The final method is one that I use in my seasonal work. I call it the

pinpoint when, during a given week, you should take a position histori-

Ruggiero/Barna Seasonal Index. This index is part of a product we call

cally. On a shorter-term basis, it can tell you how strong the bear or bull

the Universal Seasonal, a TradeStation or SuperCharts add-in that auto-

market is. In bull markets, the high occurs later in the week, in bear mar-

matically calculates many different measures of seasonality if the his-

kets, the high is earlier in the week.

torical data are available. This tool will work on all commodities and even

The final form of day-of-week analysis is conditional day-of-week

on individual stocks.

analysis. Buying or selling is done on a given day of the week, based on

some condition-for example, buy on Tuesday when Monday was a down

day. This type of analysis can produce simple and profitable trading

THE RUCCIERO/BARNA SEASONAL INDEX

patterns.

Larrv Williams, a legendary trader, developed the concept of trading

The Ruggiero/Barna Seasonal Index was developed by myself and

day-of-month analysis. This concept is very powerful for discovering hid-

Michael Barna. The calculations for this index are shown in Table 2.2.

den biases in the markets. There are two major ways to use this type of

I would like to make one point about the RuggierolBarna Seasonal

analysis: (1) on an open-to-close or close-to-close basis, and (2) more

Index: It is calculated rolling forward. This means that all resulting trades

often, by buying or selling on a given trading day of the month, and hold-

are not based on hindsight. Past data are used only to calculate the sea-

ing for N days. When a holding period is used, this type of analysis can

sonal index for tomorrow™s trading. This allows development of a more re-

produce tradable systems by just adding money management stops.

alistic historical backtest on a seasonal trading strategy.

Let™s 1l0w discuss three methods for calculating seasonality on a yearly

Besides the RuggierolBarna Seasonal Index, you can use the raw av-

basis, The first method originated in the work of Moore Research, which

erage returns, the percent up or down, and correlation analysis to develop

calculates seasonality on a contract-by-contract basis, using a calendar

trading strategies. The Ruggiero/Barna index can be calculated either by

day of the year. Moore Research converts prices into a percentage of

using the complete data set or by using an N-year window.

yearly range and then projects this information to calculate the seasonal.

The second method is the work of Sheldon Knight, who developed a

seasonal index he calls the K Data Time Line. The calculation involves

STATIC AND DYNAMIC SEASONAL TRADING

breaking down each year according to the occurrences on a given day of

the week in a given month. The steps for calculating the K Data Time

A seasonal trade can be calculated using the complete day set, some point

Line are shown in Table 2.1.

in the past, or a rolling window of data. This is true for day-of-week,

CALCULATING THE K DATA TIME LINE. TABLE 2.2 CALCULATING THE RUCGIERO/BARNA

TABLE 2.1

SEASONAL INDEX.

1. identify the day-of-week number and the month number for each day to be

Develop your seasonal and update it as you walk forward in the data.

1.

plotted-for example, the first Monday of May.

2. For each trading day of the year, record the next N-day returns and what

2. Find the 5.year price changes in the Dollar for that day, in each of the years

percentage of time the market moved up (positive returns) and down

identified.

(negative returns).

3. Add the 5.year average price change for that day to the previous day™s time

3. Multiply this 5.day return by the proper percentage.

line value. The full-year time line value starts at zero.

4. Scale the numbers calculated in step 3 between -1 and 1 over the whole

4. Trade by selecting the tops and bottoms of the time line for your entries and

trading year. This is the output value of the RuggierolBarna Seasonal Index.

exits. Buy the bottoms of the time line and sell the tops.

Seasonal Trading 47

46 Classical Market Prediction

In a seasonal pattern, a statistically significant percentage of the re-

day-of-month, and day-of-year seasonals. The question is: Which method

turns should follow the direction of the bias. For example, in a bullish

is the best? The answer depends on the commodity being analyzed. For

seasonal, the goal is to analyze the percentage of the time when the mar-

example, in markets with fixed fundamentals, the more data used and the

ket rises during the holding period.

longer they are used, the greater the reliability of the seasonal. If we were

In evaluating the percentage bias, the number of cases is a very im-

analyzing corn, we would want to go back, using as much data as possi-

portant link to the significance of the seasonal pattern. For example, on

ble. On the other hand, if we were doing seasonal research on the bond

a day-of-week pattern with hundreds of trades, 57 percent or 58 percent

market, we would not want to use any day before January 1, 1986, be-

is acceptable. On a day-of-year pattern with only 10 cases, we would want

cause, prior to 1986, the dynamics of the bond market were different.

to see 80 percent or better.

Another important issue in calculating seasonality is basing results on

Another important issue arises when evaluating a seasonal: Does the

in-sample trades versus walk forward testing. For example, if we say a

seasonal bias make sense? For example, suppose corn falls after the dan-

given seasonal is 80 percent accurate over the past 15 years, based on the

ger of crop damage from drought passes, or T-Bonds fall because of a

results of the seasonal trades over the past 15 years, that is an in-sample

quarterly refunding. If the seasonal pattern makes sense, it is more likely

result. If we use one day in the past 15 years to calculate a seasonal and

to work in the future.

then only take trades in the future using a buy and sell date calculated on

past data, and roll the window forward every day, this is walk forward

testing. More realistic results may be possible. For example, in 1985, you

CQUNTERSEASONAL TRADING

might mt have had a seasonal bias on a given day, but, years later, that day

of the year is included in a given walk forward seasonal pattern. Suppose

Seasonal trades do not always work. The question is: How can you tell

you calculate the seasonal walking forward using only data from 1970 to

whether a seasonal is going to fail and what should you do when it does?

1985. You trade in 1986 and then move the window up every year or so.

Seasonal trading fails when more important fundamental forces drive a

In 1987, you would use data including 1986 to calculate the seasonal, and

market. In 1995, the S&P500 did not have a correction in September or

you could produce a realistic seasonal model that can be used to trade.

October because of good fundamentals and falling interest rates. The

strength of the market foreshadowed the power moves of the S&P500

during late 1995 and early 1996. In another example of a seasonal failing,

JUDGING THE RELIABILITY OF A SEASONAL PATTERN

corn continued to rise during July 1995, because of the drought damage

in the Midwest. There are several ways to see whether a seasonal pattern

One of the main criticisms of seasonal trading is that it is only curve fit-

is working. For example, you can give a seasonal pattern 4 diys to work.

ting and is not based on any real fundamental influence in the market. This

Or, you can use Pearson™s correlation to measure the difference between

problem is more pronounced in trading by day of year because, often, only

the actual price movements and the seasonal. This is a very useful mea-

10 to 20 cases are available for calculating a seasonal pattern. Because of

sure in developing mechanical seasonal trading systems.

this issue, it is important to be able to judge whether a seasonal pattern

will hold up in the future. Most will not. There is no sure way to know, but

reliable seasonals do have similar characteristics. First, the returns of the

CONDITIONAL SEASONAL TRADING

seasonal pattern must be significantly above the average-day bias over the

same period; that is, if the seasonal pattern is based on the SBrP500, we

In conditional seasonal trading, you filter the cases you use in develop-

might want $200.00 a day on the long side but $100.00 could be acceptable

ing your seasonal patterns. For example, you could develop a trading day-

on the short side because the S&P500 has an upward bias. Second, one

of-month seasonal and only include cases when T-Bonds are above their

trade should not account for too large a percetitage of the profits.

Seasonal TradinP 4 9

Classical Market Prediction

48

TABLE 2.3 DAY OF WEEK AND S&PSOO.

26.day moving average. Another example of conditional seasonal trading

would be developing a day-of-year seasonal for corn but only using years Net Change Average Change Percent of

after crop damage in calculating the seasonal. This sounds like a curve fit, Day of Week (Points) (Points) Buy and Hold

but this method has worked well for Moore Research over the years.

Monday 282.69 .43 77.1%

Tuesday 8.45 .Ol 2.2

Wednesday 168.60 .25 45.8

OTHER MEASUREMENTS FOR SEASONALITY Thursday 42.35 .06 6.6

Friday -1 17.26 -.18 -31.9

The most used measure of seasonality is based on price, but seasonal ef-

fects also exist for both volatility and trend. For example, we can measure

Other markets-for example, T-Bonds-also have strong day-of-week

the average True Range/Close over the next N days based on day of week,

effects. During the period from l/1/86 to 4/12/86, the T-Bond market

month, or year. This measure will give us an idea of future volatility,

closed .07 point higher than the open on Tuesdays and -.02 point lower

which is useful for option trading as well as for setting protective stops.

on Thursdays. The day-of-week effect on the other days of the week was

Another useful way to use seasonality is to measure future trends. This

not statistically significant. The downward bias on Thursdays is caused

can be done using any trend level indicator-for example, ADX or Ran-

by the fact that most traders do not want to be long T-Bonds before a

dom Walk Index (RWI). Another good measure is using a simple differ-

major report, and many major reports, such as the monthly unemploy-

ence of ADX over the next N days relative to a trading day of month or

ment count, are released on Friday mornings just after the open. For this

year. This will tell us historically whether the seasonal effect will cause

reason, many traders sell bonds on Thursdays. This downward bias is also

a trend to get stronger or weaker. This type of information can be used

significant because T-Bonds have had an upward bias over the past ten

to filter trend-following systems.

years.

Seasonal patterns can also be used to forecast future price movements.

