Automated Trading for a Living

gmst · May 4, 2013

Quote from Allistah:

Has anyone used Tradestation's optimizer in addition to their walk forward optimizer before? Sounds like it is a whole lot better than having nothing.

A big problem with all the retail trading software out there is that the quality check is horrible. Their main aim is to roll out as many features as so that they can have more tick marks for features compared to their competitors. Whether the feature they are rolling out works as expected or not is the last of their consideration. The unsaid philosophy goes like this:

"Just roll out a feature (if it has multiple bugs, its ok). Let customers find out those bugs and tell us about them. Then we will work on them."

Sounds hard to believe????? Check out below link:
http://www.multicharts.com/traders-blog/?p=923

Go to the section BUGS FIXED -> SCRIPTS/CALCULATIONS. You will find below:

Out-of-sample intervals would get crossed during Walk-Forward Optimization.

MC introduced walk forward optimization more than a year ago, and they just admitted it was buggy. To be balanced - Multicharts is one of the better retail trading software around. Tradestation/Ninja are much in a much bad state.

Still, as a newbie retail guy, you have to go for some software because its not possible for most to write their own infrastructure. I would say choose something based on in-depth research. You may be surprised how much all the marketing and sleek websites you see is smoke once you dig a bit deeper.

Allistah · May 4, 2013

Yeah I don't have the skills or time to write my own software. I am a programmer but it's not been in the sense of writing software. Most of my skills are in writing test scripts and other things like that. All of the automated trade stuff is extremely simple for me.

So sheesh, what the hell is someone like me to do then? All I want to do is come up with some simple strategies and make sure that they're tested properly and try to run them live. Is there not an option for me besides writing my own software? There has got to be something for me.

gmst · May 4, 2013

Quote from Allistah:

Yeah I don't have the skills or time to write my own software. I am a programmer but it's not been in the sense of writing software. Most of my skills are in writing test scripts and other things like that. All of the automated trade stuff is extremely simple for me.

So sheesh, what the hell is someone like me to do then? All I want to do is come up with some simple strategies and make sure that they're tested properly and try to run them live. Is there not an option for me besides writing my own software? There has got to be something for me.

My main message is - if you are using the most basic and primitive features of these retail programs, chances are things will work as expected. However, if you want to use any fancy features of this program, it will be beneficial to be skeptical that the promised thing works as advertised. Unfortunately, bugs could be subtle. But that is the reality of today's retail software.

The best way to verify correctness of the results that you get in a program like MC/Amibroker is to model your trading strategy in a more generic program like excel/MATLAB and match trade for trade. If there is an issue, investigate. You will learn a lot of unwritten things and limitations about your trading software through this exercise.

NetTecture · May 4, 2013

Quote from ofthomas:

no, he is not... in concept yes, but in actuality he is far from it... you should not compare based on what you assume actual banks do...

he has re-invented a lot of things and if you understood grids you would see the stress points that will be ahead... the nodes checking for work vs. being scheduled the work and lacking automatic job restart... as you scale it becomes an issue... at least he is in the right path with booting off iscsi (am sure he will figure out eventually being able to provision nodes and rebuild nodes in less than 5 minutes .... an institution will just go out and create a CBA for symphony and deploy the grid foundation based on that... not create it from scratch...

dont take me wrong, he has invested a lot of time and effort and created something to suit his needs but scaling will prove interesting as the workload increases...

Agree, but 5 minutes are WAY too high. We have automatic job restart on the list. And node redeploy is already solved, too - standard for windows ISCSI boot devices, to have one master image, and differentials.

All that is off the software side, though - I could talk about using SCCM, SCOM etc. to watch things. We do not do them as we have less than a dozen nodes, and that "is it". When the need arises, all the foundations are in.

NetTecture · May 4, 2013

Quote from ofthomas:

you should be looking at entry level arrays that will enable you to present to ZFS ... keep in mind not all your data at rest is active all the time, so placing everything on SSD is not advantageous, best to tier it... imo, which is restricted to what you are saying here without knowing actual vendors and physical setup...

