Stocks & Commodities V. 3:3 (106-109): Part 1.

Concepts and Examples A Trend Following Method for All Seasons by Anthony W. Warren, Ph.D

Part 1. Concepts and Examples A Trend Following Method for All Seasons
by Anthony W. Warren, Ph.D.

Introduction Trend following methods based on plots of raw and moving average data are one of the oldest and most universally used techniques of technical analysis. However, these techniques are appropriate for only certain types of markets, characterized by great cyclical movements or trends, and are not useful in long accumulation and distribution periods when the underlying market is flat. Moreover, even in trending markets the traditional moving average trading techniques tend to be susceptible to false alarms caused by short term market reversals. In this article we discuss a trend following method which remedies most of these difficulties by (1) Establishing upper and lower trend channel lines which characterize the statistical fluctuations in the data, and (2) Utilizing a moving average or recursive filter to define trading points and trading regions OUTSIDE the uncertainty band between the trend channel lines. This concept is not really new. However, it has only recently become implementable in a reasonably economic and automated fashion, with the growth in power and usefulness of the personal computer. We will discuss the implementation aspects of this method in a follow-up article. In this article wt show the superiority of this method over the usual trend following technique based on moving averages. At its basis, our trend following concept identifies three trading regions when one should be LONG, when one should be SHORT, and when one should be OUT, i.e. noncommitted. These regions are defined by the following rules: (A) One should BUY or be LONG only when the moving average or filtered data lies below the lower trend channel line, (B) One should SELL SHORT only when the moving average lies above the upper trend channel line, and (C) One should be OUT of the market when the moving average value lies in the uncertainty region between the trend channel lines. Corollary to these rules are the optimal trading points, i.e. One should Buy or Cover a short position when the moving average penetrates the lower trend channel, and one should sell short or cover a long position when the moving average penetrates the upper trend channel line. These concepts are illustrated in Figure 1. In this figure, the straight lines represent the market price time series, the bounding lines near the price values represent the upper and lower trend channel lines, and the dotted smooth curve represents the lagged moving average for trading decisions. Four decision points are shown in the figure. After the market value has topped and has declined for some time, a SELL signal is given by the penetration of the moving average into the uncertainty region. Then, when the market drops abruptly, a SELL SHORT signal is given as the moving average value moves above the uncertainty

Article Text

Copyright (c) Technical Analysis Inc.

Stocks & Commodities V. 3:3 (106-109): Part 1. Concepts and Examples A Trend Following Method for All Seasons by Anthony W. Warren, Ph.D

Figure 1:

Figure 2:

Stocks & Commodities V. 3:3 (106-109): Part 1. Concepts and Examples A Trend Following Method for All Seasons by Anthony W. Warren, Ph.D

region. After the market has bottomed for some period of time, a COVER short signal is given, and then as the market moves up abruptly, a BUY signal is given as the moving average again moves below the lower trend channel. Implementation Issues The primary technical problem associated with the implementation of this technique is that of determining the trend channel lines. As we will show, the major problem is t hat of forecasting or extrapolation . Suppose we use a short sample length moving average as the basis for the trend lines. If we apply the moving average to the data such that the moving average output is centered at the mid-point of the input data, then the moving average may be used to represent the underlying trend. The difference between the input and the filtered data represents the statistical fluctuation about the trend. The trend lines are then determined as the moving average output plus or minus 2*Sigma where Sigma is the standard deviation of the statistical fluctuation. (Two Sigma gives an uncertainty region such that 95%, of the market values lie within the trend channels.) The only problem is that the moving average and thus the trend channels are only defined up to the current time minus one-half t he moving average length. Thus some method is needed to forecast the trend channels. One method to do this is to use the MEM forecasting method presented in earlier issues of Technical Analysis of Stocks & Commodities, e.g. December 1984. Although the MEM forecasting code gives excellent results for determining the trend channels, it is a sophisticated and somewhat time consuming analysis method. Typically, the forecasting period required is only on the order of two - five samples. Another, much simpler forecasting technique based on an Alpha-Beta recursive filter has been found to work almost as well. This filter is easy to implement (See the Appendix for a summary of the filtering equations), and is much faster and easier to use. For brevity, we postpone the implementation and usage details to a follow-on article. Example 1: Tandy Stock Analysis Frankly, last year (1984) was not a great investment y ear for the Warren household. My wife bought Tandy at a "depressed" price of $32 only to see the stock head south and end the year at a price near $25. She stuck it out though, and in February. 1985 was even considering adding to her position. Our trend following method was then applied to examine our current position. Figure 2 shows Tandy weekly data from its high point in the summer of 1983 until the end of Feb. 1985. A 25 week moving average was applied to this data and clearly shows two "false alarm'' buy points in December 1983 and August 1984, and a potential buy point in early February 1985. However, having been recently burned with moving average buy signals (See Example 2 below), I applied a full MEM analysis to this data and obtained the optimized 3*Sigma trend channels and forecasts shown in Figure 3. A 26 week lagged moving average was then applied and overplotted on the earlier data resulting in Figure 4. This figure shows two interesting points. First, one should be long in Tandy as of Feb. 1985 since the moving average penetrated the lower trend channel line, and secondly, that no "false alarm" buy signals were generated between July 1983 and January 1985 by other penetrations of the lower trend channel line! ( 1984 was a good year to short Tandy according to our trend following rules.) For comparison purposes, an Alpha-Beta filter with a smoothing time MS=4 weeks was applied to the Tandy data and the Two-Sigma trend channel lines for this filter are shown in Figure 5. (A two sigma value was used since the Alpha-Beta filter is not quite as accurate as the optimized moving average in

Article Text

Copyright (c) Technical Analysis Inc.