Besides the financial markets, other markets are influenced by strong

An example of this would be to take the current price as a base, and then

day-of-week effects. For example, since 1986, Thursday has been the

add to it the future change predicted by the seasonal. Finally, you would

most bullish day to trade silver. Because silver can be used as a measure

apply a correction factor based on the difference of the actual price

of economic strength, it would make sense that silver should have an up-

change over the last few forecasts and the seasonal forecast.

ward bias on days when T-Bonds have a downward bias.

Having discussed the issues involved in seasonal trading, let™s now

Even some of the soft commodities have a day-of-week bias; for ex-

study some examples of using seasonality for trading in several different

ample, coffee is most bullish on an open-to-close bias on Thursdays, and

markets.

it has been slightly bearish on Mondays since January 1, 1980. Believe it

What are effects of day of week in several different markets? We will

or not, in the period from l/1/80 to 4/12/96, if we do not deduct slippage

start with the S&P500.

and commissions, coffee has risen by $76,211.25 per contract by buying

The day-of-week bias in the S&P500 is revealed by measuring the dif-

at the Thursday open and exiting on the close.

ference between the close and open on a given day of the week during the

period from l/3/83 to 4/12/96, using backadjusted continuous contracts.

The results by day of week are shown in Table 2.3. Note that buy and hold

BEST LONG AND SHORT DAYS OF WEEK IN MONTH

during this period is 367.60 points.

Table 2.3 shows that you can outperform buy and hold simply by buy-

The day-of-week effect is not the same for every month; in fact, differ-

ing on Mondays and Wednesdays. We can also see that Fridays have a sig-

ent days of the week can become bullish or bearish, depending on the

nificant downward bias.

Classical Market Prediction

50 Seasonal Trading 51

month of the year. Let™s now examine how the month affects the day-of- with $50.00 deducted for slippage and commissions. This is only a taste

week analysis on an open-to-close basis. We analyzed several commodi- of the work you can do using day-of-week analysis.

ties, starting at various dates and ending on April 12, 1996. We did not

deduct slippage and commission because we wanted to judge the bias of

each market based on a given day of the week in a particular month. Table TRADING DAY-OF-MONTH ANALYSIS

2.4 shows our results.

Now that we have shown the effects of simple day-of-week analysis, The concept of analyzing the markets based on the trading day of the

let™s consider some examples of conditional day-of-week analysis, to learn month was originally developed by Larry Williams, who found that, in

how conditional day-of-week analysis works. many markets, the trading day-of-month effect is so strong that the results

One type of conditional day-of-week analysis reviews a market by day are comparable to commercial trading systems.

of week and measures what the market has done over the past five days. Let™s analyze several different markets on the basis of entering a po-

To illustrate, we will analyze the S&P500 in the period from 4/21/82 to sition on the next open after a given trading day of the month, and exit-

4/23/96, using a continuous backadjusted contract. ing on the open a given number of days later. We performed this analysis

Let a 1 mean that a market finishes higher than it opens; -1 means a on the S&PSOO, T-Bonds, coffee, and crude oil. The results of this analy-

lower amount, and a zero (0) means we do not care. Using this simple sys- sis are presented without an allowance for slippage and commissions be-

tem, with $50.00 deducted for slippage and commissions, we have found cause we wanted to test the bias of each market. Our results from the

some impressive results: start date to April 12, 1996 are shown in Table 2.5.

These results are only a sample of the power available through trading

Awrage T r a d e Win%

Net Profit

DOW DI D2 03 04 D5 Position

day-of-month analysis. Many different markets have an upward or down-

$405.24 68%

$25,125.00

Long

Monday -I -1 0 0 0 ward bias 60 percent (or more) of the time, based on a trading day of the

193.91 54

61.470.00

Short

Friday -I 0 0 0 0

month plus a given holding period. Another fact we learned from our

analysis is that the end-of-month effects in the S&P500 and T-Bonds are

In another type of conditional day-of-week analysis, we would use in- magnified when a month has more than 21 trading days; for example, the

termarket analysis in order to filter the day of the week. For example, let™s 22/23 trading day of the month produces great results but too few trades

take only Mondays when T-Bonds are above their 26.day moving average. to be reliable.

This simple pattern has averaged $249.45 per trade since April 21, 1982,

˜TABLE 2.5 SEASONAL EFFECT BY TRADING DAY OF MONTH.

Trading Day *“W.?ge

TABLE 2.4 DAY OF WEEK IN MONTH EFFECTS.

Commodity Start Position of Month Hold N e t Prolit Win% Trade

Commodity Start P o s i t i o n D a y oi W e e k M o n t h W i n % A v e r a g e T r a d e N e t P r o f i t

S&P500 412 1182 Long 17 5 $140,850.00 68% $1,354.33

$221.09 $15.255.00

Sept. 61%

cofiee l/l/80 Long Thursday S&P500 412 1 I82 Short 2 2 47,775.oo 55 459.38

19.248.75

278.97

,une 70

1 /I 180 Short Friday

Colfee T-Bonds l/1/86 Long 15 8 66,625.OO 63 550.00

10.125.00

289.29

May 66

Long Tuesday

T-Bonds l/1/86 T-Bonds l/1/86 LO”g 15 5 5,306.25 65 441 .oo

13.658.00

290.60

Mar. 57

Friday

T-Bonds t/1/86 S h o r t T-Bonds l/II86 Short 3 4 27,875.OO 56 230.37

427.73 23,525.OO

July 69

Thursday

l/3/83 tong

S&P500 Coffee l/1/80 Long 10 3 71,362.50 64 432.50

29,5*5.00

536.82

Dec. 65

LO”g Monday

S&P500 l/3/83 Coffee l/1/80 Short 14 7 70.826.25 62 429.25

20,225.oo

374.54

Thursday Dec. 63

l/3/83 Short

S&P500 Crude 4121182 tong 15 7 26,310.OO 6, 185.28

Seasonal Trading 53

52 Classical Market Prediction

This revelation should not make you think seasonality does not work,

DAY-OF-YEAR SEASONALITY

but it should point out that, when evaluating a seasonal, you need to cal-

culate and evaluate in a walk forward manner. Many seasonals are reliable.

of week and month, let™s turn to day-of-

Now that we have discussed day

For example, let™s look at the beginning-of-year rally. On January 26,

year analysis. Day-of-year seasonality requires more. comprehensive

1988, our seasonal, starting with data on April 21, 1982, shows an average

analysis in order to judge the reliability of a given pattern, because many five-day return of 3.14 points and a market rising 75 percent of the time.

patterns will have only 20 or fewer occurrences. In 1996, the same seasonal showed an average return of 3.61 points and

In addition, many of the seasonal patterns change over time. Fig- still a 75 percent accuracy. In 1996, this seasonal made over $7,000.00 in

ure 2.1 shows both average five-day returns and the percentage of the only 5 days.

time the market rose for the S&P500 futures during the period around One of the best ways to judge the reliability of a seasonal pattern is to

the crash of 1987, based on a seasonal calculated using data starting on look over the average returns and the percentage of accuracy over the

April 21, 1982. The seasonal for October 7, 1987, shows an average gain years. Seasonals that remain constant over the years are more reliable.

of 2.10 points and a percentage up of 100 percent. If we had been trading

seasonality back in 1987, we would have been long, not short, during this

time. Even a seasonal using data starting in 1960 would have yielded a USING SEASONALITY IN MECHANICAL TRADING SYSTEMS

long position in early October 1987.

Let™s now test several methods for evaluating a seasonal mechanically.

We will begin by using the S&P500 cash, starting in 1960. We will then

wait until we have 2,500 days™ data to take a seasonal trade. In our sim-

ple experiment, we will view the S&P500 only from the long side. Our

goal is to find low-risk buy points for the S&P500. We found in our re-

search that a seasonal pattern must be at least 70 percent reliable. For the

S&P500, using a holding period of 8 days and a seasonal return of .03

percent or greater produced the best results for that period. The .03 per-

cent represents about 2 points, based on an S&P500 with a value of

$600.00. Finally, we took the seasonal trades only when the S&P500 was

below its six-day simple moving average. These rules produced the re-

suks shown in Table 2.6.

The table shows that these seasonal trades offer above-average returns

for the S&P500. Based on these seasonal patterns, the market rises dur-

ing the next 8 days almost 70 percent of the time. One of the most im-

portant elements in getting these results is the fact that we collect 2,500

days™ seasonal information before taking a trade. Having this much data

improves the reliability of the seasonal. These seasonal patterns were

found using the Universal Seasonal, TradeStationTM and SuperChartsTM

FIGURE 2.1 The S&P500 futures average 5.day returns, and the and have adjusted themselves over the years. How have these seasonal

percentage of the time the market rose over a 5.day period by trading day patterns performed lately? Very well. They have not had a losing trade

of year, for the period arqund the crash of˜l987. The data used represent since October 1992.

the complete history of the S&P500 futures contract up to 1986.

Classical Market Prediction

54 Seasonal Trading 55

TABLE 2.6 S&P500 SEASONAL SYSTEM while being in the market about 40 percent of the time. Because the

BASED ON AVERAGE RETURNS OVER .03%. S&P500 has an upward bias, a -.20 value could still represent a market

with positive returns over that holding period.

