP (SOM)"

Barbara Shaw
6 years ago
Views:

1 CLASSIFY STOCK USING KOHONEN SELF ORGANIZING MAP (SOM) Copyright Arryex, 2012, Independent Investor Objective: Stock selection based on fundamental data classification Tools: Neuroshell2 ( Ward Systems Group, Inc) and Matlab, ( Mathworks) Introduction The Kohonen Self Organizing Map network is a type of unsupervised network, which has the ability to learn without being shown correct outputs in sample patterns. These networks are able to separate or mapping data into a specified number of categories. There are only two layers: an input layer and an output layer which has one neuron for each possible output category. The maximum number of output can be categorized when selecting Network Architecture Options by entering that number of neurons in the output layer. The training patterns are presented to the input layer, then propagated to the output layer and evaluated. One output neuron is the "winner". The network weights are adjusted during training. This process is repeated for all patterns for a number of epochs chosen in advance. The architecture is different with Back propagation which having hidden layers.

2 Stock Fundamental Data All companies who listed in the stock exchange should inform to public all information related to the company profile, financial report, corporate action, etc. This information should be available in the news paper, online broker trading interface, government website, etc. Here is the sample information Using above fundamental data from several companies who has similar industrial type, we implement Kohonen SOM to select which best company that may provide good investment return. JASICA CLASSIFICATION Based on JASICA (Jakarta Stock Exchange Classification Index), there are 9 Sectors that grouping companies based on its industrial type: 1. Agricultural, 6 subsectors 2. Mining, 5 subsectors 3. Basic Industry and Chemical, 9 subsectors 4. Miscellaneous Industry, 7 subsectors 5. Consumer Good Industry, 7 subsectors 6. Property, real Estate and Building Construction, 3 subsectors 7. Infrastructure, Utilities and transportation, 6 subsectors 8. Finance, 7 subsectors

3 9. Trade Service and Investment, 9 subsectors. Within this paper, author would like to classify stock within Sector 9, and Subsector 91 Wholesale (Durable & Non Durable Goods). The companies within this group are: AIMS AKRA ASIA CLPI DSSA EPMT FISH GREN HEXA INTA INTD KARK KONI LTLS MDRN MICE OKAS SDPC SING SUGI TGKA TIRA TMPI TRIL TURI UNTR WAPO WICO PT Akbar Indo Makmur Stimec Tbk PT AKR Corporindo Tbk PT Asia Natural Reseources Tbk PT Colorpak Indonesia Tbk PT Dian Swastatika Sentosa Tbk PT Enseval Putera Megatrading Tbk PT FKS Multi Agro Tbk PT Evergreen Invesco Tbk PT Hexindo Adiperkasa Tbk PT Intraco Penta Tbk PT Inter Delta Tbk PT Dayaindo Resources International Tbk PT Perdana Bangun Pusaka Tbk PT Lautan Luas Tbk PT Modern Internasional Tbk PT Multi Indocitra Tbk PT Ancora Indonesia Resources Tbk PT Millennium Pharmacon International Tbk PT Singleterra Tbk (Sebelumnya: Singer Indonesia Tbk) PT Sugi Samapersada Tbk PT Tigaraksa Satria Tbk PT Tira Austenite Tbk PT AGIS Tbk PT Triwira Insanlestari Tbk PT Tunas Ridean Tbk PT United Tractor Tbk PT Wahana Phonix Mandiri Tbk PT Wicaksana Overseas International Tbk Here are the latest available company data (3M 2012)

4 Companies Fundamental Data (JASICA, subsector 91, Wholesale) Data Normalization All above data are normalized between 0 to 1

Implementing Kohonen in Neuroshell 2 To classify data fundamental, we will classify all companies into 3 groups based on each data Sales Growth, Gross Profit Growth,

Neural Network Configuration: We have 2 layers Slab 1 and 2 Slab 1 has only 1 neuron Slab 2 has 3 neurons Neural Network Configuration One Dimensional Classification Result

the lowest value become the winner for that stock. Example for WAPO has outarray[0]= 0.269866224, outarray[1]= 0.43551024, or outarray[2]= 0.5840374.

