     jross
Member
Username: jross
Post Number: 12
Registered: 03-2004Rating: 
Votes: 1
|
| | Tuesday, May 25, 2004 - 11:29 pm: |  
|
Hey Joe! What do you know about something called the Channel-Breakout System?
To be quite honest with you, I know enough about it to stay clear of the concept when it is placed into a system.
The channel-breakout system is the heart and soul of most high priced trading systems. It consists of a series of stops, set each day at some computed distance above the current high and below the current low of the market, to form channel bands that contain between them almost all of the market's price bars. The trading rule is simple and visual: buy when the upper channel band is penetrated by the market price, and sell when the lower channel band is penetrated.
The concept of channel breakouts was first introduced as volatility bands. There are a number of systems based on this concept including Bollinger Bands, which is the latest rendering of the concept. Bollinger Bands were preceded by Keltner Channels.
I have nothing against using channels as a way to select trade entries. But I object to them being placed into a mechanical trading system.
Almost every market swing is preceded by a “spike” or “impulse” day. Often this initial day is followed by a temporary stalling out, a small retracement, or even a whipsaw. You can adjust the channels to leave either more or less room between the bands and the price action. You can find the optimum distance that will generate accurate signals based on those impulse days, while still being far enough from the market to avoid most of the false signals.
There are a number of ways to determine the distance of the channel bands from the price action. The earliest methods used a fixed dollar amount distant from a moving average. Next came a setting of the bands based on a percentage distant from the moving average. Various mutations of this have occurred. E.g. Create a channel based on a moving average of the highs + a percent or fixed amount for the upper channel. Take a separate moving average of the lows – a percent or fixed amount for the lower channel.
At the expense of repeating myself, the basic setup can be varied in a surprisingly large number of ways. There must be a formula, or algorithm, that determines the distance from the market at which to place the bands. The algorithm can be very simple, like: “place the upper band 2 points above yesterday's close, and the lower band 2.5 points below yesterday's close”, or it can be much more complex and involve any number of mathematical functions and market variables.
The heart and soul of using channel breakouts is the ability to adapt automatically to changing market conditions. When the volatility rises, the targets move further away from the current price, and you usually avoid getting whipsawed. When the volatility falls, the channel width shrinks and the targets move closer, and you can get an early entry at the next breakout.
Therefore, the best channel breakout methods are based on volatility, and how accurately you are able to calculate it. There are many ways to compute volatility and most, in fact I could say all but one has it wrong. The most correct way I’ve ever seen to compute volatility goes to my friend Len Yates at OptionVue.
How do I know Len’s method is the best? When trading options using a model, OptionVue produces far and away the most accurate results. No one even comes close.
First we must calculate, for each day, the Average Volatility (AV). I will be showing you how Len calculates volatility further ahead. The AV is a measure of volatility which takes into account the daily point spread between the highs and the lows, adjusted to include gaps. The AV is calculated for a specified number of days, so you could have a 1, 2, 3, 4, 5, 10 or any length ATR. Tomorrow we will go into which is best.
Best results are usually obtained with AV lengths of from 1 to 7 days. If too many days are used for the AV, the system becomes unresponsive to changes in volatility. The AV is measured in the same time frame as the underlying—days, weeks, hours etc.
Next we assign a coefficient by which we'll multiply the AV. A coefficient is just a multiplier we arrive at through optimization. After we have multiplied the coefficient times the AV we arrive at a value, also in points or ticks, which we'll call the Point Move.
Best results are obtained with coefficients from .2 to 1.4. This is not unlike setting a coefficient i.e. multiplier for the Volatility Stop Study which I showed in Trading the Ross Hook. If the coefficient is too small we will get too many small trades. If it's too large we will get too few trades and miss much of the action.
The Point Move will vary in direct proportion to the current volatility. The coefficient, once we choose its value, does not change from day to day, it is a constant. But the AV varies with the market. The Point move is the product of a variable (the AV) times a constant (the coefficient).
Each day after the close we can update the AV using Len’s volatility method and prices for the past N days, where N is the number of days length for the AV. We then find the Point Move, by multiplying the AV times the coefficient, and we add the Point Move to today's close to get tomorrow's long target. We also subtract the Point Move from today's close to get tomorrow's short target.