3124171

First trade

612El96 Using all of the data in calculating a seasonal is not always the best

Ending date

571.01

Buy and hold solution. My research has shown that this decision depends on the com-

Total points made 270.39

modity being analyzed. In corn, or other commodities with fixed funda-

1,368

Days in market

mentals, the more data the better. In commodities like T-Bonds, a moving

47.4%

Buy and hold

window will work better. Let™s lwlw use a moving window to develop pat-

2 1 .67%

In market

terns in the T-Bond market. To calculate our seasonal, we used data start-

-28.90

Drawdown

69% ing on September 28, 1979. We developed the seasonal by using a walk

Win percentage

forward method with various window sizes, holding periods, and trigger

levels. We tested seasonality only on the long side, in order to simplify

our analysis. We collected 2,000 days™ seasonal data before generating

The Ruggiero/Barna Seasonal Index combines both average returns our first trade. We found that the best window size was a rolling window

and percentage of accuracy into a standardized indicator. When we ran of 8 years, with a 6-day holding period and a trigger level above -.20 to

this indicator across the same data, we found that the Ruggiero/Barna generate a buy signal. We filtered our trades by requiring a 6-period mo-

Seasonal Index can outperform the market based on seasonality. Once mentum to be negative to take a long trade. Our first trade was taken on

again, we waited 2,500 days before taking our first trade. Our data period September 20, 1988. Our data ran to June 28, 1996. The results for these

is the same length as the one used in the previous example, and it started parameters over about 7.75 years of trading, without slippage and com-

on January 4, 1960. We used a holding period of 5 days and a trigger of missions, are shown in Table 2.8.

-.20 on the Ruggiero/Barna Seasonal Index. We took the seasonal trades

only when the S&P500 was below its IO-day moving average. The results

using these parameters are shown in Table 2.7. COUNTERSEASONAL TRADING

Table 2.7 shows that, without taking a short position, the Ruggierol

Barna Seasonal Index can outperform buy and hold by over 30 percent Many times, the market does not follow its seasonal patterns. Being able

to detect an aberration and use it for knowing when to exit a trade, or

even for trading against the seasonal, can give you a big advantage. Let™s

examine 5-day average returns in the T-Bond market, and the correlation

TABLE 2.7 SEASONAL S&P500 SYSTEM RESULTS

between the seasonal returns and the current actual returns. We will use

BASED ON RUCCIERO/BARNA SEASONAL INDEX.

a 15.day Pearson™s correlation in our correlation analysis. Figure 2.2

2/l 9171

First trade

6128196

Ending date

573.89

Buy and hold

TABLE 2.8 T-BOND RESULTS BASED ON

761.75

Total points made

THE RUCCIERO/BARNA SEASONAL INDEX.

2,576

Days in market

132.73%

Buy and hold Net profit $57,593.75

40.6%

In market Win% 71

-44.68

Drawdown Average trade $282.32

6wo

Win percentage Maximum drawdown -$7,656.25

Seasonal Trading 57

56 Classical Market Prediction

In markets that have stronger seasonal influences, such as the corn

market, taking the trade in the opposite direction to the seasonal pat-

tern when the seasonal pattern fails can produce great results. Let™s test

one of the classic seasonal patterns. We will sell corn on the first trad-

ing day after June 20 and exit this position on November 1. This sea-

sonal trade has produced using cash corn prices dating back to June 2,

1969: the equivalent of $2X,487.50 on a single future contract, which

represents an average of $1,095.67 per year and 65 percent profitable

trades. The problem with this system is that, during several years (e.g.,

1974, 1980, 1993, and 1995), we would have suffered large losses. Let™s

now see what would happen if we go long on the corn market once we

know the seasonal has failed. We go long corn after July 21 if our trade

is not profitable. If we take a long position, we will not exit to the first

trading day of the following year. Using this method-going long the

first trading day after July 21 if the short trade is not profitable-pro-

duced $38,537.50 on a single contract. The winning percentage did drop

I

Feb MB,

NW m?c s3 wr to 58 percent overall. The drawdown as well as the largest losing trade

did improve. Using this seasonal failure method increased the net profit

FIGURE 2.2 T-Bonds average 5.day return versus trading day of year,

and cut the drawdown on the classic seasonal for shorting corn.

and the correlation of actual market conditions to this seasonal. The

This is just one example of how using counterseasonal trading can be

failure of the seasonal rallies in February 1996 led to one of the sharpest

a powerful tool for traders. Research in counterseasonal trading is one of

drops in the T-Bond market™s history.

the most interesting and profitable areas in seasonal research.

Seasonal patterns do not relate only to price; they can also relate to

volatility. We calculate seasonal volatility by finding the next S-day av-

shows both S-day average returns and their correlation to the actual price

erage (true range/price) x 100 for every given trade day of the year. This

action for November 1995 to April 1996.

measure predicts volatility based on seasonality. The calculation has sev-

As shown in Figure 2.2, T-Bonds have positive historical 5-day returns

eral uses. The first is for trading options. If volatility is going lower on a

from late February to mid-March. After that, T-Bonds have near-

seasonal basis, you will want to sell premium. Another use for this infor-

zero/negative returns until the end of March, in anticipation of the fed-

mation is in setting stops based on historical average true range. If sea-

eral income tax day (April 15). In 1996, during this seasonal strength,

sonal volatility increases, you will want to widen your stops.

the market decorrelated from its seasonal normal and dropped over 4 full

Figure 2.3 shows the average 5-day seasonal volatility for T-Bonds

points in the next 5 days-an example of how a seasonal failure can lead

from December 1995 to June 1996. T-Bond volatility has a peak in early

to explosive moves. This move accelerated during the seasonal flat-to-

January and falls in early February, before it rises again. During the first

lower period during the month of March.

three quarters of March, T-Bond volatility drops, reaching a low during

Seasonal trades can be filtered by entering only seasonal trades when

the last week of March. Based on seasonality, there is high volatility dur-

the correlation between the seasonal and the current market conditions is

ing mid-May and June.

above a given level or is higher than some number of days ago. This logic

The final type of seasonal analysis is the seasonal trend. Seasonal trend

would have protected against several bad seasonal trades in the T-Bond

analysis works as follows. For each trading day of the year, we returned

market in 1996.

Seasonal Trading 59

the average ADX value N days into the future. This indicator is a very

good tool to add as a filter for trend-following systems. Figure 2.4 shows

a price chart of T-Bonds and a 5-period lookahead of a lo-period ADX

seasonal. Note that T-Bonds do not trend seasonally in early December

and do not begin to trend again until February. As T-Bonds moved in a

trading range during late 1995 and early 1996, the ADX seasonal was

low or falling. When the trend seasonal started to rise, T-Bonds started

to trend to the downside.

This chapter has given you a brief look at the power of seasonal trad-

ing. In later chapters, we will combine some of these ideas with other

forms of analysis in order to predict future market direction.

FIGURE 2.3 T-Bonds versus seasonal average volatility for the period

December 1995 to June 1996.

FIGURE 2.4 The seasonal trend index, based on trading day of year for

T-Bonds. The downtrend in the T-Bond market during February 1996 was

part of the seasonal trend tendency.

58

Lone-Term Patterns and Market Timine for Interest Rates 61

To study the relationship between T-Bill yields and inflation, we re-

searched monthly data going back to 1943. Most major increases in short-

term rates occur when the inflation rate is a negative real rate-that is,

3 it is greater than the T-Bill yield. It last happened in 1993, just before the

start of a severe bear market in bonds. In general, rising premiums on

T-Bills lead to lower rates, and falling premiums lead to higher rates. We

Long-Term Patterns and studied many different ways of comparing inflation to interest rates and

have found that one of the best methods is to use a ratio of interest rates

Market Timing for to inflation. During the past 53 years, on average, T-Bill yields have been

about twice the average inflation rate.

Interest Rates and Stocks The relationship between long-term interest rates and inflation is not

as reliable as the one between short-term interest rates and inflation. In

general, the spread between inflation and long-term interest rates is be-

tween 300 and 400 basis points. Currently, it is about 380 basis points or

3.80 points as of early April 1996. The ratio between long-term interest

rates and inflation is currently about 250 percent; for example, a 3 per-

cent inflation rate would relate to a 7.5 percent long-term bond. This re-

lationship has varied over the years. Long-term rates were kept

artificially low during the mid-1970s. On January 31, 1975, long-term

This chapter will show you how to use fundamental data to predict long-

rates were at 5.05 percent, which was only about half of the actual infla-

term trends in both interest rates and stock prices.

tion rate. Another example occurred during the early 196Os, when infla-

This type of long-term analysis is very important for people who

tion was under 2 percent and long-term bond rates were about 4 percent.

switch mutual funds, as well as anyone with a variable rate loan. It is also

This was only a 2.00 point difference, but the ratio of long-term interest

important for short-term traders because many systems are based on buy-

rates to inflation has recently ranged from 220 percent to 260 percent.

ing pullbacks in the long-term uptrends of both stocks and bonds, which

This type of premium is common during long periods of economic growth

started during the early 1980s. When these bull markets end, these sys-

with low inflation. This concept is very important because it means that

tems will stop working-with disastrous results.

a 1 percent increase in inflation can produce a 2.5 percent increase in

long-term interest rates.

