Price Pattern Detection using Finite State Machines with Fuzzy Transitions

Transcription

1 Price Pattern Detection using Finite State Machines with Fuzzy Transitions Kraimon Maneesilp Science and Technology Faculty Rajamangala University of Technology Thanyaburi Pathumthani, Thailand Chutima Prasartkaew Science and Technology Faculty Rajamangala University of Technology Thanyaburi Pathumthani, Thailand Abstract A price pattern is one popular technique of investment technical analysis. It is a strong signal to predict the direction of market prices and will show the good entry-exit point for investors. However, it is loosely pattern. So, it is difficult to cover detected all patterns by computer programming. This article presents a methodology through applying the Finite State Machine with Fuzzy Transitions (FSM-FT) to recognize the price pattern. The result is a flexibility increasing on detected pattern. Moreover, we can found some pattern, that not easy to observe by expert person. Keywords- pattern detection; price pattern; Investment decision system; finite state machine with fuzzy transition; FSM- FT. I. INTRODUCTION The price patterns are the technical analysis for that very high reliable when compared to other tools. They have high efficient and favors to forecasting the movement of market prices and applies to detect the signal for entry position because it can be described by the market psychology[1]. However, the detection of the Price Patterns requires the expertise of the investor to find it. Because, their definition is very broad loose. Moreover, the patterns will be changed when we change the period of chart. The objective of investor to forecasting the trends of prices is finding the entry point, direction of holding position and exit point to take profit and stop loss. It has two major methods to forecast it, fundamental and technical analysis. The fundamental is analyzed on the economic information and the technical is analyzed by prices movement[2]. The research on prices forecasting is often to work on technical analysis, because it is easier to understand. The technical analysis has three types[3]. The first is using line and graph it call the technical indicator. It base on mathematic methods, for example, RSI, MACD, StdDev[3]. The second is observed from price movement pattern like U shape, V shape, double top, double bottom and the head and shoulders [1]. The last one is observed the chart pattern like the candlestick and the point and figure chart. The study of methodology to forecasting, they have any methods, the example are neural network[4][5], hidden Markov model [6], support vector machines[7], hybrid ARIMA and neural network[8], hybrid ARIMA and support vector machines[9], hybrid rough sets theory and genetic algorithms[10] and fusion the hidden Markov model, neural network and genetic algorithm to forecast it[11]. Model that has all of above is based on the shape memory. However, the price patterns are composite by prices peak sequence and movement of peak. It is loosely patterns. So, it does not easy to detected by normal type of pattern recognition. Because, the price pattern can represents by the pattern of sequence of prices peak. It shows that, we can design the Price Pattern detection system using finite state machine with the series of prices peak. However, the conditions to change states are loosely definition. So, the proper methodology for solving this problem is a finite state machine with fuzzy transition. Finite state machine with fuzzy transition (FSM-FT) is a finite state machine that applies the fuzzy logic to the decision to state changed in the machine[12]. The condition of the state changed is represented to fuzzy relationship, and the machine will stay only one state in the quantum time. It difference to fuzzy state machine (FuSM). The FuSM can visit more than one state in the quantum times. This paper represents the methodology to apply the FSM-FT to detect price pattern, which refer to the head and shoulders price pattern[1] because it is favors pattern using detect entry signal by investor. The experiment is applied on the five minute times period of EURO-US$ exchange rate from Feb 3, 2014 to Jun 30, The result compares the number of the head and shoulders when find by expert and this price pattern detection system, the number of the success pattern and false pattern. Moreover, the result shows the advantage of the system, that it can detect the pattern which difficult to observe by expert on current times period, but can found when using the smaller period. For example, it can establish a pattern at 5 minutes times period but the export cannot observe it, but they can observe it when change times period to 1 minute. II. HEAD AND SHOULDERS PATTERN DETECTION DESIGN The first step of the article describes to the head and shoulders price pattern characteristic. The patterns of head and shoulders have two patterns, bearish and bullish pattern. The bearish pattern look like the group of mountains, it has three peaks, highest peak is the second peak and after third peak, the price will come down. In the other hand, the bullish has invested pattern of the bearish pattern.

