Futures Trading Signal using an Adaptive Algorithm Technical Analysis Indicator, Adjustable Moving Average'

Transcription

1 Futures Trading Signal using an Adaptive Algorithm Technical Analysis Indicator, Adjustable Moving Average' An Empirical Study on Malaysian Futures Markets Jacinta Chan Phooi M'ng and Rozaimah Zainudin University of Malaya Abstract. Adaptive technical analysis indicators are useful in current changing markets. This research shows an adjustable technical trading algorithm, Adjustable Moving Average' (AMA'), which generates the appropriate trading signal according to the prevailing market states either ranging or trending market state. The empirical tests are conducted, backtest and live test on Malaysian futures contracts, namely futures contracts on the FBM KLCI30, (FKLI) and the crude palm oil (FCPO). AMA' adjusts automatically to avoid some whipsaws in range market and to enter into the new trends early in trend market. AMA' outperforms the threshold passive buy and hold and generates more net returns in both: 1. backtest from 2/1/1996 to 31/12/2007 on FKLI, with net return of 1228, compares to 447 of the buy-and-hold strategy, 2. live-test from 2/1/2008 to 31/12/2012, on FKLI, with net return of 666 compares to 146 of the buy-and-hold strategy, and live-test from 2/1/2008 to 31/12/2012,on FCPO, with net return of 3079 compares to -572 of the buy-and-hold strategy. Keywords: Malaysian Futures Contracts, Algorithmic Trading, Technical Analysis Indicators, Adaptive Moving Averages, Adjustable Moving Average Prime AMA' 1. Introduction In recent years, interest in high frequency trading, online trading algorithms and time series shows a marked increase (Masteika, Rutkauskas 2012[10]). Accompanying this interest is the increase in the number of researches into new computation trading algorithms that the research find useful for investment timing decisions (Lukac et al. 1988[9]; Brock et al.,1992[3]; Irwin, Park 2009[7]) They investigate the possibility of abnormal returns in different financial markets using various algorithmic trading models and find that computational techniques are useful in timing investment decisions in time series. Lukac et al. (1988)[9], Brock et al. (1992) [3] and Irwin, Park (2009) [7] demonstrate that technical trading systems generally generate abnormal returns larger than those by passive buy-and-hold strategy. These studies test different mechanical trading systems and find evidence consistent with technical analysis being able to identify trends for the purpose of trading profitably. Balsara et al. (1996) [2] highlight the need for new trading systems that adjust in tandem with the prevailing market condition. Gandolfi et al. (2008) [6] research on dynamic volatility concept in the trading systems address this need. In current changing market conditions, adaptive technical indicators are used to differentiate trend from range trading. Currently many markets, including Malaysian futures markets, are changing regularly from trending to ranging and vice versa. If this type of market scenario is likely to continue over the next few years, then algorithmic adaptive indicators are particular suitable trading instruments to use to decipher between ranging market to avoid some whipsaws and trending market for early entry into new trends. The research, therefore, addresses the constant and persistent problem that has baffled traders over the centuries on when the market is trending and when the market is ranging. It is important for the professional trader to know when the market is trending as different trading strategies are employed in different market conditions.

2 Other works on adaptive moving averages include Kaufman Adaptive Moving Average (Kaufman, 1998[8]) and Chande Market Oscillator (Chande, 1997[5]). Both found that the adaptive moving averages perform better than the threshold buy-and-hold and better than the simple moving average. The objectives of this research are to ascertain that in the long run: 1. algorithm trading systems perform better than passive buy and hold, 2. simple moving average and standard deviation trading systems generate net profits, 3. for the periods 1996 to 2007, AMA' generates more net profits for FKLI and for the periods 2007 to 2011, AMA' generates more net profits for FKLI and FCPO than some of the other common technical analysis indicators. The special feature of an adaptive algorithm trading system is its ability to automatically adjust its parameter to be a large variable in range trading period, and to be a small variable in trend trading period. This adaptive trading system functions according to the inherent algorithm which first recognises the state the market is in, ranging or trending, and then adjusts the parameter accordingly. The underlying concept of this algorithm, Efficacy Ratio is, it dynamically and automatically varies the parameters of technical indicator(s) to suit the current market state. Efficacy Ratio = Long Term Standard Deviation/ Short Term Standard Deviation...Equation (1) Efficacy Ratio increases the length of the moving average when the market ranges, and decreases the length of the moving average when the market trends. AMA' is an adaptive moving average that employs the concept Efficacy Ratio to determine its length. AMA' is a set of adjustable moving average trading rules that quantitative traders can use on the model trading desk. 2. Data Analysis The data that we are using are the closing prices of the futures contract on FTSE Bursa Malaysia Kuala Lumpur Composite Index ( FKLI ), from 15 th December 1995 when it first commences trading to 31 st December The data source is from Bursa Derivative Malaysia. We note its open, high, low, close and volume. We concentrate and test the returns using the closing prices to check for leptokurtosis. The main characteristic in time series that traders are interested in is volatility or huge returns which are found in the heavy tails of the daily price changes' distributions. Therefore, this research will first chart FKLI daily price changes' distribution. The purpose of analysing FKLI daily price changes is to find if the distribution exhibits leptokurtosis. If the distribution exhibits excess kurtosis from the normal distribution, then it can not be inferred that FKLI daily price changes are random. Contract Mean Standard Deviation Skewness Kurtosis FKLI ' Table 1: FKLI Daily Return Statistics from 15/12/1995 to 31/12/2012 The kurtosis of for FKLI daily price changes exhibits excess kurtosis from the normal distribution of 3, therefore, it can not be inferred that daily returns are random. The frequency distribution of FKLI daily returns shows heavy tails distribution as is shown in Figure 1.

