Automatic Pattern Search

Good post with nice ideas about using statistics. Excellent!

I think you do not need > 9999 bars in search range. I think search range is confusing label. It has do do with the most recent trade of found patterns and it must lie in the search range.

"This setting does not affect the number of price bars, or data file records, used when calculating the performance of patterns. APS uses all the available records in a data file to back-test patterns during a search and determine whether they fulfill the criteria specified on the workspace."

I was also confused but read this explanation first:

Using APS FAQ

Thanks, I read the help file and it makes more sense now. In a file containing one year of 15 minute forex data you have more than 60,000 bars so to ensure that the patterns you search for occur in the final month of the data a setting of 5000 for the search range would be required - 500 is possibly too restrictive for high frequency data but fine for daily data. I understand now that the entire file is used to search for patterns independent of the search range. Will continue playing around with the stats ideas to try and produce something useful....

How come you have more than 60,000 bars in one year of 15 minute bars?
There is 96 bars in a day. My provider gives me 24,934 bars for year to date.
Am I missing something?

Sorry I got my numbers wrong - I had 65000 bars in a file that had a couple of years data in it. EURUSD 15 min for 2007 gives me 24542 bars using the Metatrader platform. Thanks for pointing that out. So would require a search range of just over 2000 bars to cover the last month of data.

Good question. I think the answer is (probably) that on the contrary there are some disadvantages to that like the possibility of your stop getting hit during slow markets.

I will take a closer look at your probability equations and provide some comments. I also think that for about 25,000 bars a search range in the order of 4,000 bars is good enough. Maybe 2,000 bars is a bit restrictive.

As far as considering delay patterns first and then the ones without delay to calculate probabilities this is because I don't want to wait to execute the trade. However, I will run some tests with the reverse conditions as Alex pointed out.

mballagan, I think with just a few patterns it is very difficult to come up with a statistically significant system. If you cannot run an extended search due to execution time limitations try combining different timeframes. For example:

- 1 hour determines the trend
- 30 min is used for confirmation
- 15 min (I prefer 20 min by the way) is used for trade entry along the direction of the trend.

To me, this is the real power of APS, i.e. the ability to run different timeframes and configure different systems in just a few hours.

By the way, Michael Harris talks about timeframe combinations in his new book I don't recall the chapter but it was one of the best ideas I've ever come across.

Ron

Hi,

I am trying the latest APS demo version with the following parameters:

QQQQ last 10 years of daily bars. Total 2402 bars.
Split the file in two parts 80% for pattern search and 20% for forward testing:
So file #1 has 1921 bars, file #2 - 481 bars.

T/S file: 7,7
Exits: %
Inputs: Open
Delay: Off
Profitable for long: > 0.66
Profitable for short: > 0.66
Trades: > 20
Max conseq losers: < 4
(By the way, in my demo it says "conCecutive", which is strange for 7 y.o. product)

Search range: 500
Extended search: ON
Incl patts w/equal values: Off

2 hours later it found 39 patterns

Then I backtest each pattern using my second data file.

18 patterns remain profitable,
14 turned into losers
7 become even

So overall, the whole system become neutral.

I'd like to hear your comments





Further testing results:

same parameters as in first test, but Trades: > 30

(so now my parameters are the same as in the article Tradingpatterns.com Michael Harris article

8 patterns found

in forward test: 6 turned into losers, 1 remained profitable and 1 even

Ten years of data cannot be just 2402 bars. You are probably missing about 150 data points.

Anyway, your results do not agree with the results shown here:

Tradingpatterns.com Michael Harris article

Concecutive I haven't even noticed that but a google search shows some people spell it that way. Here's an example:

WikiAnswers - What is the most concecutive home runs

I will run some QQQQ tests and report back.

I just did the same test with the latest APS demo.

I used the QQQQ data file that came with the demo. I created 2 different files in 2 different directories as follows:

In sample: 19920409 - 20060222

out-of-sample: 20060224 - 20080222

Results of extended search with the same parameters you used:

48 patterns found (not 8 as you say).

In out-of-sample testing:

33 winners, 2 of them with 100% success rate, 9 of them have a profit factor > 1.70. Excellent results!

15 losers: 7 of them only marginally with success rate 50%.

This means you are left with at least 33 patterns to play with.

Results for regular search:

9 patterns found.

6 Winners

3 losers

You must have done something fundamentally wrong to get the results you got.