5 Implementing Kohonen in Neuroshell 2 To classify data fundamental, we will classify all companies into 3 groups based on each data Sales Growth, Gross Profit Growth, Operating Profit Margin, Net Profit Margin, Market Capitalization, PER, BVPS, PBV, ROA, ROE, EV/EBITDA and Debt/Equity. Neural Network Configuration: We have 2 layers Slab 1 and 2 Slab 1 has only 1 neuron Slab 2 has 3 neurons Neural Network Configuration One Dimensional Classification Result Based on Sales Growth Kohonen Classification Result and Generated C code Based on Sales Growth Each stock will have three values (outarray[0], outarray[1], or outarray[2]), the lowest value become the winner for that stock. Example for WAPO has outarray[0]= , outarray[1]= , or outarray[2]= The lowest value for WAPO is outarray[0] then it classified as Group 1. If the minimum value is outarray[1] then it classified as Group 2, and if the lowest value is outarray[2] then it classified as Group3. From 28 stocks, we classified in 3 groups then we have

Kohonen Classification Result Based on Sales Growth Refer to above classification result, Kohonen mapped the Sales Growth data into the three ranges: 1. Sales Growth 1 to 0.

6 Kohonen Classification Result Based on Sales Growth Refer to above classification result, Kohonen mapped the Sales Growth data into the three ranges: 1. Sales Growth 1 to 0.01, as Class 3, considered as negative sales growth. 2. Sales Growth 0.03 to 0.42, as Class 2, considered as positive small sales growth. 3. Sales Growth 0.75 to 2.56, as Class 1, considered as positive big sales growth. A good company should have a positive sales growth. By referring to this classification, we can select a stock with medium or big sales growth (FISH and AIMS) and avoid stock with negative sales growth.

7 Two dimensional Classification Based On Sales Growth and ROE 2 Dimensional Classification Based on ROE and Sales Growth

A good company should have a positive sales growth and positive ROE.

growth and positive ROE (AIMS, FISH, HEA, INTD, SING and WAPO) avoid stock with

data to classify We get 6 stocks in Group 1, 6 stocks in Group 2, and 16 stocks in

8 A good company should have a positive sales growth and positive ROE. By referring to this classification, we can select a stock with medium or big sales growth and positive ROE (AIMS, FISH, HEA, INTD, SING and WAPO) avoid stock with small/negative sales growth and ROE Multidimensional Classification Here we use all data to classify We get 6 stocks in Group 1, 6 stocks in Group 2, and 16 stocks in Group 3. The first 6 stock is classified with all positif data and can be considered as a good stock out of 28 stocks.

Implementing Kohonen in Matlab Similar data the used in Neuroshell 2, we will classify using Matlab

Margin, Net Profit Margin, Market Capitalization, PER, BVPS, PBV, ROA, ROE, EV/EBITDA and

To establish classification we need to make a Matlab script that consist of: Data set, 12 rows x 29

9 Implementing Kohonen in Matlab Similar data the used in Neuroshell 2, we will classify using Matlab all companies into 3 groups based on each data Sales Growth, Gross Profit Growth, Operating Profit Margin, Net Profit Margin, Market Capitalization, PER, BVPS, PBV, ROA, ROE, EV/EBITDA and Debt/Equity. To establish classification we need to make a Matlab script that consist of: Data set, 12 rows x 29 columns (in xls, csv or mat file) Importing Data script (from xls or csv) Culstering using Self Organizing Map (selforgmap) Training the net View the net and classification result Data set Matlab Result

The cluster_index variable is showing the class of each input As shown above Hits and

10 The cluster_index variable is showing the class of each input As shown above Hits and cluster_index, we have 25 stocks are classified in Group 1, 3 stocks in Group 2 (AKRA, DSSA and HEA) and 1 stock in Group 3 (UNTR). CONCLUSSION Using Kohonen Self Organizing Map, we can classify all stocks into several groups based on stock fundamental data. By knowing which group a stock been classified we can avoid to buy a stock which having negative or small values on fundamental data. On above sample using Kohonen Neuroshell 2, we can select 6 good stocks out of 28 stocks. While using Matlab, we can select 3 or 4 good stocks. Further study can be used to improve classification, example by adding actual return of the stock value over certain period, actual cash deviden, etc.

LAMPIRAN 1 Perusahaan Sampel Penelitian Tahun

LAMPIRAN 1 Perusahaan Sampel Penelitian Tahun 2008-2010 No. Perusahaan Kode Tahun 1. PT. Gozco Plantation Tbk. GZCO 2008 2. PT. Indika Energy Tbk. INDY 2008 3. PT. Adhi Karya (Persero) Tbk. ADHI 2008 4.