In May 1996, the Treasury Department discussed issuing a bond that

INFLATION AND INTEREST RATES

yields a fixed number of basis points over the rate of inflation. This

would be a smart move because it would reduce the cost of borrowing

It is commonly known that interest rates are positively correlated to in-

over the next few years. During the early 199Os, the Treasury moved its

flation. As inflation rises, so do interest rates. In general, this relation-

own borrowing to the short end of the yield curve just before short-term

ship is true, but it is not constant. We will examine this relationship using

rates dropped to a low of 3 percent. When it looked as though short-term

3-month T-Bill yields and yields on the longest government bond. We will

rates were going to start to rise, the Treasury suggested the issuing of an

compare these yields to the l-year inflation rate, calculated by taking a

inflation-based bond.

12.month percentage change in the˜consumer Price Index (CPI). These

This type of bond would save the Treasury money during periods of

data, as well as the other fundamental dataused in this chapter, were sup-

long-term growth and moderate inflation. During these periods, the

plied by Pinnacle Data Corporation and are part of their index database.

60

Classical Market Prediction

62 LonpTerm Patterns and Market TiminK for Interest Rates 63

premium between interest rates and inflation can be expected to remain TABLE 3.2 RESULTS OF INFLATION

over 200 percent. For example, suppose the inflation rate rises to 4.0 per- AND SHORT-TERM INTEREST RATES.

cent from its current 2.8 percent. On an inflation bond purchased at a

Net basis points 32.17

400-basis-point premium, the yield would rise from 6.8 percent to 8.0

Rise in basis points 16.34

percent. Our research has shown that during moderate increases in in- Fall in basis points 15.83

flation, long-term rates can retain over a 200 percent premium to infla- Average trade 2.34

Largest loser -.79

tion. In 1996, the ratio was 243 percent. Based on my model of long-term

Trades 14

yields to inflation, the long-term bond yield would increase from 6.8 per-

Percent correct 86%

cent to 9.72 percent. Under these conditions, this new inflation bond, is-

sued in January 1997, would save the government 1.72 percent in interest

per year. effect of inflation on longer-term rates is not as strong as it is on shorter-

term rates. Using the same general model with different trigger levels,

we can predict longer-term rates using inflation, but not as well as

PREDICTING INTEREST RATES USING INFLATION shorter-term rates. The results as well as our new model, for the period

from l/l/71 to 4/l/96, are shown in Table 3.3.

Let™s llow use the interaction between inflation and short-term interest

rates to develop a long-term 3-month T-Bill yield model. Inflation be-

came a better measure of interest rates after 1971, when the U.S. gov- FUNDAMENTAL ECONOMIC DATA FOR PREDICTING

ernment allowed the price of gold to float and dropped the gold standard INTEREST RATES

to back the U.S. Dollar. Table 3.1 shows how inflation can be used to

model short-term interest rates. Given this interaction between interest rates and inflation, how many other

This is a very robust model for short-term interest rates since the fundamental factors affect both long- and short-term rates? Using data

United States abandoned the gold standard in 1971. The results from Jan-

uary 1, 1971, to April 1, 1996, are shown in Table 3.2.

TABLE 3.3 RESULTS OF INFLATION AND LONG-TERM RATES.

Even more amazing, the average correct signal lasts 24 months and the

average wrong signal lasts only 2 months. This model has not produced a Ratio=l-(Inflation/Yield)

losing signal since January 3 1, 1986. InflatYieldOsc=Ratio-Average(Ratio,201

This is a good model of how inflation affects short-term interest rates. If Ratio<.25 or InflatYieldOsc<O and Yield>Yield 4 months ago then long-term

Let™s MIW apply the same general model to longer-term interest rates. The interest rates will rise.

If Ratio>.35 or InflatYieldOso.45 and Yield<Yield 4 months ago then long-

term interest rate5 will fall.

TABLE 3.1 INFLATION BASED SHORT-TERM NOTE MODEL.

Summary:

Results

Ratio=l-(Inflation/Yield)

Net basis points 20.55

lnflatYieldOsc=Ratio-Average(Ratio,ZO)

10.88

Rise in basis points

If Ratio<.2 or InflatYieldOsc<O and Yield>Yield 3 months ago, then go-day

Fall in basis points1 9.67

interest rates will rise. Largest error -.64

If Ratio>.3 or InflatYieldOso.5 and Yield<Yield 3 months ago. then go-day Forecasts 17

interest rates will fall. Percent correct 71%

-,

64 Classical Market Prediction Long-Term Patterns and Market Timina for Interest Rates 65

supplied by Pinnacle Data Corporation, let™s see how various fundamental This model has produced about 100 basis points per year for the past

factors can be used to predict interest rates. We will use money supply, 26 years. The average trade is about 23 months. The last signal this sys-

consumer confidence, and unemployment data to build our models. tem gave was in October 1995, when it predicted a drop in short-term in-

We start by showing how changes in the money supply affect interest terest rates.

rates. We use three standard measures of money supply: Money supply is highly predictive of short-term interest rates but not

as predictive of long-term interest rates. Using the same basic model with

Ml = money stored as cash and in checking accounts. different parameters did not produce results as good as those when pre-

dicting short-term rates. Current M2Chg was compared with the reading

M2 = Ml plus money stored in time deposits, such as CDs.

16 bars before, and current yields were compared with those 3 months be-

M3 = Ml and M2 plus assets and liabilities of financial institutions,

fore. The model did a fairjob of predicting long-term rates. For example,

which can be easily converted into spendable forms.

it produced 17.44 points since 1970 and was 65 percent accurate on 26

trades. The draw down was only -1.64 points, and the average trade was

In general, the greater the amount of money in the system, the more the

.67 points. These are good results but not as good as for the prediction of

economy will grow. This growth translates into higher interest rates.

shorter-term rates.

Let™s now develop a simple model using the monthly change in M2.

With this background in money supply and inflation, we can now dis-

When M2 is in an uptrend and rates have begun to increase, then rates

cuss how some other measures of economic activity can be used to pre-

will continue to increase. If M2 is in a downtrend and rates have begun

dict interest rates. Let™s start by using consumer sentiment to predict

to fall, then rates will continue to drop. Using the same period as ear-

short-term rates. Our model is based on the fact that if consumers are

lier-January 1971 to April 1, 1996-we can develop a short-term in-

positive, then growth will increase and, consequently, interest rates will

terest rate model based on M2. Our rules are as follows:

rise.

Our model compares consumer sentiment and T-Bill yields between a

1. If M2Chg > M2Chg [6] and 90-day Yields > 90-day Yields 11

given number of bars. The rules are:

months ago, then 90 days interest rates will rise.

2. If M2Chg < M2Chg [6] and 90.day Yields < 90-day Yields 11

1. If CSenti > CSenti [12] and CSenti > CSenti [l l] and Yields >

months ago, then 90 days interest rates will fall.

Yields [4], then rates will rise.

The results of this simple system since 1971 are shown in Table 3.4. 2. If CSenti < CSenti [12] and CSenti < CSenti [ll] and Yields <

Yields [4], then rates will fall.

TABLE 3.4 RESULTS OF MONEY SUPPLY Table 3.5 shows us the results of this model during the period from

AND 90-DAY INTEREST RATES. 4130156 to 4/l/96.

Using consumer sentiment to predict short-term interest rates, the

26.67

Net basis points

average winning position is 20 months and the average losing position is

14.19

Rise in basis points

12.48

Fall in basis points 12 months. The model predicted that rates would start to fall on May 31,

1.91

Average trade 1995, when 3-month rates were at 5.72 percent. As of March 31, 1996,

-1.13

Largest loser this position is profitable by .73 points.

14

Forecasts Let™s now look at how unemployment information-specifically, the

79%

Percent correct

average duration of someone™s unemployment-can help to predict

Low-Term Patterns and Market Tim& for Interest Rates 67

Classical Market Prediction

66

TABLE 3.5 CONSUMER SENTIMENT How can we use unemployment claims to predict short-term interest

AND SHORT-TERM RATES. rates? Our system for timing short-term rates-is based on unemployment

claims. The rules are:

Net basis points 34.19

Rise in basis points 18.56

Fall in basis points 15.63 1. If Claims < Claims [ll] and Claims > Claims [14], then interest

Average move 1.37 rates will rise.

Largest loser -1.34

2. If Claims > Claims [l 11 and Claims < Claims [14], then interest

Trades 25

rates will fall.

Percent correct 84%

This simple model was tested on T-Bill yields in the period from

l/31/56 to 3/31/96 and produced the results shown in Table 3.7.

short-term interest rates. The theory is that the longer someone is with- This simple model does a great job of predicting short-term interest

out a job, the slower the economy will be, and interest rates will drop in rates. Its last trade was predicting a drop in rates on July 31, 1995. On

order to stimulate the economy. long-term interest rates, the same model produces 18.20 points and wins

Our simple model is based on unemployment duration as a factor in 67 percent of its trades, with an average trade profit of .61 point. It is

predicting 90-day T-Bill rates. The rules are: profitable on both long and short moves. Even though this model did not

perform as well as it did on short-term rates, it still shows that unem-

1. If NoJobDur < NoJobDur [3] and Yields > Yields [6], then interest ployment claims are predictive of interest rates.

rates will rise. Our research showed that, when using fundamental-type data, it was

easier to predict short-term (rather than long-term) interest rates on a

2. If NoJobDur > NoJobDur [3] and Yields < Yields [6], then interest

weekly or monthly basis. I think the reason for this is that short-term in-

rates will fall.

terest rates are based on current economic activity, and longer-term rates

are also affected by the perceived effects of future activity.