Stocks & Commodities V. 3:3 (106-109): Part 1. Concepts and Examples A Trend Following Method for All Seasons by Anthony W. Warren, Ph.D

Figure 3:

Figure 4:

Stocks & Commodities V. 3:3 (106-109): Part 1. Concepts and Examples A Trend Following Method for All Seasons by Anthony W. Warren, Ph.D

Figure 5:

Figure 6:

Stocks & Commodities V. 3:3 (106-109): Part 1. Concepts and Examples A Trend Following Method for All Seasons by Anthony W. Warren, Ph.D

estimating the underlying trend, and thus the value of "Sigma" is larger for the Alpha-Beta filter.) The trend channels for this technique are very similar to that of the MEM analysis except for the last few time points, but even here the Alpha-Beta results seem quite reasonable. Finally, we applied a lagged Alpha-Beta filter comparable to the 25-week moving average and obtained the overplot shown in Figure 6. This figure may be interpreted exactly like that of Figure 4, and shows that our trend following method can be implemented either with MEM trend channel forecasting or with Alpha-Beta filtering and forecasting. Example 2: Fidelity Select Technology Analysis The second example is also from the author's experience, and concerns a trading decision made in January 1984. At that time, the author was examining the Fidelity Select Technology fund with the view of adding to an established position in the fund. Figure 7 shows a plot of the weekly fund closing values from November 1982 to January 6, 1984, and an 11 week shifted moving average. After the fund peaked in June 1983, the fund values channeled in a fairly narrow trading zone, and appeared to be starting a second up-move from the secondary bottom in Nov. 1983. The moving average indicator appears to corroborate this analysis as the fund values rise above $26 in early January. Subsequently the author added to his position in the fund and the market promptly responded by reversing direction and headed south for another seven months. Figure 8 shows the nosedive that occurred from late January to mid-February 1984. Now let us examine the same fund data using the trend following method outlined earlier. Figure 9 shows the results of plotting the Alpha-Beta two Sigma trend channels and an Alpha-Beta data filter with lag similar to that of a 25 week moving average. Notice that our trend following method gives a sell signal in August 1983, and no further buy or sell signals through the end of January 1984! This chart shows clearly that the author's buy analysis in January 1984 was premature. (In fact an extension of the data period would give a sell short signal in February 1984, although in actuality one cannot short a Mutual fund.) Summary This article has presented a trend following technique which is inherently superior to conventional moving averages in minimizing trading risk. Of course one still must use protective stops and portfolio risk management techniques, to protect against unforeseen market swings. However, the technique of trading only when the lagged trend line remains outside an uncertainty region minimizes the probability of being whipsawed. It is the author's experience that this is far more important than techniques which maximize profits. (Far too often I have seen the profits of one season erased by losses on the next.) This technique has two other advantages which are of equal importance: * The method is fairly easily implemented using simple Alpha-Beta filters, and * The method is basically a charting technique which is easily mastered and interpreted. As a result, I have sworn off using crossing moving averages for trade analyses, and I would heartily recommend the same for you! Appendix: Summary of Alpha-Beta Filtering The Alpha-Beta filter implemented in the above article consists of equations for current sample data estimation, data momentum estimation, and data smoothing. Suppose z(i) denotes the raw price sample at the i'th time point, x(i) denotes the current trend estimate and v(i) denotes the momentum estimate at the

Article Text

Copyright (c) Technical Analysis Inc.

Stocks & Commodities V. 3:3 (106-109): Part 1. Concepts and Examples A Trend Following Method for All Seasons by Anthony W. Warren, Ph.D

Figure 7:

Figure 8:

Stocks & Commodities V. 3:3 (106-109): Part 1. Concepts and Examples A Trend Following Method for All Seasons by Anthony W. Warren, Ph.D

Figure 9:

Stocks & Commodities V. 3:3 (106-109): Part 1. Concepts and Examples A Trend Following Method for All Seasons by Anthony W. Warren, Ph.D

ith time point. The basic data filtering equations are given by: y(i) = x(i + v(i (one step forecasting) 1) 1) x(i) = y(i) + alpha*(z(i) y(i)) v(i) = v(i + beta*(z(i) 1) y(i)). The smoothed value used to represent the data trend is given by s(i ms) = x(i) MS * v(i), where MS is the smoothing time in samples. The author chooses the alpha and beta smoothing constants using the following design equations alpha = 2 / (MS + 1.5) beta = alpha / (MS + .5), i.e. the alpha and beta values are simple functions of the smoothing time MS. If the data z(i) is given for i=1,...N then the smoothed values are defined above for i=l,...N-MS. The trend is forecasted up to the current time point using the linear recursion: s(N+1 = x(N) (k * v(N) k=1,...MS. k) 1) The details for implementing trend following with this method will be forthcoming in my next article.

Article Text

Copyright (c) Technical Analysis Inc.