Yesterday I ended the day with a large long position in ES. After a nice up move like we had I like to hold overnite to see if it's likely to keep going in the same direction. My models indicated today wasn't likely to have a large range or continue in yesterday's direction. They basically said to step aside for today and forget about trading. They didn't help me figure out what to do with my position so I did a little test with Tradestation. I have 20 years of daily data in the SP so I can check just about any day against a pattern. What I did was look at the pattern for the day (using daily data) and try to see if it had any bias. I used 3 identifiers that I thought described yesterday. 1). The close was greater than any of the past 20 days. 2). The range was more than 1.75 times larger than the previous day. 3). The open was lower than the close from 2 days ago. I tested the pattern against the past 20 years. I found 62 days that matched. Of the days 63% were down days and the average down day was 60% greater than a up day. Using this info. I decided to exit my long position in overnite trading. So far it looks like it was a pretty good decision. I've used the pattern matching quite a few times over the past 5 years and found it to be worth taking a look at when I need a clue.
Thanks for the great post acrary. That is an excellent use of daily values for getting a probability of the future. I have been searching around for a good source of daily and intraday data for doing research and so far Tradestation looks like the best.
I certainly do not have capability and ability to do something similar to what you did in your research . I use 5, 15 and 30 min charts and if there is a divergence at the close of a day, my bias is to the reverse of a today's trend . Or exiting position, expecting a reversal or at least a pullback. Walter
Interesting post again, acrary ... I just tried to find out if this result is statistical significant. Unfortunately I couldn't find an approach (e.g. t-test), since stddev is unknown. Do you have any idea how to calculate ? Dierk
You could use a t-test for the % Up since up days and down days follow an approximate normal distribution. I did a t-test to make sure it could be done. I imported all the daily data into a excel spreadsheet. I then marked a one for a up day and a zero for a down or unchanged day. Then I grouped the days into 62 days each and summed the up days for each 62 day period. There were 84 groups in my data. The mean number of up days was 32.22 with a standard deviation of 3.11. In my pattern test I had 23 up days in 62 samples so that's what we're testing against. The critical value for 61 degrees of freedom at .05 is 1.999. In plugging in the values I get: T = (32 - 23)/ (3.11/sqr(62)) T = 9/.3949 T = 22.79 which is way past the critical value of 1.99 so this is highly significant. It wouldn't take much work to write a quick little program to find the mean and stddev from the daily data and do the T-test if you wanted to do this every day.
just came to my mind. do not know if anyone could be interested. step 1: define the pattern ingredients in this case it is the relation of the last close to the previous 20 closes, the relation of the last open to the close on t-2 and the relation of today's range to yesterday's range. step 2: define the criteria ttest below 0.05 (i use this way to look at ttest - which is: the smaller the better) hitratio above 60% payoffratio (=avg gain to avg looser) above 1.3 step 3: trade if criteria is met now this whole process could be done several times, thus yielding different patterns with different ingredients. the whole thing would be a selflearning pattern recognition tool. walter i have a spreadsheet, which could enable you to test things of the described nature, but it is quite big. if you PM your email i will send it over. it will need some explanation, but is rather intuitive i guess. peace