For the period from 4/30/52 to 3/30/96, this simple model produced

the results shown in Table 3.6.

This model does not work as well as some of our other models, but it

does show that the unemployment duration is predictive of short-term

interest rates.

TABLE 3.7 RESULTS OF UNEMPLOYMENT

TABLE 3.6 RESULTS OF UNEMPLOYMENT

CLAIMS AND SHORT-TERM RATES.

DURATION AND SHORT-TERM RATES.

Net basis points 37.1 2

Net basis points 26.63

Rise in basis points 19.90

Rise in basis points 14.80

Fall in basis points 17.22

11 .a3

Fall in basis points

Average move 1.43

.86

Average move

Largest loser -1.37

-2.14

Largest loser

Forecasts 26

Forecasts 31

Percent correct 77%

Percent cOrrect 5 8%

68 Classical Market Prediction

Long-Term Patterns and Market Timing for Interest Rates 69

A FUNDAMENTAL STOCK MARKET TIMING MODEL

in the prime rate. An ideal interest-rate stock-timing model would com-

bine all three of these factors. We developed a model that used all of these

We have seen how fundamental data can predict interest rates. Let™s now

concepts but only traded the market on the long side. This model was

see how we can use it to predict the Dow Jones Industrials Average tested for the period from 8/11/44 to 4/12/96, using weekly data. During

(DJIA). this time, the DJIA rose about 5,490 points in about 2,640 weeks, or a lit-

We need a model that combines the prime rate, the Federal Reserve tle over 2 points per week. The rules for our model and the results for

(Fed) discount rate, and long bond yields. Our model is inspired by Mar- our test period are shown in Table 3.8.

tin Zweig™s book, Winning on Wall Street. Zweig discuSes how a cut in This long-term time model is an incredible tool for asset allocation. It

the prime rate is bullish for stocks as long as the prime rate is below a

performed as well as buy and hold while being exposed to the market

given level.

only 50 percent of the time.

Another important factor is the Fed discount rate. When the last move The models in this chapter are just a starting point for using funda-

in the Fed discount rate is a cut, that is very bullish for stocks. During the

mental data to predict both interest rates and stock market timing. Many

past 50 years, there have been several false moves in the prime rate. The more relationships can be discovered using this type of fundamental data.

number of false moves drops to almost zero when the change in the prime The data can give you a long-term view of future market direction and

is in the direction of the last Fed move in the discount rate. can be used as a filter in developing trading systems.

The prime rate and discount rate have a strong effect on stock prices,

but so do long-term interest rates. Often, the market will lower rates be-

fore the Fed or the banks make a move. For example, during most of 1995,

the stock market rallied because of a drop in long-term interest rates.

This drop occurred months before the Fed cut rates, which led to a drop

TABLE 3.8 RESULTS OF A FUNDAMENTAL

MARKET TIMING MODEL.

If Prime < 12, Prime Rate is cut, and last Fed discount move was lower, then

buy at close.

If Prime Rate is raised, then go flat.

If Long Bond Yield sets a 42.week low, then buy at close.

Results Summary:

4,141 .17 + I ,367.63 open = 5,508.80

Net points

Trades 21 + open trade

9 0 + open

Win%

197.20

Average trade

-193.71

Maximum drawdown

66.89

Profit factor

1,319

Weeks in market

100.3%

Percent buy and hold

Percent of time ins market 49.99%

Tradine Usine Technical Analvsis 71

1. Calculate a short-term moving average.

2. Calculate a longer-term moving average.

3. When the short-term moving average crosses above the long-term

4 moving average, then buy.

4. When the short-term moving average crosses below the longer-term

moving average, then sell.

Trading U . Moving-average crossover systems work well in trending markets but

.lYSlS

Technical should not be used when the markets are not trending. Critics of techni-

cal analysis will apply a moving-average crossover system to all market

conditions and will optimize the lengths of the moving average to find the

most profitable set of parameters. Anyone who does this will lose money

because many studies, over the years, have shown that the most profitable

moving-average lengths over the past 10 years are most likely to lose

money over the next 3 to 5 years. Let™s look at an example. We optimized

the moving averages for the D-Mark for the period from 2/13/75 to

12/31/89. We then tested them for the period from l/1/90 to 5/31/96. The

Technical analysis, as the term applies to trading, studies the use of price

performance of the best three pairs of moving averages from 2/13/75 to

data or chart patterns to make trading decisions. This chapter explains

12/3 l/89, as well as their performance since 1990, are shown in Table 4.1.

why many people unjustly criticize traditional forms of technical analy-

(A deduction of $50.00 has been made for slippage and commissions.)

sis. It then describes some profitable ways to use technical analysis to de-

The most profitable moving-average pairs during the earlier period

velop mechanical trading strategies.

lost money during later periods. If we had selected our optimized para-

meters based on a robust pair of parameters+ne in which small changes

WHY Is TEcHNKAL ANALYSIS U N J U S T L Y CRITICIZED? in parameters produced little change in performance-we would have

Many of the people who say that technical analysis does not work are

TABLE 4.1 MONEY AVERAGES THEN AND NOW.

either fundamentalists or skeptics who believe that the markets are ran-

Len1 Len2 Net Profit Average Trade Win% Drawdown

dom and cannot be predicted. They point to the fact that the published

rules in many introductory books on technical analysis are not profitable Results from 2/13175 to 12/31/89

when hacktested. This is true; but these simple rules are not the indica-

6 10 $90,137.50 $495.26 43% -$7,187.50

tors used by professional traders. When used correctly, technical analy- 10 20 89,125.OO 521.21 49 -7,475.oo

sis can make money for traders. Let™s take a close look at some examples 10 22 87,750.OO 555.00 49 -8,962.50

of the misuse of computer-based technical analysis.

Results Since 1990

6 10 -1,012.50 -5.53 37 -21,425.oo

Moving Averages

10 20 -13,300.00 -147.78 37 -22,7.50.00

10 22 -29j375.00 -373.44 38 -37,125.oo

As a first example, consider a simple moving-average crossover system.

This system works as follows:

70

72 Classical Market Prediction Trading Using Technical Analysis 73

found that moving-average systems are profitable in many markets. We reversals and take them regardless of current market conditions, we will

also can improve our results if we trade these systems only when a given lose money. When they are used properly, they can be valuable tools.

market is trending. When used correctly, moving-average crossover sys- The classic chart patterns-head and shoulders, triangle, and so on-

are used by many discretionary traders. The problem is, we cannot test

tems are valuable trading tools.

how profitable these patterns are without developing mechanical defini-

tions. The key is to develop mechanical methods for detecting the pat-

Oscillators

terns. Trying to detect market patterns in a mechanical way is the best

The classic oscillators are stochastic and a price-based oscillator that was way to use technical analysis, because we can see how well each method

developed by Welles Wilder (RSI). If we use the simple rules that are of analysis works. We also can tell when each method works best and can

often published for both of these types of oscillators, we will lose money. develop mechanical definitions for these specific conditions.

Let™s now take a look at several technical methods that can be used to

These rules are as follows:

produce very profitable trading systems. These methods are often the

1. Buy when the oscillator crosses above 30 from below 30 core logic behind the best commercial systems on the market,

2. Sell when it crosses below 70 from above 70.

PROFITABLE METHODS BASED ON TECHNICAL ANALYSIS

Why do these rules lose money? First, most people use a fixed-length

stochastic or RX George Lane, who popularized stochastics, adjusts the

Gap Analysis

period used in these indicators to half of the current dominant cycle. Sec-

ond, the standard rules will enter a trade about one-third of a cycle too Gap-based patterns are among the most powerful of all trading patterns.

late. If the cycle is long (30 days or more), being a few days late will still

These patterns are produced based on breaking news or changes in mar-

produce profitable signals. If the cycle is only 10 days, then the market

ket conditions that occurred overnight. The news may turn out to be not

rises for 5 days and falls for 5 days. In this case, being 2 days late on both as important as originally thought, or it may be proven to be wrong. When

the entry and the exit means that, even in a perfect world, the trader is on this happens, the gap closes. If prices continue in the direction of the gap,

the right side of the market for only one day. This is why the rules do not then the news was real and a major move in the direction of the gap could

work. George Lane uses divergence between an oscillator and price and be beginning. Gap patterns are represented in Figure 4.1.

generates most of his trading signals via a cycle-tuned indicator. (The

concept of divergence is discussed later in this chapter.)

Key Reversal Days

1

s?,, here

BUY her

Another classic pattern that will not backtest well when misused involves

key reversal days. These are days on which the market sets a new low and

then closes higher than yesterday™s close. We then generate a buy signal

I-

at yesterday™s high on a stop. If the market sets a new high and closes

below close, we sell at yesterday™s low on a stop. The problem is that key

I-

reversal should be used only to trigger a signal once the market is set up Sell gap

Buy WP

for a given move. For example if the market is overbought and we get a

FIGURE 4.1

bearish key reversal, that would be a good sell signal. If we look at key A standard gap buy-and-sell pattern.

