Performance of Moving Average Trading Rules in a Volatile Stock Market: The Russian Evidence

ABSTRACT

This article examines the profitability of dual moving average crossover (DMAC) trading strategies in the Russian stock market over the 2003–12 period. It contributes to the existing technical analysis (TA) literature by testing, for the first time, the applicability of ordered weighted moving averages (OWMA) as an alternative calculation basis for determining DMACs. In addition, this article provides the first comprehensive performance comparison of DMAC trading rules in the stock market that is known as one of the most volatile markets in the world. The results show that the best trading strategies of the in-sample period can also outperform their benchmark portfolio during the subsequent out-of-sample period. Moreover, the outperformance of the best DMAC strategies is mostly attributable to their superior performance during bearish periods and, particularly, during stock market crashes.

KEY WORDS:

• market efficiency
• moving average
• ordered weighted average
• portfolio performance
• quantifier guided aggregation
• technical analysis
• trading rules

Notes

1. The profitability of technical trading rules in emerging stock markets has been examined by Ojah and Karemera (1999); Ratner and Leal (1999); Chang, Lima, and Tabak (2004); MacKenzie (2007); Fifield, Power, and Knipe (2008); Krausz, Lee, and Nam (2009); and Moosa and Li (2011), but the Russian market is not included in any of these studies. To our knowledge, the only exceptions are Chong, Cheng, and Wong (2010) and Ni, Lee, and Liao (2013), who both briefly examine the profitability of only a few technical trading rules as a basis of index trading in the BRIC (Brazil, Russia, India, and China) stock markets. The former study compares the annual returns of nine technical trading rules which do not include any DMAC strategies to the buy and hold returns of the corresponding indices. The latter focuses on only yhree DMAC combinations that, according to the authors, are often employed by practitioners. Moreover, the results of Ni, Lee, and Liao (2013) are restricted to average holding period returns obtained by following buy/sell signals, and the significance of the results is benchmarked against zero return instead of the B&H return as done commonly in TA literature (e.g., see Brock, Lakonishok, and LeBaron 1992; Hsu and Kuan 2005; Cheung, Lam, and Yeung 2011). Moreover, neither of the above-mentioned closest peer-group studies takes transaction costs into account. In addition, they use another stock index as a proxy for the Russian stock market (i.e. the RTS index) instead of MICEX. However, we prefer MICEX in our study for several reasons: The overall liquidity of MICEX companies has been higher than that of RTS companies due to the higher turnover in MICEX, although there is a significant overlap in the constituents of these two indices (e.g., see Grigoriev and Valitova 2002; Roschkow, Marsh, and Todorovic 2009 for details). The higher liquidity implies a better pricing efficiency, which supports the use of MICEX for our purposes. Second, the constituent list of the MICEX index has been updated on a regular basis from the beginning of its calculation, unlike that of RTS, which enables the calculation of a passive B&H benchmark return for the trading portfolios of constituent stocks. Third, unlike RTS data, MICEX data is available in the original currency (i.e., in rubles) throughout the 2003–12 sample period.

2. Throughout the study, we report the annualized volatilities calculated from the monthly returns. The reader should note that the annualized volatilities calculated from daily returns are generally much higher. For example, in this case, the corresponding volatility is as high as 37.28 percent p.a. based on daily returns.

3. As the longest moving average employed in our study is 200, the maximum number of preceding daily quotes required is 200.

4. Over the sample period, we also set an additional filter rule according to which stocks included in the analysis were not allowed to have more than fifteen equal prices in two subsequent trading days during the year preceding the investment period. If this rule was violated, we excluded such stocks from our sample until the above-mentioned condition was satisfied. The previous year price history was employed for this purpose to avoid look-ahead bias.

5. To eliminate “whiplash” signals in cases when shorter and longer MAs are close to each other, we test all DMAC rules with a 1 percent band, which requires that the shorter MA must exceed the longer MA by o1 percent before a buy signal is implemented. Correspondingly, the shorter MA must go below the longer MA by 1 percent to trigger a sell signal.

6. An alternative method is to calculate technical trading returns with one-day lag by beginning with the closing quote one day after a technical signal is initiated, consistently with e.g., Bessembinder and Chan (1998), Fong and Yong (2005), and Pätäri and Vilska (2014). However, this method also includes a trade-off because the trader would lose the one-day return if the timing of the buy signal is right. Correspondingly, in the case of the right timing of the sell signal, he/she would cut his/her losses one day later than when using the more commonly-used assumption employed in this article. We also calculated trading returns with one-day lag as a robustness check. Consistent with McKenzie (2007), the results were very similar for both calculation variations. Therefore, and for the sake of brevity, we only report the results based on the first-introduced assumption.