I hate to tell you but right now we keep the data on a non-ssd setup (basically it is dumped on a file server that does other things). Unless when I do OLAP type queries neither the data store nor the SQL databases are particularly busy. IO on the SQL side is rarely above 10mb/second, and that is RARE. It is not like I am a moron, you know.

The SSD is a transparent raid controller cache - this IS tiered. Helps quite a lot with spikes..... even on the OLAP side you do not use all the data all the time.

Allistah · May 4, 2013

Are there any other tips that you would give to someone like me? I work full time and I want to transition over to trading for a living. Since I have been in the software industry for 21 years, I have a lot of experience with hardware and software.

What about tips on making strategies? I've heard that the more simple the better. Any tips around that?

Are you guys making this work? Fully automated trading?

NetTecture · May 4, 2013

Quote from gmst:

My main message is - if you are using the most basic and primitive features of these retail programs, chances are things will work as expected. However, if you want to use any fancy features of this program, it will be beneficial to be skeptical that the promised thing works as advertised. Unfortunately, bugs could be subtle. But that is the reality of today's retail software.

The best way to verify correctness of the results that you get in a program like MC/Amibroker is to model your trading strategy in a more generic program like excel/MATLAB and match trade for trade. If there is an issue, investigate. You will learn a lot of unwritten things and limitations about your trading software through this exercise.

Yes. Some ninja examples:

* Stock filles are never out of the bar. You say "2 tick slippage". Cross is at extreme of the bar - fill gets 0 slippage. Not that anyone tells you. You need to program a cutom execution for that.
* There is no sensible way to analyze optimizations visually or compare them to earlier runs. That is brutal - you have to go throug them in excep (Export/Import) to see how the statistics of trades changes when you do a program chance. We REGULARLY do compare executions between backtests when I go and tell a programmer to optimize the code - code optimization should result in better runtime, cleaner code, but not changed behavior.

Plus until recently when it was downloading data and there was an error for a day - well, 0 trades. Nice. Not that you would get an error.

Another REALLY stupid thing:

When your strategies goes exceptional (i.e. throws an exception) it is stopped. The log has "has thrown an exception". They obviously do not consider it proper to provide BASIC debugging information there - which is standard for .NET.

And another REALLY funny one:

By default they write log files with every execution. That is quite nice..... until you run a really nice backtest and realize you just filled up your hard disc with 60 or 70 gb of logfiles that noone will ever look at any way. Happened here

This all is "we make it, we don't use it" type of buggy. Oh, and Ninja, the culmination of Retail software, is STILL using .NET 2.0 - which is way way way outdated and WAS way way way outdated when they released it.

Allistah · May 4, 2013

Quote from NetTecture:

Agree, but 5 minutes are WAY too high. We have automatic job restart on the list. And node redeploy is already solved, too - standard for windows ISCSI boot devices, to have one master image, and differentials.

So what would you recommend for a time frame then? I'm more of someone that would like to trade on a smaller timeframe.

What about strategies? I can code some crazy stuff but I've heard that the more simple they are, the better.

I just have to figure out how to make this work. Must. Make. It. Work. :-)

NetTecture · May 4, 2013

Quote from Allistah:

So what would you recommend for a time frame then? I'm more of someone that would like to trade on a smaller timeframe.

What about strategies? I can code some crazy stuff but I've heard that the more simple they are, the better.

I just have to figure out how to make this work. Must. Make. It. Work.

The 5 minute was for a node install, not strategy. Use whatever you can make work. If you need to go retail, try them out. Just make sure you know what you get into. I have someone I sometimes talk to who is making a lot of work on Ninja - but his order management code alone (to work around ninja issues) is larger than most strats (1500 lines, iirc). It just is not an easy undertaking, and no - the software wont help. And you will need some seriously good data.

If you plan doing more than simple instrument strats most retail is dead anyway. It is hard to make spreads and you can outright forget option strategies. Just not doable.

QuantWizard · May 4, 2013

Quote from Allistah:

Has anyone used Tradestation's optimizer in addition to their walk forward optimizer before? Sounds like it is a whole lot better than having nothing.

As written in another post, Tradestation works for simple indicators, however these are all virtually obsolete these days. You also need more control to test thoroughly, which can only be done in a real programming language. The sooner you get past that learning curve, the better.