2 Figure 1. Head and shoulders price pattern Figure 1 show the bearish (A) and bullish (B) head and shoulders price patterns. It has six states compose the pattern. The states are peak of the prices that alternate between a high peak and low-peak call S 1, S 2, S 3,, S 6. In the bearish pattern, if S 1 is lower than S 2, S 2 is higher than S 3 and S 3 higher than S 1 ; we call S 2 is the first shoulder. After that, if S 4 high over S 2 and S 4 lower than S 2 ; we call S 4 is the head. Moreover, if S 6 is more over S 5 but lower than S 4 : we call S 6 is last shoulder. When the pattern has completed by one head and two shoulders, the head and shoulders are presented. It is a signal for an open position and prices will be down. The bullish pattern is similar to the bearish pattern but, it has the direction of peaks are opposite to the bearish pattern. The entry signals will be created when the system found the last shoulder. However, the head and shoulders are not exact patterns. Sometime, prices do not move follow by the pattern direction. In this case, the price will be move to near the line drawn through points S 3 and S 5, we call this line is the neckline. If price is cross the neckline, it is confirmed pattern signal. In the other hand, if prices direction is reflex when it move nearly neckline, it is a cancel signal, the investors will be close position to cut loss. Figure 2 show the prices movement of this case. Figure 2. Head and shoulders cancel signal From the foregoing, this pattern detection should be three output signal types. First, the pattern found signal, second and third are confirmed and cancel signal, respectively. The pattern found signal is telling to should entry at this time. The types bearish or bullish are told to short or long position which we should open. The confirm signal is telling to should hold this position for take profit and cancel signal is telling to should close this position for cut loss. From pattern definition and state output, we can design the state linguistics state machine shows by figure 3 and figure 4. Figure 3. Bearish linguistics state machine Figure 3 show the bearish linguistics state machine, we can rewrite to decision statement follow this. - If state is S 0, and a new bottom peak is coming; state changed to S 1, other then, stay in S 0. - If state is S 1, the new top peak is coming and the price at new peak much higher than the price of current peak; state changed to S 2, other then, stay in S 1. - If state is S 2, a new bottom peak is coming and the price at new peak lower than the price of current peak; state changed to S 3, other then, stay on S 2. - If state is S 3, the new top peak is coming, the price at new peak higher than the price of current peak and the price at new peak little higher than the price at last peak; state changed to S 4, other then, stay return to S 2. - If state is S 4, a new bottom peak is coming and the price at new peak lower than the price of the last peak; state changed to S 5, else-if a price at new peak little lower price at current peak and little higher than the price at last peak; state return to S 1, other then, stay on S 4. - If state is S 5, the new top peak is coming, the price at new peak higher than the price of current peak and the price at new peak little lower than the price at last peak; state changed to S 6 and system send the head and shoulders bearish signal to output, else-if a price at new peak higher than the price at last peak; state return to S 4, other then, stay on S 5. - if prices are move to lower neckline; the system send the confirm signal to output and return to S 0, else-if prices are move to higher neckline; the system sends the cancel signal to output and return to S 0, other then, stay on S 6.

3 The bullish linguistics state machine is similar the bearish linguistics state machine, but the condition to change state is opposite. It shows by figure 4. Figure 5 and 6 show an example. They are the bearish of head and shoulders price patterns on EURO-US$ exchange rate. Figure 5 occurs at Jul 17, 2014 from 01:20 to 02:30. It has S 1 price at , S 2 price at and S 3 price at The standard deviation is calculated by 200 times sampling. The value is in the range of to In this case, the little higher is S 3 - S 2 = and much higher is S 3 - S 1 = Figure 6 occurs at Jul 7, 2014 from 18:35 to 18:35, S 1 price at , S 2 price at , S 3 price at and the standard deviation is in the range of to In this case, the little higher is S 3 - S 2 = and much higher is S 3 - S 1 = It can observe that, the value of little higher is high when the deviation higher. That it has a problem when define the fuzzy relationship by never concern to the standard deviation. So, the fuzzy relationship in this case should by relate to price movement index ( ) can find by this equation. Figure 4. Bullish linguistics state machine The next step is to define the value of linguistics in state changed condition to fuzzy membership function. In this case have six linguistics values: higher, lower, little higher, little lower, much higher and much lower. To see that, the linguistics values of this case are comparison values between two points. Over that, the values have changed when standard deviation of prices has changed. where price is the difference of prices, S N is standard deviation[13] can find by (1) (2) where N is the number of observed values, {x 1, x 2,., x N } are the observed values, x is the mean value of these observations. This paper using N is 200. The fuzzy membership functions of state transitions present by figure 7. Figure 5. Bearish pattern at 17/07/ :20-02:30 Figure 7. Fuzzy membership functions of state transitions Figure 6. Bearish pattern at 07/07/ :35-19:35 III. PEAK FINDER UNIT The pattern detection system want to the preprocessing to get the peak of price, call the peak finder. It will be find all peaks in prices chart to build a list of peaks and send it to pattern detection.

4 This paper is setting the parameter of MACD and standard deviation follows, - MACD: fast MA=4, slowma=8, Signal=9 - Standard deviation period = 200. Figure 8. System block diagram The input of system is the prices series in the prices chart. The prices chart is capture prices information by the range on the times and represent in the bar chart or candlestick chart. So, it does not represent only one value per bar. It has four values are highest price, lowest price, open price and close price. The function of peak finder unit is find the top of highest price of all bars in the up- trend and bottom of lowest price of all bars in the down- trend, after that, and price and time information of peak to the peak list. The up-trend and down-trend considers from MACD [2] and standard deviation [2][13]. If MACD is value higher than MACD signal and standard deviation, the trend is up. If MACD value is lower than MACD signal and negative value of standard deviation, the trend is down. If absolute of MACD value is not over standard deviation, the trend is sideway. The functional of peak finder unit is described by flow chart in figure 9. IV. EXPERIMENTAL The experiment focuses to number of head and shoulders price pattern detected by the purpose method, compare to finding this price pattern by expert persons. The expert persons in this experiment are three investors who have much experience on the price pattern. The point of price pattern will be needed to get the consensus from all persons. The exchange data is using EURO-US$ exchange rate from Feb 3, 2014 to Jun 30, The system is developed on the MetaTrader 4 platform [14] and exchange range data is downloading from its historical data. The result is show with table I. TABLE I. NUMBER OF HEAD AND SHOULDERS PRICE PATTERN Month Feb Mar Apr May Jun found by Total the purpose confirm technique cancel found by expert persons Total confirm cancel The purpose technique is found pattern equal the expert persons in March and May. In February, the purpose can found pattern more than expert persons is 2 point. In April and June, the purpose can found pattern more than expert persons is 1 point. However, the cancel pattern that found by purpose technique and expert persons is equality. That show, the patters that found by purpose technique and do not found by expert persons are the confirm pattern. We can plot a graph for represent the number of found patterns by purpose technique and expert persons in figure 10. Figure 9. Flow chart of peak finder unit Figure 10. Number of pattern detected by purpose technique and detected by expert persons.