3 Below to to to to to to to to to to to to to to to to to to to to 100 Above Fig. 1. Frequency distribution for FKLI daily returns from 15/12/1996 to 31/12/ Methodology 1. First, 7 best practice trading methods selected based on the abnormal returns are used for backtests for FKLI for the past period from 2 nd Jan 1996 to 31 st Dec Then, to test the robustness of these systems, live test for the current period for both FKLI and futures on crude palm oil contracts (FCPO) from 2 nd Jan 2008 to 31 st Dec 2012 are conducted. The trading systems used are based technical trading rules specified by Brock et al. (1992) as well as innovations of moving average and standard deviations bands. The technical trading rules specified by Brock et al. (1992) include Variable Length-Moving-Average technical rules, and Fixed-Length-Moving Average rules. The 7 (3 common and 4 innovated) trading systems are: 3.1. Simple 21 Days Moving Average (MA) The most common mechanical trend trading system is the simple moving average which Brock et al. (1992) refers to as Variable Moving Average (1,21,0%). 1 refers to the closing price, 21 refers to 21 periods moving average and 0% refers to 0% from the simple moving average. The method to construct this simple moving average trading system is to calculate the average of 21 daily closes and compare that to the current close. If the current close is above the 21 day moving average, then the signal is to buy. If the current close is below the 21 day moving average, then the signal is to sell and 21 Days Moving Average Crossover (MAC) Another common mechanical trend trading system is the moving averages crossover which Brock et al. (1992)[3] refers to as Variable Moving Average (3,21,0%). 3 refers to the 3 periods moving average, 21 refers to 21 periods moving average and 0% refers to 0% from the averages. The method to construct this moving averages crossover trading system is to calculate the average of 3 daily closes and the average of 21 daily closes. If the 3 day moving average is above the longer 21 day moving average, then the signal is to buy. If the 3 day moving average is below the 21 day moving average, then the signal is to sell. Both the previous moving average(s) systems are fixed length moving average(s) and the lengths, 3 and 21 are arbitrary chosen Moving Average Convergence Divergence (MACD) MACD is a timing model, that Appel invented during the late 1970s.(Appel, 2005[1]). MACD is created by subtracting a longer-term exponential moving average from a shorter-term exponential moving average. MACD can be calculated using the difference between the 12-day short-term moving average minus the 26- day moving average. Appel introduces a further component of MACD: Signal Line which the average of the difference between the shorter-term and longer-term moving averaes. Signal line can be calculated by taking the 9-day exponential averages of MACD. A buy signal is generated when MACD crosses up above Signal Line and a sell signal is generated when MACD crosses below the Signal Line Kaufman Adaptive Moving Average (KAMA)