If either target is hit during the day tomorrow, we enter the market at that respective target. This means we must either leave a resting buy or sell order with our broker, or keep an eye on the market intra-day.
Picture the long and short targets as forming bands above and below the market. We buy at the top band when it is penetrated and sell at the bottom band when it is penetrated. As you may realize, this is different from Bollinger’s Bands (BBs). With BBs, we do just the opposite—we sell at the upper band and buy at the lower band. The reason for this is that with BBs, the bands are set at statistical Standard Deviation. With channel bands, the bands are set at average volatility (AV).
In the basic system the upper and lower bands are always equidistant from the close. If the market needs to move 1.5 points up to trigger a buy, then it must move 1.5 points down to trigger a sell. The bands themselves will widen or narrow as volatility changes, but regardless of their width, the Close will always be at the center.
A more advanced version of the system uses separate coefficients to set the distance from the close to the upper and lower bands. This allows the system to better capture moves that are triggered by trending markets. For example, in an uptrend, reducing the value of the “up” coefficient will bring the upper band a little closer down to the market, and make the whole system more sensitive and more easily triggered to go long. At the same time the “down” coefficient can be increased to also drop the lower band a little, making it harder to trigger premature sell signals—which are often due to spike bottoms that occur just before a rally.
Traditionally, the basic breakout system has been used as a reversal system: you would go short as soon as you exited a long position, and long as soon as you covered a short. However, when using the advanced version, with separate up and down coefficients, it makes much more sense to divide the system into a long trading mode and a short trading mode. Each mode will be optimized to better take advantage of bull or bear trends, and, therefore, each will use a different set of parameters.
Occasionally you will find that the same set of parameters turns out to be optimum for both sides of the market, but this is rare other than in the currencies. This demonstrates what a poor practice it was “in the old days” to use a reversal strategy. By splitting the system into long and short modes, and by using separate up and down coefficients for each mode, you will not be constantly in the market, exposed to unproductive risk. Your money can collect some interest while the system waits to jump on the next opportunity.
VOLATILITY CALCULATIONS
Perhaps the “purest” method of measuring volatility is statistical analysis of the recent price history of the asset itself. Proponents of statistical analysis claim that theirs is the only way to arrive at the “true” volatility for an asset, and that the use of market implied volatility is circular. I agree with this view entirely.
While it may be true that statistical analysis is the only way to measure true recent volatility, market implied volatility might be a better indication of future volatility than the statistical method, because it is sensitive to current perceptions “on the floor.”
Many traders will compute their options model in reverse by manually plugging in the market implied volatility from the floor, and I agree that for computing future volatility, there can be no better figure.
However, the implied volatility figures produced by options models (as opposed to the market implied volatility from the floor) are not acceptable because those figures are only as good as the figures for statistical volatility which were used in the derivation of the model to begin with. That is the situation in which implied volatility becomes circular and becomes increasingly inaccurate.
Options are currently priced to reflect expectations about the future volatility of the underlying asset. These expectations are evident almost instantly when you measure volatility the “implied” way.
FORECASTING FUTURE VOLATILITY IS A KEY TO OPTIONS MODELING
Implied volatility is admittedly, somewhat circular. If everyone relied exclusively on implied volatility, the slightest error (i.e., rounding errors) could cause us all to drift further and further from the true volatility of an underlying asset until there was no resemblance between assumed and actual volatilities. Perhaps an occasional look at statistical volatility is important for “staying on course.”
In any case, forecasting future volatility is a key to options modeling.
STATISTICAL VOLATILITY
Perhaps the best way to measure statistical volatility (SV) is through the “extreme value method.” Using just daily high and low prices, the extreme value method is roughly five times more precise than the more common “close-to-close estimator. And the extreme value method has the further advantage of being unaffected by possibly missing days. You can go away for a week or two, and come back and pick up where you left off. The SV number might be a bit “stale,” but the formula deals with the missing day just fine. Truthfully the calculation of SV involves more than just the basic “extreme value formula.” There is an algorithm involving several steps. These steps represent refinements made over the years. These refinements were necessary because the ‘extreme value method’ has some inherent weaknesses. These will be dealt with later.
