The purpose of the following work is to create an investment strategy operating on the historical data from the Polish stock exchange. Such a system may support decisions making by the human during investments on a stock market.

The following problem will be solved by the usage of unsupervised clustering algorithm bases on the neural networks - Self-organizing Kohonen Maps. Its adventage comparing to the supervised algorithms, is that the unsupervised learning may provide better results, as no action is defined for the particular situation on the market. Instead, the most similar configurations of shares’ prices may be identified and labeled.

Because the nature of each stock market's company is different, it is not possible to build a single model for all of them. Therefore, the separate model will be trained for each company, as the particular enterprise is unique and its behavior in time does not depend on the other companies.

The whole process consistes of two main phases which are: data preparation and modeling. During the first phase a set of parameters describing a company will be created in a form of timeseries dataset. Those parameters will be callculated from the available coefficients described in the next section. In the next phase, which is modeling, the similar groups of vectors (clusters) will be created, basing on dataset features in particular timestamp. Then, each cluster will be marked as 'buy', 'sell' or 'hold'. The labels will be assigned to the clusters according to their mean return value. At the picture bellow we can see a good moments for selling or buying a stock:

According to the tresholds due to which we set clusters labels, a strategy can be more or less agresive. Tresholds don't need to be symetric as they reflect the character of the investor and his inclination towards taking the risk.

To make our model working correctly we need to operate on the stocks which has significant fluctuations in their prices. In an opposite case, it is highly probable that we would mark every moment in the timeseries as 'hold'.

The example companies which has been chosen comes from stock exchanges like: HKG (Hong-Kong), NASDAQ (USA) and WSE (Polish). The market summary of each company can be seen at the pictures above:

The self-organizing map (SOM) is an automatic data-analysis method. It is widely applied to clustering problems and data exploration in industry, finance, natural sciences, and linguistics. The SOM is related to the classical vector quantization (VQ), which is used extensively in digital signal processing and transmission. Like in VQ, the SOM represents a distribution of input data items using a finite set of models. In the SOM, however, these models are automatically associated with the nodes of a regular (usually two-dimensional) grid in an orderly fashion such that more similar models become automatically associated with nodes that are adjacent in the grid, whereas less similar models are situated farther away from each other in the grid. This organization, a kind of similarity diagram of the models, makes it possible to obtain an insight into the topographic relationships of data, especially of high-dimensional data items. If the data items belong to certain predetermined classes, the models (and the nodes) can be calibrated according to these classes. An unknown input item is then classified according to that node, the model of which is most similar with it in some metric used in the construction of the SOM.

Input data is calculated using extract_features' method read_and_extract_features. The method accepts the name of file received from the stock market, and writes output to the file named as the second parameter. The output of the method contains following fields:

• Index
• Date
• Close/Last - the closing price
• High - the highest price during that day
• Low - the lowest price during the day
• f1 - logarithmic RoR (natural algorithm) of ratio between current and previous day
• f2 - logarithmic RoR (natural algorithm) of ratio previous day and 2 days before
• f3 - logarithmic RoR (natural algorithm) of ratio of price from 2 and 3 days before
• f4 - logarithmic RoR (natural algorithm) of ratio of price from 3 and 4 days before
• f5 - RoR of ratio of closing price from current day and 2 days before
• f6 - RoR of ratio of closing price from 1 and 3 days before
• f7 - gradient of 5-day price trend (starting from current day)
• f8 - gradient of 5-day price trend (starting from 5 days ago)
• f9 - gradient of 10-day price trend (starting from current day)
• f10 - difference between first and second feature
• f11 - difference between first and third feature
• f12 - Relative Strength Index (RSI)
• f13 - Moving Average Convergence/Divergence (MACD) for 12 and 26 days
• f14 - difference between current day's closing price and Moving Average for 12 days
• f15 - 14-day Rate of Change (RoC)
• f16 - Commodity Channel Index (CCI) for last 20 days