74 Classical Market Prediction Trading Using Technical Analysis 75

We would buy when the market gaps down below the low on the open TABLE 4.3 RESULTS OF RUNAWAY GAPS.

and then crosses above yesterday™s close. We would sell when the mar-

Net profit $80.150.00

ket opens above yesterday™s high and crosses below yesterday™s close. Profit long $17,650.00

We can also have a second set of buy-and-sell signals. If the market gaps Profit short $62,500.00

Trades

up and rallies at some percentage of yesterday™s range above the open, 384

Average trade $208.72

then we should buy. If the market gaps down at some percentage of yes-

Win% 55

terday™s range and then continues to fall, then we should sell. Most gap

Drawdown $14,750.00

patterns are based on the OOPS pattern developed by Larry Williams:

Sell at yesterday™s high on up gaps, and buy at yesterday™s low on down

gaps. Williams also developed the strategy of buying or selling in the di-

rection of the gap if the market continues to move in that direction. Gaps This simple filter improves the performance of runaway gaps in the

are a good tool for day trading or short-term trading (1 to 3 days). Let™s S&P500. This is only one example; there are many other filters that might

look at some examples of using gaps for developing trading models. work as well. The point is, runaway gaps do work. What happens when a

To trade the S&P500, we analyze runaway gaps. We perform our analy- gap fills based on the OOPS pattern? We will sell when we open higher

sis using daily bars and accessing the next open in TradeStation, using a than yesterday™s high on a stop at the high. We will buy when we gap

Dynamic Link Library (DLL) written by Ruggiero Associates. Our rules, lower than yesterday™s low on a stop at yesterday™s low. We will exit on

written in EasyLanguage, are shown in Table 4.2. the close. We tested this system on the S&P500 during the period from

This simple system buys on a stop order and then exits on the close 4/21/82 to 6/28/96. These results are shown in Table 4.5.

without holding any overnight positions. We tested this system on the Note that filling gaps works much better on the long side. Using a

S&P500 in the period from 4/21/82 to 5/31/96, allowing $50.00 for slip- money management stop can help this system. If we use a simple $600.00

page and commissions and a breakout equal to .40. The results are shown money management stop, it greatly improves our results, as shown in

in Table 4.3. Table 4.6.

Runaway gaps produce good results on the short side and substandard Our results show that, when developing a system based on filling a

results on the long side. What happens if we apply a very simple filter gap, most of the problems involve large losing trades. Using a simple

for both up and down gaps? We will require the S&P500 to close lower money management stop, we are able to greatly improve the profitability

than 5 days ago. Our results over the same period, with $50.00 deducted of the OOPS pattern so that it works as well as most high-priced systems

for slippage and commissions, are shown in Table 4.4. costing thousands of dollars.

TABLE 4.4 RESULTS OF RUNAWAY GAPS

WITH SIMPLE FILTER.

TABLE 4.2 RUNAWAY GAPS.

Net profit $86,150.00

Profit long

Inputs: BreakoutC.3); $23,650.00

Profit short $62,500.00

If NextOpewHigh then buy at NextOpen+Brakeout*Average(TrueRange,3) stop;

Trades 237

If NextOpen<Low then sell at NextOpen+Erakeout*Average(TrueRange,3) stop;

Average trade $363.50

ExitLong at close; Win% 59

ExitShort at close; Drawdown -$8,925.00

77

Trading Using Technical Analysis

Classical Market Prediction

76

TABLE 4.7 RESULTS OF CLASSICAL CHANNEL BREAKOUT.

TABLE 4.5 RESULTS OF FILLING THE GAP.

AWL?ge Profit

$73,075.00

Net profit

Market Start Date Net Profit Open Trade Win% Trade Drawdown Factor

$58,750.00

Profit long

Profit short $14,325.00 D-Mark li2lSO $ 5 6 . 6 6 3 . 7 5 $3.286.25 5 0 % $ 5 4 4 . 8 4 -$22,075.00 1.57

588

Trades 1.23

Heating oil l/2/80 21,903.91 5,333.58 44 1 8 2 . 5 3 -19,772.67

$124.28

Average trade 27.763.52 1.17

Lumber 1 /z/so 5,087.68 31 222.11 -49,514.67

56

Win% Swiss Franc l/1/80 71.666.25 5,148.75 1.59

50 6 3 9 . 8 8 -12,858.75

Drawdown -$19.275.00 1 O-year note l/3/83 52,116.OO 5.030.00 47 606.00 -7,116.OO 1.84

T-Bonds l/V80 1.44

59,614.OO 8,499.OO 39 5 2 7 . 5 6 -18.990.00

-91.00

Crude oil l/3/84 59,898.OO 49 777.90 -8.061.00 2.61

l/1/80

Coffee 36 813.21 -29,121.OO 1.52

101,650.63 2,568.38

Gap patterns work in markets with a high daily range, for example, the

S&P500, coffee, and T-Bonds.

something thaw is hard to d-buy at high prices and sell at low prices. In

Breakout Systems

trading, it pays to do the hard thing. Most people can™t, and most people

Breakout systems are among the best methods of technical analysis.

lose money.

There are two major classes of breakout systems: (1) channel breakout We tested the simple channel breakout system on several different mar-

and (2) volatility breakout. These systems work very well in trending kets (allowing $50.00 for slippage and commissions) and our results from

markets and have been used by most of the top traders in the world. Let™s

start date to May 17, 1996, using continuous backadjusted contracts, were

first discuss the channel breakout system, which works as follows: as shown in Table 4.7.

These results show that this simple system is profitable in each mar-

1. Buy at the Highest (High, 20) + 1 point stop.

ket used in our tests. If we had traded one lot of each commodity listed

2. Sell at the Lowest (Low, 20) - 1 point stop. from the start date until May 18, 1996, we would have averaged over

$28,ooO.O0 a year since 1980, and, as of May 17,1996, we would have had

In this form, the system is just a modification of Donchian™s Weekly a combined open equity of over $33,000.00 on all eight commodities.

Rule: Buy at four-week highs and sell at four-week lows. The system When trading a basket of commodities, as in this example, the total draw-

has been discussed in countless publications. It works because it does down of the portfolio is less than the maximum drawdown produced by

the individual commodities, because of the effects of diversification. The

concept of trading a basket of commodities with a channel breakout sys-

TABLE 4.6 RESULTS OF FILLING THE GAP

tem is the heart of the Turtle trading system, but our system still needs

WITH A MONEY MANAGEMENT STOP.

money management rules, filters, and more advanced exit methods that

$145,950.00

Net profit we will discuss in later chapters.

$84,450.00

Profit long The volatility breakout system, made famous by Larry Williams,

$61.500.00

Profit short

works by buying or selling when the market breaks above or below its

588

Trades

open or previous close by a given percentage of the previous day™s range.

$248.21

Average trade

Let™s look at an example-a purchase at a given percentage above today™s

53

Win%

-$8,150,00 open. The rules are:

Drawdown

Trading Usinp. Technical Analvris 79

78 Classical Market Prediction

1. Buy at opening price+.6 x Average (True Range, 3) stop. TABLE 4.9 RULES FOR TREND MODE.

2. Sell at opening price-.6 x Average (True Range, 3) stop. If ADX crcxses above 25, then the market is trending.

1.

2. If ADX crosses below 20, then the market is consolidating.

This simple system buys when the market breaks 60 percent of the av- 3. If ADX crosses below 45 from above, then the market is consolidating.

erage true range above today™s open, and it sells when the market breaks 4. If ADX rises from below 10 on 3 out of 4 days, then the market will start to

60 percent of the 3-day average true range below today™s open. Let™s see trend.

how this simple system works on the T-Bond markets. We tested this sys- 5. If a trend is based on rule 4, it remains in effect until the S-day difference in

tem from September 27, 1979, to May 31, 1996. The results, after de- ADX is less than 0.

ducting $50.00 for slippage and commissions, are shown in Table 4.8.

This system had an equity curve profit similar to the channel breakout

on the D-Mark. It worked very well until September 1994, and then went

into a drawdown until the end of 1995. From January 1, 1996, to May 31,

14-day ADX to detect a trend and consolidation mode are as shown in

1996, this system made over $lO,OOO.OO on one contract.

˜Table 4.9.

Different technical methods work well in different types of markets.

These rules are a modified version of the classic rules for using ADX,

We call these types modes. Developing indicators and systems to identify

which simply say that the market is trending when ADX is above 25. My

modes is one of the most powerful uses of technical analysis. Let™s lxlw

rules handle two conditions: (1) the exhaustion of a trend, and (2) the

discuss trending and the countertrend mode, as well as something called

early detection of a trend.

the breakout mode.

Figure 4.2 shows the T-Bonds market and a trend indicator described

above for the period from l/1/96 to 5/31/96. You can see a trend end,

Market Modes

based on an exhaustion of the trend, when ADX crosses from above

to below 45 during late March and continues downward through the end

The most important market mode involves identifying when a market is

of May 1996. The market started to trend on February 26, 1996, and

trending and will continue to trend or knowing whether a market will

dropped 3 points in less than 6 weeks.

consolidate. One of the best tools for detecting the strength of a trend is

an indicator called Average Directional Movement (ADX), originally de- How do we use this trend mode indicator in a trading system? We start

by using the modified channel breakout system. We have not optimized

veloped by Welles Wilder. Based on my research, the rules for using a

the parameter and will use a 17.day breakout to enter a trade and a lo-

day breakout to exit a trade. We tested this system on T-Bonds, from

9/28/79 to 5/31/96. The results for both the simple system and a system

TABLE 4.8 RESULTS OF VOLITITY filtered with our trend mode detector filter are shown in Table 4.10. (A

BREAKOUT SYSTEM FOR T-BONDS. deduction of $50.00 was made for slippage and commissions.)

