TRADING TECHNIQUES The S&P 500 Seasonal Day Trade Is there a particular day of the week within a month that offers the most opportunity? Here s a trading system based on the best days of the week for trading Standard & Poor s 500 futures. Much research has been done on seasonal trading. Originally, the concept derived from the theory that there had to be certain times of the year when it is better to buy and other times when it is better to sell. If you take a commodity and run a simulation in which you pair up all possible entry days with all possible exit days in a given year, you should be able to locate the best days to trade. If you perform this task over several years of data, it is possible some patterns will appear to identify those days with the highest probability of success for both long and short trades. In 1994 I was present at a speech given by statistician Sheldon Knight that shed new light on this old trading strategy. Mark James by William Brower, C.T.A.
Knight runs a company using statistics in processing financial data and has more than 30 years experience in computerized analysis of stocks and futures. During the presentation in 1994, Knight reasoned that seasonal trading approaches were ineffective because commodity futures prices react to government reports and other events based on a monthly calendar, not on a yearly one. Many government reports are released not on the nth day of the calendar year, such as the fifth of every month, but on predetermined positions within a given month. For example, employment data comes out on the first Friday of every month, while money supply figures come out every Thursday. Likewise, the producer price index comes out on the second Wednesday of the month and export sales numbers are released on the fourth Thursday. Armed with this insight, Knight forged the concept of a seasonal trading system based on the day of week in a given month. He developed the K-data timeline method, which, according to him, was almost a successful position trading system, except for the sizable drawdowns. This was later improved upon to include the influence of the actual first notice days. Recalling Knight s work in this area, I decided to use his concepts to develop trading systems. A TRADING SYSTEM I chose to build a system that day trades the Standard & Poor s 500 futures contract. The trading system is based on two concepts. First, the entry rules use some very basic patterns, which I will explain later, and second, I use a unique filter to determine the day of the week in the month for the actual day trades. This custom filter, which is called the Day of the Week in Month (DoWinMo), was created for TradeStation. Figure 1 denotes the listing of the day of the week in a month by two-digit numbers. Each DoWinMo is designated by a two-digit number, and all end in a number between 1 and 5, representing the day of the working week. The list includes numbers as high as 55, because the first 5 stands for the fifth occurrence. All months other than February have at least two days of the week that appear five times. For my entry rules, I created eight different entry patterns and tested each pattern using the DoWinMo as a filter for each of the possible 25 DoWinMo numbers. The eight patterns are as follows: Pattern 1 If tomorrow s open minus 30 points is greater than today s close, then buy at market. Pattern 2 If tomorrow s open plus 30 points is less than today s close, then buy at market. Pattern 3 If tomorrow s open minus 30 points is greater than today s close, then sell at market. Pattern 4 If tomorrow s open plus 30 points is less than today s close, then sell at market. Pattern 5 If tomorrow s open plus 10 points is less than today s low, then buy at today s low stop. Pattern 6 If tomorrow s open minus 20 points is greater than today s high, then sell at today s close stop. DAY OF THE WEEK IN A MONTH DoWinMo listing 11 = 1st Monday 12 = 1st Tuesday 13 = 1st Wednesday 14 = 1st Thursday 15 = 1st Friday 21 = 2nd Monday 22 = 2nd Tuesday 23 = 2nd Wednesday 24 = 2nd Thursday PATTERNS AND FILTERS DoWinMo numbers Pattern 1 11, 15, 23 Pattern 