THE HEART OF THE ALGORITHM IS THE “EXTREME VALUE FORMULA”
SV = .627 * sqrt(365.25) * ln( H/L)
Where:
SV = statistical volatility
sqrt = the square root function
ln = the natural logarithm function
H = the high of the day
L = the low of the day
This formula gives you the statistical volatility number for one trading day, normalized to a year (all volatility numbers are normalized to one year).
The result of using the above formula on each of ‘n’ trading days is ‘n’ different SV numbers. These may then be combined by averaging. We use 20 trading days, and we use an exponential averaging method to give the most weight to the SV of the most recent trading day, and successively less weight to the other days as you go back in time.
HERE IS THE EXPONENTIAL AVERAGING ALGORITHM
a = 1.0
alpha = 0.92
numer = 0.0
denom = 0.0
daily volatility = H-L
For each trading day, do the following three instructions:
1. numer = numer + (a * daily volatility)
2. denom = denom + a
3. a = a*alpha
The final step is: Average SV = numer/denom
In this algorithm, ‘a’ represents the progressively smaller weight, and is used as a multiplier against each daily SV reading. You can see that ‘a’ starts out a 1.0 and becomes smaller as it is repeatedly multiplied by 0.92.
REFINEMENTS
An inherent weakness of the extreme value method is that opening gaps are ignored. For example, if a stock were to trade in a ½ point range one day (pretty low volatility), then open the next day 5 points higher and trade in a ½ point range once again, this 5 point opening jump would completely escape the notice of the extreme value method, as it only looks at each day’s high-low range.
To compensate for this, and to specifically deal with the Japanese Index (JPN) and others like it (whose range is never reported), but merely a ‘last’ which jumped from the day before), we need a system of combining pairs of consecutive trading days into a single two-day event before measuring SV.
When measuring SV from the high and low of a two-day period, the basic formula becomes:
SV = .627 * sqrt(336525) * ln( H/L ) / sqrt(2)
This is the same formula as before, plus a division by the square root of 2. Additionally, since we are using two-day pairs, we must include 21 days in order to get 20 individual measurements.
So when analyzing a sequence of daily records, we measure SV from successively overlapping two-day periods, but only if they are two consecutive trading days (this calls for a function which knows about weekends and holidays). When measuring SV from a day for which there is no subsequent trading day information, we apply the single-day formula.
One more important thing needs to be discussed, then you’ll see the whole picture.
The reality of real-world data is that it can contain occasional errors or spurious data. Bad data can be as simple as apparent price drop from a stock split, or the jump that accompanies a roll-over in futures contracts. Or, it can be bad information from a quote vendor.
To filter out bad data, we do three things:
First, each record is checked to make sure it contains both a High and a Low, and to make sure that the High is greater than or equal to the Low. Any record not meeting those conditions is thrown out.
Second, check for individual days of bad high/low prices in context, and throw them out. (e.g., if three consecutive days of high/low prices are 144/142, 76/7, 142/140, the middle one is obviously bad.)
Third, after all the SV numbers are compared, toss out the highest and lowest of these, prior to averaging.
Finally, refuse to return an average SV from too few samples. The fewest number of samples to use is 5. There must be at least 5 individual SV number or else the program must return a blank average SV.
HERE IS AN OUTLINE OF THE COMPLETE ALGORITHM
1. Filter out bad records on an individual basis.
2. Filter out bad records in context
3. Stepping from the oldest record to the earliest, compute SV from a combination of this record and the next, if the two represent consecutive trading days. If this record and the next are not consecutive trading days, compute SV from this day’s record alone.
4. There is now a collection of 20 SV readings. See if there are at least a minimum of 5 of these readings. If so, go on; otherwise, the program cannot compute average SV.
5. Toss out the highest and the lowest SV readings.
6. Combine the remaining SV readings using the exponential averaging method described above.
As you can see, this is a robust algorithm which is capable of dealing with a wide variety of data, even ‘dirty’ data.
A proper options program should track both implied and statistical volatility for every asset for which the user maintains a file, and uses these figures automatically (unless placed in ‘manual’ mode) for the volatility input to the Matrix and all forms of analyses.
However, the user must be given his choice of using statistical or market implied volatility, or any proportionate mix of the two. He should be able to input his choice of volatility figures in the manual mode.
Joe Ross
(Message edited by JRoss on May 25, 2004)
|
