Building an ATS - Logbook

[CT10Gov]

Quote from 2rosy:

thats what I was getting at. I meant that for personal use infrastructure (especially in python) is not difficult and can be done in a few days. Nothing needs to be fast considering you're going over the internet and most likely have a slow feed.

I totally agree with you are rationale. (You are a python guy? I really tried to get into it; I like it a lot; but I found the lack of a good editor somewhat limiting; And the lack of good thread support. I'm spoiled by visual studio I guess. But having an interpreter rocks).

Anyway, I think the instinct to build 'infrastructure' rather than strategy is a strong one: with the former, at least you see something being done everyday; With the latter, just a lot of negative results for a long while.

 

[tiagor]

It's nice to see this thread has sparked some discussion. My take on the subject of time investment on the SHELL (oms) vs GHOST (algo) is that both are important.
I understand there is a larger group of people that will have a better chance of succeeding in building a workable oms than a profitable algo. However, with some algos (HFT comes to mind), a deficient oms will trace the line between success and failure, independently of the algo potential.

Progress report: The simulation engine of the SHELL is nearly complete. I guess a few more days will allow for completion and debugging and I'll be able to start doing heavy work on bringing the first GHOST to life. I've been working on the code for a week now, with only a couple of hours available each day. From my personal experience I'm inclined to say it is definitely possible to build a working ATS prototype in some days if you know what you're doing and don't aim for something too complex.

Benchmarking the GHOSTs: As I said in a previous post, I'll be making backward tests on the GHOSTs with both in-bound and out-bound data to avoid overfitting of variables. I'll use the sharp ratio as a measure to compare them. I was wondering what is considered a good measure of correlation between the in-bound and out-bound tests and an acceptable sharp to separate the wheat from the chaff. Other suggestions for strategy performance measurement are also welcome.

As a final note, its been funny to see that many of you have adopted the concepts of SHELL and GHOST on the discussion. Guess they intuitively fit their purpose.

 

[CT10Gov]

Quote from tiagor:

However, with some algos (HFT comes to mind), a deficient oms will trace the line between success and failure, independently of the algo potential.

You are not going to be building HFT.

By the way, GHOST and SHELL isn't intuitive, but it's actually kind of catchy given the reference to the classic anime. Also, I like writing in all caps. Makes it feel like we are talking about much more important things than little bit of code that buys and sells random crap.

 

[sle]

Quote from Kevin Schmit:
The shell (OMS) is the difficult part, especially if your ATS generates complex multi-leg, algorithmic, and/or contingent orders that, once filled, need to be dynamically hedged. Dealing with partial fills, hanging legs, etc can be convoluted and error prone. I have never seen a non-pro OMS that gets these things right.

Why would a non-HFT trader want to build an OMS on his own? There are plenty of products out there that do that do it well enough and your job, as a trader, is to find where alpha is, not to optimize order handling for the 15th time around. I am pretty sure there are even some open or semi-open source OMS systems out there, so there is no need to reinvent the wheel.

I am pretty sure 99% of people here are not latency sensitive and can do their trading via Excel or Matlab API and there are plenty of trading systems with complete and intelligent OMS APIs for these two products. Just saying...

 

[tiagor]

Quote from sle:

Why would a non-HFT trader want to build an OMS on his own?

Speaking only for myself, it all comes down to control.
If I'm allowing an automated system to do things for me, I better know it will do exactly what I want it to do and when I want it to.
I rather have fewer fancy useless functionalities and tweak what I feel is needed than have an OMS that may be great but I don't know what exactly its doing in the background.

Just my 2 cents. I can accept it won't be an issue for most though.

 

[tiagor]

Quote from tiagor:

Benchmarking the GHOSTs: As I said in a previous post, I'll be making backward tests on the GHOSTs with both in-bound and out-bound data to avoid overfitting of variables. I'll use the sharp ratio as a measure to compare them. I was wondering what is considered a good measure of correlation between the in-bound and out-bound tests and an acceptable sharp to separate the wheat from the chaff. Other suggestions for strategy performance measurement are also welcome.
[/B]

I was afraid this would easily go unnoticed on the previous post since the discussion seems to be orbiting around developing an oms from scratch or not. Since I'm practically beyond that phase, help is appreciated on the benchmarking issue

Summarizing the issues I'd like some inputs:

1. Is a good correlation between backward testing in-bound and out-bound data enough of a good measure of consistency before forward testing on demo account?

2. How good is a good correlation?

3. What is a considered a good Sharpe ratio for an ATS?

4. Any further suggestions?

 

[themickey]

I built my own ATS from the ground up this year.
I have coding skills learned from Amibroker but I use Excel for the ATS.
Two computers, 4 monitors to do the job.
At a guess, I've spent 1500 hours designing, coding and testing this one Algo.
It only trades one futures contract, the SPI, not designed for anything else and is now complete and running.
The Algo includes features such as recognizing holidays, emergency circuit breakers, large gaps, losing positions, fast moving market, market chop, going long, going short, quantity of trades, fully auto, manual intervention (placing trades or stops), manual emergency liquidate features, plus a couple of other manual components which means the system can go to semi auto rather than full auto. This as I look at another system and key in levels where I calculate the market will stall - trading chop being the deadliest sin.
The final product is not how I envisaged when started. My original concept was flawed so I had to come up with a new idea and this has morphed from a caterpillar to a butterfly to use an analogy. There are 4 scenarios to allow a long or short trade. There are 8 conditions to be met prior to allowing a trade. There are 14 different auto parameters, any one of which will cause the system to exit a position.
I think the scariest part of the past year was when i discovered my original concept was fatally flawed and after spending considerable time on that, was extremely concerned I had walked down a blind alley. I felt then I was maybe destined to spend the rest of my life finding the allusive pot of gold at the end of a rainbow and going nowhere. Luckily I worked out the puzzle, but I'll say this, I've been trading for over 20 years and designing for most of that. The difference being that was stocks, now I'm doing futures. I work a full time day job and so my typical day when coding is out of bed 3-4AM, go to work 8AM, come home 5PM, no dinner (hot lunch at work) and work till 8PM then lights out. Work all weekend, no socialising hardly the past 11 months. Job's finished, now time to relax and find a tasty lady.

 

[CT10Gov]

Quote from tiagor:

1. Is a good correlation between backward testing in-bound and out-bound data enough of a good measure of consistency before forward testing on demo account?

You mean in-sample and out-of-sample?

Quote from tiagor:


2. How good is a good correlation?

I don't think anyone has a definitive answer on that.

Quote from tiagor:


3. What is a considered a good Sharpe ratio for an ATS?

Depends on what kind of strategy it's running.

 

[tiagor]

@themickey: thank you for you contribution. I'm glad it turned out ok for you and hope the system meets your expectations on live trading. I must say I'm actualy amazed you implemented all those rules only with excel. Depending on excel to do all that would definitely freak me out (and I must say like excel a lot).
By the way.. do you have and estimation of the Sharpe ratio for your system?

 

[tiagor]

@CT10Gov: Yes, I mean in-sample, out-of-sample data. Regarding the correlation and Sharpe ratio I didn't expect a definitive number to begin with. But certainly there is an interval of confidence that can/has be used.