This simple trend filter works very well at filtering breakouts. The

Net profit 5163JW3.75

original results were not remarkably profitable, but by entering a break-

$114,856.25

Net profit long

out only when we are in a trend mode, we improved profit, cut drawdown,

Net profit short $49,012.50

1,282 and almost tripled the average trade.

Trades

43

Win% We used T-Bonds data to develop the levels of ADX for our indicator,

Average trade 5127.82

SO they are not optimal for all markets and time frames. Still, the levels

-$22,256.25

Drawdown˜

selected for the example are a good starting point, and the values can be

Tradine Usine Technical Analvsis 81

Classical Market Prediction

optimized by combining this indicator with a channel breakout system

and then optimizing those results.

During an interview with market wizard Linda Raschke, she discussed

something she called a “breakout mode”-a period of low volatility fol-

lowed by a period in which the market is in equilibrium. When the mar-

ket enters the period of equilibrium, there is often an explosive move.

Let™s develop a breakout mode indicator, which can serve as a good

filter for a channel or volatility breakout system.

The market is in equilibrium when technical indicators are confused,

so we begin by building a confusion indicator.

First, we study simple market momentum. We use three different mo-

menta, with durations of 5 periods, 10 periods, and 20 periods, respec-

tively. When these three momentums do not all have the same sign, we

can say they are in confusion. Second, we look at the classic overbought

and oversold indicators: stochastic and RX We will use both a 9-period

and a 14-period SlowK. This is the slow unsmoothed stochastic value.

war Ma7 For us to be in a breakout mode, the values of the two periods must be be-

APT

tween 40 and 60.

FIGURE 4 . 2 T-Bonds, with both the Ruggiero trend mode index and the

Third, we look at volatility because the market often makes large

ADX.

moves after a period of low volatility. To judge periods of low volatility,

we develop a volatility trend indicator. As coded in TradeStation™s Easy-

Language, this indicator works as follows:

Value1 = Volatility(l0);

TABLE 4.10 MODIFIED CHANNEL If value1 = Highest (valuel, 20), then value2 = 1;

BREAKOUT WITH AND WITHOUT OUR

If value1 = Lowest (valuel, 20). then value2 = -1;

TREND FILTERS SHOWN IN TABLE 4.9.

Plot1 (value2, “Volbreak”),

$46,X2.50

Net profit

180

Trades

When this indicator is at -1, the volatility is low enough to sustain a

$258.68

Average trade

40 breakout. Our final indicator will be the efficiency with which the mar-

Win%

-$21,012.50 ket moves. Our indicator, coded using the EasyLanguage Quick Editor, is

Drawdown

as follows:

Results with Trend Filter:

$66,125.00

Net profit

(Close - Close[Len])/summation(abs(Close - Close[ l])),Len)

85

Trades

$777.94

Average trade

Let™s use a period length (Len) of 10. In this case, our efficiency in-

Win%

Drawdown dicator signals a breakout mode when its value is + .20.

82 Classical Market Prediction Trading Using Technical Analysis 83

TABLE 4.12 MODIFIED CHANNEL BREAKOUT WITH

Now we combine these indicators to develop a breakout mode index.

DIFFERENT LENGTH BREAKOUT MODE FILTERS.

Our indicator will simply sum how many of these conditions are true and

then take the average of this simple sum over a given number of days. The Breakout Index Len Net Profit Trades Win% Drawdown

longer the average of this indicator stays at or above 2, the stronger the re-

54% -$l o,ooo.oo

10 $85,537.50 125

sulting move will be. Let™s use the modified channel breakout we dis-

16 73,887.50 91 63 -5,437.50

cussed earlier, and then compare the resulting system using this filter. 18 57,637.50 75 60 -6,937.50

We will buy or sell at a 17-day high or low and will exit on a lo-day high 20 52,912.50 72 60 -6,937.50

26 58,075.OO 57

or low. Let™s take a look at this system for the D-Mark futures. 65 -5,462.50

28 57,600.OO 50 66 -4,800.OO

We tested this simple system on the D-Mark in the period from 2/13/75

30 53,012.50 46 67 -4,050.oo

to 5/31/96, allowing $50.00 for slippage and commissions. The results

are shown in Table 4.11,

We can modify this system so that we will take the trade only when our

breakout mode indicator signals a breakout mode. The system will enter

reduces the profit slightly (small stop) or improves profits and drawdown

trades only when the breakout mode indicator is greater than or equal to

2. (We tested moving averages between 4 and 30 in order to determine our (larger stop). For example, using a $250.00 stop, we made $40,950.00 on

breakout mode indicator.) We found that this filter helps the overall per- 60 trades with only a -.$2,962.50 drawdown. If we had used a $l,OOO.OO

formance of the system over the whole optimization range. Table 4.12 stop, we would have made $55,000.00 on 46 trades, winning 67 percent

of them, with only a -$3,225.00 drawdown. These results show the po-

lists several of the combinations tested, to give an idea of how this filter

performed. tential of this breakout mode filter.

The table shows the power of our breakout mode indicator. We reduced

drawdown, and the winning percentage, in most cases, has risen to over Momentum Precedes Price

60 percent. This simple test shows the potential of a breakout mode index,

One of the most valuable concepts in trading is: Momentum precedes

but how low is the level of adverse movement when a breakout mode fil-

ter is introduced into the system? We will test a breakout mode index price. This concept is one of the major themes in Linda Raschke™s trad-

ing methods. Figure 4.3 shows the Yen from January 1995 to July 1995.

length of 30 for our average, and then add a simple money management

This was the last major top in the Yen. The subgraph shows an oscillator

stop. We range the stop from $200 to $1,000. in steps of 50. Over this

constructed by taking the difference between a 3-day and a IO-day sim-

complete range of parameters, the stop either reduces the drawdown and

ple moving average. Figure 4.3 shows that the momentum makes a lower

high while prices are still making higher highs. Once the top is made,

prices fall about 8.00 full points, or $lO,OOO.OO within the next 5 weeks.

TABLE 4.11 MODIFIED CHANNEL BREAKOUT

SYSTEM FOR D-MARK FUTURES. How can the concept of momentum preceding price be used to develop

a trading system? A simple momentum-precedes-price system, coded in

Net profit $74,587.50

TradeStation™s EasyLanguage, is shown in Table 4.13.

Net profit long $39,325.00

Net profit short $35,262.50 The rules in Table 4.13 say to buy when the oscillator sets a new high,

204

Trades and sell when the oscillator sets a new low. This is mandated because

48

Win% momentum precedes price and a new high or low in an oscillator should

$365.63

Average trade be followed by a new high or low in price. We filtered the oscillator so that

-$15,800.00

Drawdown

we would buy only when the oscillator is above zero and sell only when

84 Classical Market Prediction Trading Using Technical Analvsir 85

TABLE 4.14 MOMENTUM FIRST

RESUtTS ON THE YEN.

Net profit $135,550.00

Profit long 978J712.50

Profit short $56,737.50

Trades 192

Win% 43

Average trade $705.99

Drawdown -$11,437.50

net profit over the period from 8/2/76 to 6/28/96. Using a LookBack rang-

ing from 30 to 34 and a StopLen ranging from 10 to 14, a cluster of com-

binations produced between $125,000.00 and $137,000.00. We selected

a LookBack of 32 and a StopLen of 12 as the parameters to use when

trading the Yen. Our results during the period from 8lUl6 to 6128196 are

shown in Table 4.14.

95 Fsb Mr .bm $4

ap Me”

The results in Table 4.14 show that the concept of momentum preced-

FIGURE 4.3 The 3-10 oscillator made lower highs while the Yen made

ing prices can produce a profitable trading system.

higher highs during early 1995, just before the major top in the Yen.

In this chapter, we have shown that the classic rules for various tech-

nical indicators, although often published in books, do not really work

and are not used by professional traders. For example, George Lane does

the oscillator is below zero. This prevents buying into a bear market and

not use a fixed-length stochastic; he tunes it by using the current domi-

selling in a bull market.

nant cycle. This method substantially improves performance. Other prof-

We then optimized this system on the Yen, using a range of 20 to 40,

itable and reliable methods used by professional traders include gaps,

in steps of 2, for LookBack, and a range of 2 to 20, in steps of 2, for

channel breakout systems, and momentum precedes price.

StopLen. Over these ranges, we found many very profitable combina-

We also showed how markets have many different modes that can be

tions; for example, 56 of 110 tests produced more than $100,000.00 in

detected by combining several different technical indicators into a com-

posite indicator. As composites, the indicators are more predictive than

when used by themselves. They can improve existing trading systems or

TABLE 4.13 MOMENTUM FIRST SYSTEM.

he used to build new ones.

LookBack(32),StopLen(12);

Inputs:

km: Osc(0);

OK = Average(Close,3)-Average(Close,lO);

If OK = Highest(Osc,LookBack) and Osc > 0 then buy at open;

OK = Lowest(Osc.LookBack) and OK < 0 then sell at open;

if

ExitShort at HighesNHighStopLen) Stop:

ExitLong at Lowest(Low,StopLen) Stop;

The Commitment of Traders Report 87

The COT report was first published during the 1970s and was halted

during 1982. From 1983 to November 1990, the report was released

monthly. It was then published twice a month until October 1992, when

5 the current biweekly reporting schedule was adopted.