4 In order to vary these moving averages according to market conditions, Kaufman (1998)[8] proposes to apportion weights to the current data, and past smoothened data series according to Efficiency Ratio in the formula below: KAMA t = a ER C t + (1-a ER) KAMA t-1 where a=[(er x (2/3-2/31))+2/31] 2 and ER=(C t -C t-n )/Absolute Sum of(c t C t-1 )...Equation (2) C t is the most current close and C t-1 is the previous close. If the current close is above the KAMA t, then the signal is to buy. If the current close is below the KAMA t, then the signal is to sell Adjustable Moving Average' (AMA') KAMA uses Efficiency Ratio to determine the weight of the current data and past smoothened data series whereas AMA' adjusts the length of the moving average for each different period according to the prevailing Efficacy Ratio. Efficacy Ratio = Long Term Standard Deviation/ Short Term Standard Deviation...Equation (1) If the current close is above the AMA t, then the signal is to buy. If the current close is below the AMA t, then the signal is to sell. However, these systems are turn and reverse systems, which mean that the trader trades all the time, even in range periods when the trader gets a lot of whipsaws BB Z-Test-Statistics (BBZ) To avoid trading unprofitably during range periods, BBZ (Chan, 2005[4]), to trade when volatility increases, when the price moves above +1 or below -1 standard deviation band. The method to construct BBZ is to calculate the 21 day moving average and 1 standard deviation. The next step is to add 1 standard deviation to the 21 day moving average to get the upper band and to deduct 1 standard deviation from the 21 day moving average to get the lower band. If the close is above the upper band, then the signal is to buy and when the close is below the upper band, the signal is to exit long. If the close is below the lower band, then the signal is to sell and when the close is above the lower band, the signal is to exit short. BBZ can be programmed into the trading system as follows: 1) Under System Tester, key in: Under Enter Buy, Close>BbandTop(Close, 21, Simple, 1) Under Exit Buy, Close<BbandTop(Close, 21, Simple,1) Under Enter Sell, Close<BbandBot(Close,21,Simple,1) Under Exit Sell, Close>BbandBot(Close,21,Simple,1) 2) Run Simulation Tests on the data (FKLI and FCPO). 3) View Results after the test to check for amount of profit, no of trades, profit versus unprofitable trades, average gain versus average loss per trade, maximum consecutive gains versus maximum consecutive losses. Table 2: Sytem Instructions for BBZ 3.7. Optimised BBZ (Opt BBZ) Fixed length BBZ (21,1) produces result that only favour trends which begin when prices move beyond the 1 standard deviation bands from the 21-days simple moving average. For other periods, when market is moving directionally very fast (or not moving directionally at all), 21 periods and 1 standard deviation may not be the most optimal parameters to use. Therefore, optimisation is performed to find the optimised parameters that produce the best results. Optimisation is a series of simulations with different parameters with the intention of selecting the most optimal parameters that generate the most profit with the least number of consecutive losses. The system tester then generates the most optimised moving average and

5 standard deviation. In system tester, steps 1 to 4 are repeated, replacing 21 with Opt1 and 1 with Opt2. The most optimised parameters for FKLI and FCPO for 2008 are 34 day moving average and 0.8 standard deviation. 4. Empirical Results Table 3 presents the backtests results for FKLI between 15/12/1996 and 31/12/2007. Based on Table 3, the results show that 6 of the 7 algorithm trading systems generated more profits than the passive buy-andhold (BH) strategy for FKLI. AMA records the highest return of 1,228, followed by MA (1,037), BBZ (835), MAC (830) and KAMA (786). BBZ Opt, due to the optimisation nature, shows an ideal highest profit of 2,279. Only MACD (258) fails to outperform the BH strategy (447). FKLI BH MA MAC MACD KAMA AMA' BBZ BBZ Opt Sum Table 3: Backtest Results on FKLI from 2/1/1996 to 31/12/2007 Table 4 is the summary of FKLI livetests results for the period of 2/1/2008 to 31/12/2012. Unlike backtest results, all the 8 tested models outperform the BH strategy (146). KAMA (690) registers the highest return of 690, followed by AMA (666), MA (623), MAC (622), BBZ (463) and MACD (177). In this period, the ideal BBZ Opt shows highest profit of 763. FKLI BH MA MAC MACD KAMA AMA' BBZ BBZ Opt Sum Table 4: Live Test Results on FKLI from 2/1/2008 to 31/12/2012 If we combined the results of backtests and livetests, then 7 of the 8 trading systems outperform the BH strategy. AMA (1894) is the most profitable trading system, followed by MA (1660), KAMA (1476), MAC (1452) and BBZ (1298). Only MACD (435) fails to outperform the BH strategy (593). In this period from BBZ Opt shows the highest profit of 3,042.