2 12, 15, 25, 42 Pattern 3 14, 41 Pattern 4 23 25 = 2nd Friday 31 = 3rd Monday 32 = 3rd Tuesday 33 = 3rd Wednesday 34 = 3rd Thursday 35 = 3rd Friday 41 = 4th Monday 42 = 4th Tuesday 43 = 4th Wednesday 44 = 4th Thursday 45 = 4th Friday 51 = 5th Monday 52 = 5th Tuesday 53 = 5th Wednesday 54 = 5th Thursday 55 = 5th Friday FIGURE 1: Each DoWinMo is designated by a two-digit number, and all end in a number between 1 and 5, representing the day of the working week. The list includes numbers as high as 55, because the first 5 stands for the fifth occurrence. All months other than February have at least two days of the week that appear five times. Pattern 5 12, 15, 21, 33, 35, 42, >50 Pattern 6 21, 32, 41, 43, >50 Pattern 7 22, 33 Pattern 8 14, 33, 34, 42 FIGURE 2: Here are the successful patterns and the DoWinMo numbers used as filters. Since there are very few trades for the DoWinMo numbers from 51 through 55, they were combined into one test filter designated by the >50. IN-SAMPLE AND OUT-OF-SAMPLE TEST RESULTS Pattern 1 Pattern 2 Patttern 3 Pattern 4 In Out of In Out of In Out of In Out of sample sample sample sample sample sample sample sample Net profit 11,700 24,800 44,190 57,025 9,175 16,325 14,950 3,975 Total trades 67 73 130 139 58 51 25 24 % winners 48 56 62 63 52 61 60 50 Avg trade 175 340 340 410 158 320 598 166 Largest win 2,500 4,300 10,225 5,000 6,600 3,700 5,450 1,850 Largest loss 1,575 4,725 3,225 5,050 2,500 2,850 1,925 2,425 Pattern 5 Pattern 6 Patttern 7 Pattern 8 In Out of In Out of In Out of In Out of sample sample sample sample sample sample sample sample Net profit 23,775 17,175 12,675 13,825 8,025 4,225 3,450 11,125 Total trades 39 33 28 21 17 14 10 23 % winners 67 79 79 62 77 79 60 70 Avg trade 610 520 453 658 472 302 345 484 Largest win 4,025 3,625 2,250 6,850 1,900 1,825 1,200 3,675 Largest loss 1,300 2,500 825 2,100 525 1,500 1,125 1,025 FIGURE 3: Here are the test results for each of the eight patterns. Of the eight, patterns 1, 3 and 4 are suspect because they have such large discrepancies in the average trade from in-sample to out-of-sample. Patterns 1 and 3 improve in the out-of-sample test in this category, but that does not justify keeping them in the system, and so patterns 1, 3 and 4 are tossed out. Pattern 7 If tomorrow s open minus 40 points is greater than today s close, then buy at today s low limit. Pattern 8 If tomorrow s open plus 70 points is less than today s close, then sell at today s high limit.
Open + 30 points Open + 10 points tomorrow s open + 30 points < today s close buy at today s market tomorrow s open + 10 points < today s low buy at today s low stop PATTERN 2 IN CHART FORM PATTERN 5 IN CHART FORM I created a continuous contract of the S&P 500 futures (using the back-adjusted, forward-roll method) from April 1982 to March 1996 using the database available from Genesis Financial Data Services. To avoid curve-fitting, I divided the data into two test periods. The in-sample test period was from April 1982 through December 1989 and the out-of-sample test period was from January 1990 through March 1996. I tested each pattern separately, filtering trades by the DoWinMo numbers for both the in-sample and the out-ofsample datasets. Test results that deviated significantly in winning percentage and/or average profit per trade (from the sample period to the out-of-sample period) disqualified the pattern on that DoWinMo day. I kept the patterns simple because the DoWinMo filter is powerful. Each pattern is a day trade using a simple market-onclose (MOC) exit. Further, each pattern uses the open tomorrow command. Since the trading signals for daily bars must be PATTERN 6 IN CHART FORM Open - 20 points tomorrow s open - 20 points > today s high sell at today s close stop Open - 40 points Open + 70 points tomorrow s open - 40 points > today s close buy at today s low limit if: tomorrow s open + 70 points < today s close then: sell at today s high limit PATTERN 7 IN CHART FORM PATTERN 8 IN CHART FORM FIGURE 4: FINAL PATTERNS IN CHART FORM. Here are the final patterns, shown in chart form. Note that patterns 1, 3 and 4 are dropped.
