Comparing a Fixed and Flexible Bollinger Band Trading System

Nauzer J. Balsara*Judith Aburmishan, Andrey Kanalin 



Introduction: The Limitations of Fixed Parameter Mechanical Trading Systems

        A mechanical trading system is a set of rules defining entry into and exit out of a trade. There are two kinds of mechanical systems: (a) predictive and (b) reactive. Whereas a predictive system uses historical data to predict future price action, a reactive system uses the same data to react to price trend shifts. Instead of predicting a trend change, a reactive system would wait for a change to develop, generating a signal to initiate a trade shortly thereafter. The success of any reactive system is gauged by the speed and accuracy with which it reacts to a reversal in the underlying trend. In this paper, we will restrict ourselves to a study of Bollinger bands (Bollinger 1992), a popular mechanical trading system of the reactive kind.

        While prices do have a tendency to trend every so often, these trends do not seem to recur with definite regularity. Moreover, the magnitude of the price move in a trend varies over time, and no two trends are exact replications. Although the existence of trends cannot be denied, there is an annoying randomness as regards their magnitude and periodicity. This randomness is the Achilles heel of conventional mechanical systems based on fixed, market-invariant parameters, since it is virtually impossible for them to capture trends in a timely fashion consistently. Consequently, a fixed lag system may perform exceptionally in one time period, but can be woefully inadequate during a slightly different market environment (Balsara 1992).

        A fixed-parameter mechanical system implicitly expects market conditions to adapt to its invariant and market insensitive logic. A flexible-parameter system, on the other hand, seeks to modify its parameters to adapt to changes in market conditions. This paper seeks to develop a flexible alternative to the conventional fixed-parameter Bollinger band system. While it might be wishful thinking to conjecture that such a system would dominate its fixed-parameter counterpart in terms of profitability, it is reasonably safe to hypothesize that a flexible system would be more consistent in its performance. 

Top of Page

The Structure of Conventional Bollinger Bands and their Inflexibility:

        Bollinger bands are lines plotted x standard deviations above and below the moving average of close prices over the past n periods, forming an envelope around the moving average. As prices penetrate the upper band, we have a signal to buy the stock in question. Conversely, as prices penetrate the lower band, we have a signal to sell. The moving average defines the market’s pivot point at any given time, and the standard deviation bands around the moving average define the trend reversal points. The Bollinger band system keeps the trader constantly in the market, reversing from a buy trade to a sell trade. Like most other mechanical trading systems, the Bollinger band system is a fixed-parameter system, in the sense that it holds constant the lag length (n) over which the moving average of historical prices is analyzed. The lag length is chosen on the basis of back-testing historical data to determine what appears to have worked best in the past. The logic is that a moving average lag length rule which has worked best over the recent past will continue to do so in the future.

        The Bollinger band system defines standard deviation bands around a moving average of prices over a fixed lag length. This moving average defines the market’s pivot point at a given point in time. The average gives equal weight to each and every data point in the data set without regard to the directionality of the data. Consequently, as is evident from Appendix I, it is possible for two data sets to be very different in terms of their directionality and yet share the same average or pivot point. Whereas data series 1 has a clear upward trend, data series 2 has no clear trend. Notice that both data series share the same average, leading to the same exit or reversal price. Given the upward trend of data series 1, prices will have to move lower quite a bit more as compared to data series 2 simply to reach the same pivot point. Trend switches are therefore likely to be more difficult in a trending market as opposed to a non-directional market. To offset this, an up trending market (data series 1) ought to have a higher pivot as compared to a sideways market (data series 2). This is a major weakness of the conventional fixed-parameter Bollinger band system (Kaufman 1995).