5 any pattern that similar character by adapted the purpose technique to solve this problem. REFERENCES (a) (c) Figure 11. Price chart and output of price pattern detection The reason of, why the purpose technique can found some pattern but the expert persons cannot found, described by figure 11. The figure 10(a) is shows the output of the purpose technique. It is developed by MetaQuote Language 4(MQL4) [14], it look like C language. The expert persons can see only figure 10(b) and code can see like a figure 10(c). The figure 10(b) is not easy to found the pattern, because any bars are very long, the expert persons may be not found the peak from bars chart. For figure 10(c) the peak finder unit is help to find the peak of price. So, the purpose technique can found pattern more than expert persons. The figure 10(d) represent to, when we change the time period to smaller, the pattern that only found by the purpose technique, can found by expert persons too. That look like the purpose technique can found pattern that represent in smaller time period to show in current period. V. CONCLUSION This article purpose the new technique for price pattern detection. It applies the finite state machine with fuzzy transition to development. The result, it can detect the price pattern similar the expert persons and can found the pattern that found by expert persons in smaller period. However, this technique need to pre-process for find the peak before uses it. So, the performance is depending on the performance of peak finder unit too. The purpose technique applies only head and shoulder price pattern which the invastor can be uses it to find the entry point or applied to entry decision module on automatic trading system. This technique has only one example on head and sholder prices pattern, however it easier applies to (b) (d) [1] Martin J. Pring, Martin Pring on Price Patterns, McGraw-Hill, NewYork,2004. [2] Lukasz Snopek, The Complete guide to Portfolio Construction and Management, John Wiley&Sons Ltd., United Kingdom, [3] John J. Murphy, Technical Analysis of the financial markets, New York institute of finance, New York, [4] Ken-ichi Kamijo and Tetsuji Tanigawa, Stock price pattern recognition a recurrent neural network approach, Proceeding of 1990 International Joint Conference on Neural Network, vol.1 pp ,17-21 June [5] Youngohc Yoon and George Swales, Predicting Stock Price Performance: a Neural Network Approach, Proceeding of 25 th Annual Hawaii International Conference on System Sciences, vol.4 pp , 8-11 January [6] Md. Rafiul Hassan and Baikunth Nath, Stock Market Forecasting Using Hidden Markov Model: A New Approach, Proceeding of 5 th International Conference on Intelligent Systems Design and Applications, pp , 8-10 September [7] Sheng-Hsun Hsu, JJ Po-An Hsieh, Ting-Chih Chih and Kuei-Chu Hsu, A Two-stage Architecture for Stock Price Forecasting by Integrating Self-organizing Map and Support Vector Regression. Expert Systems with Applications, vol.36 pp , [8] Jung-Hua Wang and Jia-Yann Leu, Stock Market Trand Prediction using ARIMA-based Neural Networks, Proceeding of 1996 IEEE International Conference on Neural Networks, vol.4 pp , 3-6 June [9] Ping-Feng Pai and Chih-Sheng Lin, A Hybrid ARIMA and Support Vector Machines Model in Stock Price Forecasting, Omega, vol. 33, Issue 6, pp , [10] Ching-Hsue Cheng, Tai-Liang Chen and Liang-Ying Wei, A Hybrid Model Base on Rough Sets Theory and Genetic Algorithms for Stock Price Forecasting, Information Sciences, vol.180, issue 9, pp , [11] Md. Rafiul Hassan, Baikunth Nath and Michael Kirley, A Fusion Model of HMM, ANN and GA for Stock Market Forecasting, Expert Systems with Application, vol.33, issue1, pp , [12] Pietro Ducange, Francesco Marcelloni and Michela Antonelli, A Novel Approach Based on Finite-State Machines with Fuzzy Transitions for Nonintrusive Home Appliance Monitoring, IEEE Transactions on Industrial Informatics, vol.10, No.2, pp ,2014. [13] Paolo Brandimarte, Numerical Methods in Finance and Economics, John Wiley&Sons Ltd., USA, [14] Meta Quotes Software. (2000). MQL4 Reference. [Online]. Available: [15] Computational Intelligence in Scheduling (SCIS 07), IEEE Press, Dec. 2007, pp , doi: /scis