The COT report is calculated after the market closes each Tuesday.

Because of auditing restraints, two weekly reports are issued on alter-

The Commitment of nating Fridays. These electronically issued reports indicate traders™ po-

sitions for the most recent week. At publication, the data are 3 days old.

Traders Report

HOW DO COMMERCIAL TRADERS WORK?

You may be skeptical about the value of this information, considering

that it is 3 days old (and during earlier years of this data series, was even

2 weeks old). The information is still valuable because it tells how large

commercial traders work. They manage their positions by using a pro-

cess called accumulation and distribution. Because of the size of their

positions, most commercial traders are countertrend traders. They buy

Have you ever wished you knew what position great traders like John

as prices fall and then begin to sell into the rally. Because of this pro-

Henry and Paul Tutor Jones currently had open? With this information,

cess, the COT data could lead the market by 2 weeks or more.

you would, in effect, have access to all of the research they paid millions

Let™s study some examples of this process. Figure 5.1 shows the S&P500

of dollars for, and you would be able to piggyback their trades.

and the net long commercials during the stock market bottom of Novem-

This information is available to you, a few days after the great traders

ber 1994. The commercials began to build their position as the market fell,

have made their moves, in a report called the commirment of traders

and the market did not begin to turn up until they had built their position.

(COT) report, published by the Commodity Futures Trading Commis-

Another example was the large correction in the T-Bonds market, once

sion (CFTC). The COT report tells how many contracts large professional

the commercials had built their short position in late 1995 and early 1996.

traders are currently long or short.

As shown in Figure 5.2, they began to cover their short positions as the

market collapsed. This is additional evidence of how the commercials are

WHAT IS THE COMMITMENT OF TRADERS REPORT? really countertrend traders.

The COT report gives the actual numbers for the three major groups of

USING THE COT DATA TO DEVELOP TRADING SYSTEMS

traders: (1) commercial, (2) noncommercial, and (3) small traders. Com-

mercial traders are hedgers who are trading a given commodity because

Let™s now talk about how to analyze the COT data. The COT report sup-

they have a business that produces it or they need it as an ingredient or

plies the number of contracts each group is long or short, as well as the

material in a product. For instance, Eastman Kodak, which depends on

number of traders in each group. It also publishes the number of spreads

silver, may hedge against a rise in silver prices. Noncommercial traders,

put on by each group, and the net change and percent of open interest

such as commodity funds, are large speculators. Small traders-small

each group holds.

hedgers and speculators-are in a class called nonreportable.

86

The Commitment of Traders Report 89

The first step in using these data is to calculate the net long position

for each group. This is done by subtracting the number of long contracts

from the number of short contracts. When there are more longs than

shorts, the number is positive. When there are more shorts than longs,

the number is negative. The absolute levels of the net long commercials

are not important. What is important is how these levels compare to the

historical normals; for example, in markets like silver, the commercials

have never been net long.

The COT data are predictive because the large traders have a lot of

money and their inside connections give them an edge over the public. If

you were to use the positions of the small traders, the results would be the

opposite, because these traders are normally wrong.

Steven Briese™s newsletter, Bullish Review, uses an index based on

COT data to give trading recommendations. This index is an oscillator

based on the net long positions for a given trading group. We calculate

FIGURE 5.1 The S&P500 weekly versus the net commercial traders.

this oscillator as follows:

Note how the commercials built positions in early 1995, while the

S&P500 was bottoming. They began to liquidate as the S&P500 rallied.

COT Index = 100 x (Current Net-Lowest (Net, N))/(Highest (Net,

N) -Lowest (Net, N))

N is the LookBack and can vary between 1.5 and 4 years.

This indicator is scaled between 0 and 100. When calculating this in-

dicator using the commercials, 0 is the most bearish and 100 is the most

bullish. We generate a buy signal when this index is above 90, and a sell

signal when it is below 10.

When we calculate this indicator using small traders™ data, lower val-

ues are more bullish and higher values are more bearish. Compared to

the indicator calculated using commercials, this is an opposite result be-

cause small traders often buy tops and sell bottoms.

The COT report is a valuable indicator in many markets-T-Bonds,

the S&P500, the agricultural markets such as corn and soybeans, and

markets such as gold and crude oil. On the other hand, the COT report

does not do well in markets such as currencies because, in these markets,

futures represent only a small percentage of the complete market.

How can the COT report be used to develop mechanical trading

95 Feb Mar A;, M™aq JL Jd Ala & oci NO” ˜cec (6 Feb NW Aor Ml” JUI _ systems?

We tested the classic rules for the COT index in several markets and

FIGURE 5.2 Commercials begaLto liquidate their T-Bonds position as

found that they produced unimpressive results. This is not to say that the

bonds rallied during late 1995 and˜early 1996.

COT data are not predictive. Combining the COT index calculations,

88

Classical Market Prediction

90 The Commitment of Traders Report 91

TABLE 5.2 RESULTS FOR COT

using both commercials and small traders, produces some very impressive

COFFEE SYSTEM.

results, and we developed a basic model for using the COT weekly data

in this way. The rules are: Net profit $100,983.75

Trades 24

1. If (COT Index Commercials)[Lagl] > Ctrigger and (COT Index Win%

Small) < Strigger, then buy at open. Average trade ::,207.66

Drawdown -$43,530.00

2. If (COT Index Commercials)[Lagl] < Ctrigger and (COT Index Profit factor 5.88

Small) > Strigger, then buy at open.

Ctrigger is the level of the trigger used to generate buy or sell signals

The high winning percentage, average trade amount, and profit factor

based on the commercials™ version of the COT index. Strigger is the level

confirm that these results are predictive. The drawdown is very high but

for generating buy and sell signals using the small traders™ version of the

COT index. The higher the index is when calculated using the commer- it still shows that the COT data are predictive for coffee.

These two examples are only a start in how to use the COT data as a

cials and the lower it is when using the small traders, the more bullish it

filter for developing high-accuracy trading systems. The goal of this chap-

is. The lower the index is when calculated using the commercials and the

ter has been to provide the basic tools for using the COT report when de-

higher it is when using the small traders, the more bearish these data are

veloping more advanced trading systems.

for that market. Lag1 is the delay in weeks for the values used in our

model. Because the COT index calculated using commercials leads the

market at turning points, the COT data are more predictive if we lag

the index calculated using commercial traders. We tested this mode1 on

T-Bonds, using 30 for Ctrigger, 50 for Strigger, and 1 for Lagl. Our test

used data for the period from 111184 to 5/31/96. After deducting $50.00

for slippage and commissions, the results were as shown in Table 5.1.

This basic model also worked on other commodities. We tested this

model on coffee for the period from 9/7/84 to 5/31/96, and used 55 for

Ctrigger, 35 for Strigger, and 3 for Lagl. Again deducting $50.00 for slip-

page and commissions, the results were as shown in Table 5.2.

TABLE 5.1 RESULTS FOR COT

T-BOND SYSTEM.

$110.412.50

Net profit

Trades 28

Win%

Average trade :z 932.50

4lil56.25

Drawdown

Profit factor 7.45

6

Part Two A Trader™s Guide to

Statist ical Analysis

STATISTICALLY BASED

MARKET PREDICTION

A trader does mt have to be a statistical genius, but he or she should have

a basic understanding of statistics that are descriptive of various proper-

ties of the data being analyzed. This chapter gives an overview of some

of the most important statistical concepts that traders should understand.

These concepts are as follows:

1. Mean, median, and mode.

2. Standard deviation.

3. Types of distributions and their properties.

4. How mean and standard deviation interact.

5. Hypothesis testing.

6. Mean or variance with two or more distributions.

7. Linear correlation.

A trader who has a general idea of these concepts can use statistics to

develop trading systems as well as to test patterns and relationships. Let™s

now discuss each of these concepts in more detail.

95

96 Statistically Based Market Prediction A Trader™s Guide to Statistical Analvsis 97

MEAN, MEDIAN, AND MODE

The mean is another term for the average. The median of a sample is the

middle value, based on order of magnitude. For example, in the number

sequence 1,2,3,4,5,6,7,8,9, the median is 5 because it is surrounded by

four higher values and four lower values. The mode is the most frequently

occurring element.

To clarify the definitions of mean, median, and mode, let™s look at two

I

different cases:

1. 1,2,3,4,5,6,7,8,9,10.

2. 1,2,3,4,5,100,150,200,300.

In the first case, the mean is 5.5 and the median is either 5 or 6. Hence,

in this case, the mean and the median are similar. In the second case, the

I,

I,

median would still be 5, but the mean would be 85. In what is called a nor-

mal distribution, the mean and the median are similar. When the distri-

bution is not normal, the mean and the median can be different.

TYPES OF DISTRIBUTIONS AND THEIR PROPERTIES

moves occur more than they should is why trend-following systems work.

Most standard statistical methods are based on what is called anormal or Figure 6.2 shows the distribution of 5-day returns for the D-Mark from

gaussian distribution. This is the standard bell curve, which is repre- 2113175 to 7/l/96.

sented in the symmetrical chart shown in Figure 6.1, Dr. Bewit Mandelbrot, one of the patriarchs of chaos theory, suggested

A distribution can have several different properties. The first is skew- in 1964 that the capital markets follow a family of distributions he called