Top of Page

Designing a Flexible-Parameter Bollinger Band System:

        Since the pivot point and stop-setting process in the flexible-parameter approach is based on the directionality of the market, we begin by introducing an index of market momentum or directionality over the past n periods as defined by Chande and Kroll (1994) and given below:

                                                                                      Total "Up" days - Total "Down" days

Momentum index  =  ----------------------------------
                                       Total "Up" days + Total "Down" days

        A higher close today as compared to yesterday results in an "Up" day. Conversely, a lower close today as compared to yesterday results in a "Down" day. The sum of all "Up" days ("Down" days) over the past n periods gives us the total "Up" days ("Down" days) in the above formula. The momentum index as defined above ranges anywhere from +1 (a perfect up trend, as in case of data set 1) to -1 (a perfect down trend). A value close to zero implies a lack of directionality in the market.

        In an up-trending market, when the momentum index is positive, the pivot point of the current trend ought to be higher than the moving average, somewhere near the high end of the trading range (Ruggiero 1996; Ruggiero 1998). This will allow for a faster reversal from a buy to a sell, when the up trend reverses. Accordingly, a flexible pivot point is defined as the n period moving average of prices plus some fraction of the difference between the n period highest price and the n period moving average. The fraction in question is given by the current n period momentum of the market. Therefore, the stronger the directionality of the up trend, the closer the pivot point is to the highest high over the past n periods. The formula for the n period flexible pivot in an up trending market, when the momentum is positive, is as under:


Flexible pivot = Moving Average + Momentum * (Highest High - Moving Average).

In the extreme case, when the momentum is +1, the flexible pivot would be exactly equal to the highest high price over the past n periods.

        In a down trending market, when the momentum index is negative, the pivot point of the trend ought to be lower than the average over the past n periods, closer to the lowest low over the past n periods. This will allow for a quicker reversal from a sell to a buy when the down trend reverses. A flexible pivot point is defined as the n period moving average of prices minus some fraction of the difference between the moving average and the n period lowest low. Again, the fraction is given by the current momentum of the market. The stronger the directionality of the down trend, the closer the pivot point is to the lowest low over the past n periods. The formula for the flexible pivot point in a down trending market, when the momentum is negative, is:


Flexible pivot = Moving Average + Momentum * (Moving Average - Lowest Low)

In the extreme case, when the momentum is -1, the flexible pivot would be exactly equal to the lowest low price over the past n periods.

Research Objectives:

        We will check for overall differences between the results of the fixed and flexible versions of the Bollinger band system using data from the Omega Research historical database for six randomly picked futures contracts over the past eleven years, 1987 to 1998. Specifically, we wish to compare the fixed and flexible versions of the system for differences in profit performance over the entire period. Next we wish to check for the consistency of performance of the two methods, by comparing the coefficient of variation of trade profits for each of the two methods over the past eleven years.

Top of Page

Methodology for Proposed Research:

        Both fixed and flexible versions of the Bollinger band system will be programmed using Omega Research’s Trade Station software. Next the programs will be run on the nearest futures contract between January 1987 and July 1998 for each of six commodities, namely, corn, soybeans, gold, silver, Swiss Francs, and the Japanese yen. Since both fixed and flexible versions of the system are applied to the same commodity over the same time period, we can assume that both systems are subject to identical capital requirements. Moreover, both systems are programmed to trade a single contract of the commodity in question at all times, making both systems equally risky from a trading perspective. Since we propose to work with a long-term trend-following system, we have arbitrarily chosen a 90-day moving average as the basis for computing the fixed/flexible pivot price. To distinguish between meaningless fluctuations (noise) and genuine trend changes, we propose to use a 3 standard deviation band width around the pivot price to signal a trend reversal. In the case of a long trend, a close 3 standard deviations below the current pivot price signals a reversal from long to short. Conversely, in the case of a short trend, a close 3 standard deviations above the current pivot price signals a reversal from short to long. The total dollar net profit (realized and unrealized) for both the fixed and flexible approaches will be obtained from the performance summaries, and individual trade profits will be obtained from the detailed trade results generated by the Trade Station software.

