Sneak Peek: Algo Trade

My trading partner and I have been working on a pattern recognition algorithm as a side project for over a year now.  We’ve been chipping away at the various problems and bugs associated with new software, but we’re definitely getting closer to completion.  For those who haven’t read about it here’s an over simplified abstraction of what the software does from a previous post:

To visualize how the software works, imagine drawing a line graph of every single trading day (intraday) of the Dow on a separate sheet of paper.  Now take the humongous stack of paper and press the it up against a sunny window like a kid with tracing paper.  Even if the sun could shine through the stack of paper, you’d still have an unrecognizable blob, right?  Wrong.  Our software can intelligently processes HUGE amounts of data and reveal the true nature behind the noise — even if the nature of the data is in fact actual random noise (in which case it will generate no patterns, smart huh?).

There were two major hurdles involved.  One, adapting the software from its native application to financial markets; two, finding the correct way to categorize data so that we’re comparing stuff that’s likely to give us a meaningful result.   The first problem was purely technical while the second is a mixture of trading art and logic.  The software will attempt to extract patterns from any data sets with a sufficiently large sample size.  We’ve focused on comparing intraday charts that have anecdotal similarities with one another.  We call these anecdotal similarities correlating factors.  We plug all the days data that match our correlating factor criteria and let the software determine if there is a dominant pattern to be extracted.  Without getting too ahead of myself, I think we have a fairly objective and robust way of categorizing days based on a mix of the generic technical parameters.  We also have bulletproof software that does the heavy lifting of crunching all the actual data.  I’m not ready to discuss exactly how we determine our correlating factors or what algos we’re using to do the pattern recognition, but I will write more about it in the future.

Check out how the algo performed today.  This was a fairly generic study using correlating factor dates back to 2001.  Notice that the primary pattern (72% of energy) flagged a lower low at 10:30 followed by a rally into the afternoon session.  The secondary pattern (17% of energy) was very flat in the morning session and rallied into the late afternoon.  What actually transpired was a kind of hybrid of both patterns.  We rallied off the lows just after 11:00 AM and kept rallying until late in the session.

Wouldn’t it have been great to know that past sessions similar to this bounced off the morning lows and rallied to new highs into the afternoon?  Would it be valuable to know this ahead of the opening bell?  If you knew this ahead of time would you have pushed your shorts at 11:00AM or would you have covered?  Could have been dumb luck, but none of the patterns were bearish at all today.

Can we extrapolate that every day like this MUST bounce after 10:30AM and make new highs?  Hell no.  But it was pretty damn close today — and my hypothesis is that there are plenty more tradable patterns just like it waiting to be discovered.

Report generated using S&P 500 Futures on November 14, 2011 (for trading on November 15, 2011):

Note: “bearish sentiment curve” indicates that the days used to generate these charts were instances where the market gapped down off the open.  In today’s case we used these results because the market obviously gapped down off the open (this is an example of one of the many correlating factors used to categorize the data).

Actual trading of S&P 500 Futures on November 15, 2011:

Disclaimer: Nothing on this site should ever be considered to be advice, research or an invitation to buy or sell any securities, please click here for a full disclaimer.

blog comments powered by Disqus
Dynamichedge Blog