6 Fig 2: FKLI Daily Closes and AMA' for 2/1/2012 to 31/12/2012 Table 5 summarises the results of the livetests on FCPO for the period between 2/1/2008 to 31/12/ out of the 8 trading systems outperform the BH strategy. AMA (3,709) produces the highest return. MA (3,649), MAC (3446), BBZ Opt (3,149), KAMA (1,718) and BBZ (1,575) follow in this order of performance. Once again, only MACD (-664) cannot outperform the BH strategy (-572). FCPO BH MA MAC MACD KAMA AMA' BBZ BBZ Opt Sum Table 5: Live Test Results on FCPO from 2/1/2008 to 31/12/2012 Fig 3: FCPO Daily Closes and AMA' for 2/1/2012 to 31/12/2012 Similar findings are reported by Lukac et al. (1988)[9], Brock et al. (1992)[3] and Irwin and Park (2009) [7]. Therefore, it can be seen that for the period 2/1/1996 to 31/12/2012, AMA' generates the most net profits for FKLI and FCPO, compared to the other models and the BH strategy. Except for MACD, the other 6 trading systems consistently outperform the BH strategy. The profits for BBZ Opt which are the highest, are disregarded because optimisation generally ensures the best fitted parameters to past data and thus produces the maximum profits on hindsight. 5. CONCLUSIONS

7 It has been shown that one of the most important feature of an algorithm trading system to high frequency traders in model trading desks worldwide, is its ability to adapt quickly so that it can be robust in different markets and across different time frames. In designing the new algorithm trading system, the technical indicator used should show this ability. Conceptually, Efficacy Ratio is designed to adjust to both the different market conditions, range market and trend market. AMA' is designed to address some of the common problems encountered by most trend trading systems like being triggered by floods of orders generated by common trading systems (like simple moving average), being whipsawed in range market and inability to capture the trend by entering the trend too late and exiting the trend too early. In selecting the most ideal algorithm trading system, factors to take into consideration are that the trading system should not encounter large losses, or show net large loss in any of the years. The algorithm trading system should work well in practice as in testing and that it can adjust automatically to the parameter shifts. In testing, slippage should be taken into consideration. In summary, this research ascertains that: 1. The prices of FKLI and FCPO contracts tested are not random, 2. Mechanical algorithm trading systems like moving averages and AMA' can be used to capture the abnormal returns arising from trending behaviour. 3. AMA' is a robust adaptive technical trading system that can be implemented from past and current empirical evidence and it is possible that it can contribute to the profits of the model trading desk. AMA' ability to adjust according to market conditions points a new research direction for incremental machine learning trading systems. AMA' can be extended, innovated and developed into new adaptive algorithm trading systems. Possible implications from this work are that adaptive new trading indicators like AMA' and Adjusted Bands Z-Test statistics (ABZ') can be applied immediately on any professional model trading desk. However, it should be noted that there is still much to be done for future further research to find specific algorithms to automatically determine the length of the long and short term standard deviations, the preset width of the bands from the moving average and the maximum parameter for the bands. 6. References [1] Appel G Technical Analysis, Power Tools for Active Investors. Financial Times Prentice Hall. Upper Saddle River, NJ. [2] Balsara N, Carlson K, Rao N. Unsystematic Futures Profits with Technical Trading Rules: A case for Flexibility. Journal of Financial and Strategic Decisions. 1996, 9(1): [3] Brock W, Lakonishok J, LeBaron B.. Simple Technical Trading Rules and the Stochastic Properties of Stock Returns. Journal of Finance. 1992, 47: [4] Chan J. Using Time Series Volatility to Trade Trends: Trading Technique BBZ. Australian Technical Analysts Association. 2005, [5] Chande T Beyond Technical Analysis: How to Develop and Implement a Winning Trading System. John Wiley 61 Sons Inc. [6] Gandolfi G, Rossolini M, Sabatini A, Caselli S. Dynamic MACD Standard Deviation Embedded in MACD Indicator for Accurate Adjustment to Financial Market Dynamics. IFTA Journal.2008, [7] Irwin S, Park C. A Reality Check on Technical Trading Rule Profits in the U.S. Futures Markets. Journal of Futures Markets. 2009, 30: [8] Kaufman P Trading Systems and Methods. John Wiley & Sons, Third Edition. [9] Lukac L, Brorsen B, Irwin S. A Test of Futures Market Disequilibrium Using Twelve Different Technical Trading Systems. Applied Economics : [10] Masteika S, Rutkauskas AV. Research on Futures Trend Trading Strategy Based on Short Term Chart Pattern. Journal of Business Economics and Management. 2012, 13(5):