        To check for statistically significant differences in profits resulting from the two approaches over the entire time period, we will use a T-test of differences in the means of paired two-samples with the null hypothesis that there is no difference between the methods. A calculated t-value in excess of the critical value shows a significant difference between the two approaches. To check for consistency over shorter time periods, we propose to divide the sample into two approximately equal sub-periods, January 1987 to December 1992, and January 1993 to July 1998. A similar T-test will be used to test for differences between the two methods during each of the two sub-periods. Again, the null hypothesis is that there is no difference between the two methods for any given time period. Assuming that the normality assumptions behind the T-test are not entirely satisfied, we will also employ a non-parametric Wilcoxon signed-rank test of the differences in the profit performance of the two approaches.

        To check for consistency of performance under the two approaches, we will compute the coefficient of variation of completed trade profits for each commodity for both the fixed and flexible approaches over the period January 1987 to July 1998. Next, we will use a T-test and the Wilcoxon signed-rank test to check for significant differences between the coefficients of variation for the two approaches, with the null hypothesis that there is no difference between them.

Top of Page

Research Findings:

 Appendix II summarizes the total profit for both the fixed and flexible approaches for each of the six commodities over the entire period, January 1987 to July 1998. Appendix III summarizes the profit for the two approaches for each of the six commodities over the two sub-periods, January 1987 to December 1992 and January 1993 to July 1998. The results of the T-test for paired-sample differences in the mean are given in Table I(a). The t-value for differences in profits for the two approaches over the entire sample period is 2.9070, signifying a statistically significant difference in favor of the flexible method. The t-value for differences in profits over the sub-period January 1987 to December 1992 is 2.6939, again suggesting a statistically significant difference in favor of the flexible method. The t-value for differences in profits over the sub-period January 1993 to July1998 is 1.7725, suggesting a statistically insignificant difference between the two approaches.

        The results of the Wilcoxon signed-rank test are given in Table I(b). The z-statistic for differences in profits for the two approaches over the entire sample period is 2.2014, which is well beyond the critical z value of 1.645 at the 5 percent level of significance. This suggests that the flexible method returns profits that are statistically significantly higher than those provided by the fixed approach. The same result is observed for each of the two sub-periods, with a z-statistic of 1.9917 for both the 1987/1992 and 1993/1998 periods. Hence we can say that the flexible approach results in higher profits as compared to the fixed approach.

        The results of the T-test for paired-sample differences in the coefficient of variation for both methods are given in Table II(a). The t-value for differences in the coefficient of variation for the two approaches over the entire sample period is 2.8496, signifying a statistically significantly difference. The coefficient of variation is higher in case of the fixed pivot approach as compared to the flexible pivot approach. However, the t-value for differences in the coefficient of variation for both methods drops to a statistically insignificant 1.6462 for the 1987/1992 sub-period and is once again statistically insignificant at -0.9437 for the 1993/1998 sub-period. Although this could be due to the rather limited number of trades executed in each of the sub-periods, the results prevent us from emphatically concluding that the flexible approach results in a lower coefficient of variation as compared to the fixed approach.

        The results of the Wilcoxon signed-rank test are summarized in Table II(b). The z-statistic for differences in the coefficient of variation over the entire sample period is 2.2014, once again suggesting a statistically significant difference in the coefficient of variation for the two approaches over the entire sample period. However, the z-statistic is statistically insignificant at 1.3628 and 0.1048 for the 1987/1992 and 1993/1998 sub-periods. This reiterates our earlier finding that we cannot conclusively say that the flexible approach results in a lower coefficient of variation as compared to the fixed approach.

Top of Page

Conclusion:

        From the above results, it is clear that the flexible approach results in higher profits as compared to the traditional fixed-parameter approach. Although the coefficient of variation of individual trade profits is lower in case of the flexible approach over the entire trading period, this is not true when we consider trading over shorter time periods. More research is necessary before we can conclusively say that the flexible pivot approach is more efficient in terms of controlling the volatility of trade profits, a matter of particular importance to fund managers who wish to protect themselves from wide swings in trading performance. Working with a shorter than 90-day lag length for the moving average might be one way to increase the number of trades in each of the sub-periods. It is also recommended that this study be extended in the future to cover other fixed-parameter mechanical trading systems, in order to ascertain whether these results can be replicated across a broader spectrum of commonly used systems. 



Top of Page