NuclearPhynance Archive

The In-Sample data that I used initially was from 1997---2003 and out of Sample was 2003---2007



But we found out that the Ticks were NOT good in ES until late 2000. Also, AG Commodities went electronic in August 2006. So I had to use PIT data and then combine with electronic data once they went electronic. So Lot of Care was taken so I do not CURVE FIT.

Hi tptspecial,



I gather you are aware that the CME began aggregating ticks a few years ago. In addition, sammus' comments are spot on. Basically you are working with tick bars - fair enough but I gather that you are also aware that even small variations in where you start your tick count can at times cause wild variations in observed bar patterns. It becomes a question of "Where is One?"



No problemo at all using tick-bars or volume-bars (or the type of bars described in the Olsen paper for FX), but you need to be careful if you are searching for bar patterns - unless of course I have misunderstood your approach. Otherwise a sensitivity analysis would be in order.

hankscorpio,



I am NOT working with tick bars, even though I did start off with that initially. I wanted to make it dynamic and so If I were to go with constant tick bars I would have to fix the bar size initially. We all know bar size fixed for 1997----2003 will be very different now because of higher liquidity and volatility. So I decided against doing Tick bars. Same reason for Volume Bars.



Can you please expand on Sensitivity Analysis.....Is there any particular way of doing it....I decided against going the Bootstrap route as In my case, I am not sure, it will help me.

Right, so by [i]"Here is what I do: Say, There are 100000 ticks per day. depending on the price action, I aggregate them into say, 30 Bars. The number of bars formed vary each day. After aggregrating, I calculate various stats such as"[/i], you were not implying simple bog standard tick aggregation, but something a little more exotic.



A very simple test would be to just move the starting datum - i.e., where you commence the tick aggregation from - a few ticks either way, and determine whether this affects your pattern detection (I'm hoping you are not looking for Candlestick patterns here with your dynamic tick bars Wink - if you are then I reckon you're toast)



If moving the initial tick aggregation point a few ticks either way leads to a significantly different performance, then you have an issue.



Of course, you also must keep in mind that some exchanges already aggregate ticks. Have you accounted for this via your dynamic tick aggregation method?



Do you reset the datum on a daily basis? What about near 24-hour markets (ES, YM, NQ, etc), or no resets so it all depends on your arbitrary starting point?

Yes, NOT standard tick aggregation. I dont know if I can call it exotic, but different. All the model I develop, is strictly, Intraday...I.e...8:30 to 3 PM CDT. SO the aggregation starts at 8:30 and ends at 3 PM CDT.



OH No candle Stick patterns :) Tick aggregation at exchanges does not really affect much on what I am doing.

I think Hank probably meant the sensitivity of your results when you start your tick aggregation at 9am to 2:30pm instead of starting at 8:30 to 3pm or some sort of variant.



Another is to dig through the data and see why you have this edge?  Can it be explained from the empirical distributions and such?  Is there a high serial correlation with the aggregated ticks by construction or is it also present in the data? There are definitely lots of questions that you can ask and try to answer yourself.



And obviously, the easiest way to see if it works is just pay the tuition and see what happens.  Start out as small as possibe, capture the feedback, and analyze the results.



Btw, when I meant segmented in and out sample, I meant I would have divided the data randomly instead of choosing 1997-2003 as in and 2003-2007 as out, I would have done something like 1997-1999 and 2002-2004 as in sample, 1999-2002 and 2004-2007 as out sample...

Tabris,



Thanks. First, I dont think I have an edge yet. Because If I account for slippage and other things, I might just get about a tick extra through this method. I have to work on a different entry and exit methods. As you know, Just knowing that 62% of time next bar is going up is useful for GREY BOX approach, But to make it completely Black BOX is so tough.



The Pattern that I am talking about came from practical experience while watching the Order Flow on TT MD TRADER. So I kind of already knew about the pattern and so just tested it out over historical data. It is NOT something that was datamined.



Even the Bosses feel the same thing about this. So if I am ever given an opportunity to GREY box this, I will definitely use this information. I am given another week to see where I am going with this system.



As an extra information, I added an extra filter, to the basic pattern, and now the observations are reduced to about 120  in 4 years, But % increased dramatically, to 73% each. It works only in ES and ER2.



Whole weekend is gone doing this.......

Well back to answering some question after my own trading system boondoggle for the last few days...



While not perfect and of course not knowing all the details here is what I would do as a minimal first cut:



1. As mentioned before look at the next bar distribution after your data xform (tick to dynamic bars) with and without your other pattern.



2. At a minimum you should have statistical sig using the plain old t-test between the two distributions. You can also look to doing a K-S test of the two dist. In general I have seen that t-test by and large follow what bootstraps do when eval trading systems like this. Also, there is a good chance you have made the data "more normal" due to your transformation



3. You mentioned you get perhaps 30 bars per day on average using your transform. I would then create calendar time bars, eg n-minute bars, that roughly equal the same number of your tick-time bars over the sample. I would then rerun the tests and pattern to see the difference in return distribution properties. This will tell you if the tick-time is helping or not.



After you perform these tests and the pattern/system still looks interesting there is more one can do before putting capital-at-risk.

Thank You svquant,



Good Stuff..I Hope I am as good in statistics as you seem to be. Learnt a lot. It is running in simulation now. We will see what happens.



Thanks

Nice one tptspecial.



BTW, if we don't hear from you, we'll know that you are well on your way to becoming a Brazillionaire Dollar Bill