7. Many different weighting schemes are used for exponential weighting (e.g., see Gardner Jr., 1985, 2006). In this article, we use a widely-used variant as follows: w_i = 2/(i + n), where i refers to the serial number of the freshness indicator of the price quote within the timespan over which a moving average is calculated and n is the length of the particular moving average. Thus, the lowest weight within the timespan (i.e., 1/n) is given to the oldest quote, while the newest gets the highest weight (i.e. 2/(1 + n). We use this simple weighting scheme because it takes into account the wide range of MA lengths employed in our empirical tests.

8. The performance metrics are calculated on the basis of monthly returns in order to avoid some nondesirable characteristics of daily return distributions (e.g., high kurtosis and higher autocorrelation). In addition, the reader should note that the trading strategies chosen for closer examination are not the top three strategies among the all trading strategies examined, but instead, the best strategies among those that are based on three different DMAC variants employed (i.e., SMA, EWMA, and OWMA).

9. The percentage of transaction costs is based on the estimate given by the Moscow Exchange. We use equal proportional transaction costs throughout the full sample period, though in the real world, the transaction costs would have been higher during the first five-year sub-period. This approach was chosen for two main reasons: First, if an investor had liked to follow the best strategies of the first sub-period during the second sub-period, he/she would have paid the transaction costs according to the cost level of the second sub-period, not according to that of the first sub-period. Therefore, it is justified to also take into account the decrease in transaction costs in the calculation of returns of the DMAC strategies during the out-of-sample period (i.e., during the first sub-period). Second, we would like to keep our programming code as simple as possible, because in any case, we can only have an estimate of the real transaction costs.

10. The choice of DMAC combination used for comparison and demonstration purposes is arbitrary, and it is not the best-performing DMAC combination based on any of them being compared. We also made similar comparisons for other length combination of DMACs between the three DMAC variants examined and ended up with similar results according to which OWMA-based strategies with high p values generated the greatest number of trading signals, whereas the corresponding number was lowest for OWMA-based strategies with p = 0.5.

11. Pätäri and Tolvanen (2009) employed SKASR in their hedge fund study under the name of “adjusted Sharpe ratio,” which by its content was exactly equal to the SKASR used in this study. The name SKASR was adopted by Pätäri (2011) to avoid confusions with the other closely-related definition for “adjusted Sharpe ratio” introduced by Pézier (2004).

12. The Ledoit-Wolf test can take account of asymmetries in return distributions being compared, as well as the impact of autocorrelation in return time-series, which is particularly important to control when using the emerging market data (eg., see Harvey 1995; Kinnunen 2013). Because of the complexity of the test procedure and space limitations we do not describe the Ledoit-Wolf test in more detail here, but recommend the interested reader to see the original article (Ledoit and Wolf 2008. The corresponding programming code is freely available at: http://www.econ.uzh.ch/faculty/wolf/publications.html).

13. This implies that we get the results for 12 × 13,333 = 159,996 active trading strategies. Due to the abundance of the results and the fact that the great majority of the trading strategies underperform against their passive benchmark portfolio during the in-sample period, we only report for each DMAC variant the proportion of those strategies that have generated higher returns than their passive benchmark portfolio to the total number of MA length combinations (which is always 13,333 for each DMAC variant). The corresponding proportions are also reported on the basis of SKASRs.

14. The reader should note that the reported results are for those DMAC strategies that were the best within each DMAC variant during the in-sample period. None of those DMAC combinations were the best during the subsequent out-of-sample period. However, the results for the best strategies during the latter period are not reported in detail, because it would have been impossible to identify such DMAC strategies beforehand. The variability of the best rules over time is consistent with the earlier literature (e.g., see Sullivan, Timmermann, and White 1999; Ready 2002; Szakmary, Shen, and Sharma 2010; Pätäri and Vilska 2014).

15. The last sub-period from of May 24, 2012 to the end of December 2012 is classified as bullish despite the fact that the cumulative return of the MICEX index from the previous trough does not exceed 20 percent by the end of 2012. However, the rising trend continued until January 2013, when the cumulative return of 20 percent from the previous trough was exceeded before the trend turned down again.

16. The reader should note that the aggregate log-scale returns are used only to demonstrate the dependency of the relative performance of DMAC trading portfolios compared to the B&H portfolio on the general stock market trend. Thus, they are not real-world returns due to the discontinuity of the time span depicted.