S&P 500 DoWinMo SEASONAL In Out of Combined sample sample Net profit 82,240 97,250 179,490 Total trades 200 212 412 % winners 66 66 66 Avg. trade 411 459 436 Avg. winner 923 1,116 1,022 Avg. loser 582 819 704 Largest win 10,225 6,850 10,225 Largest loss 3,225 5,050 5,050 Consec. win 8 15 15 Consec. loss 3 4 4 Intraday drawdn. 5,175 9,050 9,050 Profit factor 3.08 2.65 2.82 Slippage & comm. 100 100 100 FIGURE 5: Here are some of the system s performance results. The high percentage of winning trades, high average trade and respectable profit factor are the system s strong points. 180000 160000 140000 120000 100000 80000 60000 40000 20000 0 S&P 500 DoWinMo seasonal net profit and max intraday drawdown 5/28/82 1/31/83 9/30/83 5/31/84 1/31/85 9/30/85 5/30/86 1/30/87 9/30/87 5/31/88 1/31/89 9/29/89 5/31/90 1/31/91 9/30/91 5/29/92 1/29/93 9/30/93 5/31/94 1/31/95 9/29/95 FIGURE 6: SYSTEM EQUITY CURVE. Here is a simulated track record for trading the system since 1982. The lower line is the maximum intraday drawdown. provided the day before the trade is taken, all references to DoWinMo numbers identify the day before the trade. There are no restrictions on how many trades may take place in a day, but only one trade is allowed in the same direction at the same time. No stops of any sort are used. Figure 2 represents the successful patterns and the DoWinMo numbers are used as filters. Since there are very few trades for the DoWinMo numbers from 51 through 55, they were combined into one test filter designated by >50. PATTERN-BY-PATTERN RESULTS Figure 3 lists the in-sample and the out-of-sample test results for each of the eight patterns. Of the eight, patterns 1, 3 and 4 are suspect because they have such large discrepancies in the average trade from in-sample to out-of-sample. Patterns 1 and 3 actually improve in the out-of-sample test in this category, but that does not justify keeping them in the system, and so patterns 1, 3 and 4 are tossed out of the system. Figure 4 displays the final patterns in a chart form. When testing a system, adhering to strict rules is important to avoid the Texas sharpshooter syndrome, well known to statisticians. Say a sharpshooter is blindfolded and allowed to fire 1,000 rounds at the side of a barn some 100 feet away. After doing so, he takes off the blindfold and approaches the barn, locating the area with the highest concentration of bullet holes. Then he draws bull s-eye rings around that spot. To an observer who happens on the scene after the rounds are shot and the bull s-eye rings are drawn, an erroneous conclusion may be drawn about the sharpshooter s accuracy based on the spread of the bullet holes and the concentration at the bull s-eye location. In reality, of course, the conclusion has no validity. This is an easy trap to fall into when testing a trading system, FIGURE 7: JUNE S&P FUTURES. The system enters (indicated by the arrow on the left-hand side of the bar) based on the pattern and exits at the close. Here are two trades during April. FIGURE 8: JUNE S&P FUTURES. Here are six trades during May.
since all this historical testing is akin to drawing the bull s-eye rings around the bullet holes after the fact. We need some means to improve the likelihood that our conclusion is not mere illusion. Requiring comparable results in both the in-sample and out-of-sample period is one way to do this. The best test, however, is to see the system work on real-time data. THE LONG AND THE SHORT I have listed some of the system s performance results in Figure 5 and an equity curve in Figure 6. The high percentage of winning trades (66%), high average trade ($436) and respectable profit factor (2.82) are the system s strong points. The $9,050 maximum intraday drawdown (MIDD) is actually quite good in relation to the net profit. My rule of thumb is that a net profit to MIDD ratio above 10 is very good; our ratio is almost 20. The system had most of the profit on the long side ($138,440), which is not surprising, considering we threw out two sell patterns and one buy pattern. The short side made $41,050 and was 68% profitable with an average trade of $500. Interestingly, the largest MIDD occurred on the long side. The short side only reached $3,175. Figures 7 and 8 present some recent trades. The system s biggest weakness might be that some would have trouble with the largest losing trade ($5,050). Most intraday traders like to keep the stops closer to $1,000, but money management stops only serve to degrade the performance of the system. A $1,000 stop cuts the combined net profits to less than $120,000 and increases the drawdown to $11,600. I ran the optimizer and found that larger stops, in general, improved performance. Money management stops tend to work better with systems with a low percentage of winning trades. IN SUMMARY Using the Day of Week in Month can be a powerful filter for pattern-based day-trading systems, and using a split dataset to generate in-sample and out-of-sample tests can help prevent the Texas sharpshooter syndrome. A major question remains as to how well this system will work in real-time trading. Of course, it is still possible this entire exercise is a sophisticated curve-fit. However, in the June 1996 S&P contract, and since completing system testing and development, I have had six out of six winning trades for a net profit of more than $17,000 on one lot. The concept seems to be working. William Brower, CTA is president of Inside Edge Systems, publisher of TS Express newsletter. He trains TradeStation users, writes programs and tests systems. S&C