Market Models

[stephencrowley]

Quote from ATSTRADER:
I never seen successful AI model, but I know some people who are testing such system now and claiming it should work... [/B]

Eh, what are you defining as AI? Because statistics and AI are the same thing, but "AI" is just statistics that is much more complicated than what most people perceive statistics to be.

 

[OddTrader]

Quote from ATSTRADER:

Just three questions:
Was it real trading or paper?
How many trades the winner made?
To consider the trading approach successful minimum 1000 real trades should be done...
Did NFA verified your claims?

You probably don't know that are not my claims.

 

[DrChaos]

the preferred term now is "machine learning".

I would translate this to be reasonably sophisticated quantitative modeling which is oriented towards Bayesian type of probabilistic prediction, as opposed to classical hypothesis testing.

Compared with 'regular' statistics, the problems, data and underlying models often involve high degrees of freedom and nonlinearity, and thus results depend profoundly on the insight of the data analyst, model structure, explicit and implicit assumption. The "machine learning" part is an attempt to fit the parameters without overfitting noise; the "human learning" part of the problem is more important to success.

Also success functions are also "soft" in the sense of somewhat better some of the time is OK.

 

[rufus_4000]

Quote from stephencrowley:

The smallest indivudal you knew was trading 2.5 mill? How much did he start out with? I would think you only need the 25K PDT minimum unless i'm missing something.

Also, these blow ups sound interesting. Is hedging not so common for intraday bots? Were they just leveraged so far that a small factor threw them off? I'm interested to know how risk is approached in this type of "nonstandard" trading, not textbook stuff anyway.

Sorry, I was in my institutional mindset, and after talking to a friend who wants to start to trade on her own. When I asked her how much start up capital she would need, she said 10M, and I guess in that mindset we both thought the figure sounds perfectly reasonable.

In terms of leverage, for a professional firm, I know that they have only about $1.5M in capital, and frequently have offsetting 80M or so positions, that doesn't sounds unreasonable at all. The risk associated with high frequency trading can be sufficiently big, I have seen $1.4M wiped out in span of about 40-50 mins, but that firm was putting a system that in my opinion, not sufficiently well backtested into production, only to have the system blow up on the 2nd day. Daily losses of 300-500k is very common, they are not even worth mentioning. It is very hard to hedge again small spikes intraday.

As I have mentioned before else where on ET, back in '99 I was doing cap structure arb, playing the COMS/PALM spin off in a hedge fund, and had on positions of 45M or so, I lost around $2M before I knew what hit me. But that wasn't automated, and that was the start of me starting to focus more and more on automated trading.

 

[nononsense]

Quote from DrChaos:

the preferred term now is "machine learning".

... the "human learning" part of the problem is more important to success.

 

[ATSTRADER]

Quote from DrChaos:

the preferred term now is "machine learning".

I would translate this to be reasonably sophisticated quantitative modeling which is oriented towards Bayesian type of probabilistic prediction, as opposed to classical hypothesis testing.

Compared with 'regular' statistics, the problems, data and underlying models often involve high degrees of freedom and nonlinearity, and thus results depend profoundly on the insight of the data analyst, model structure, explicit and implicit assumption. The "machine learning" part is an attempt to fit the parameters without overfitting noise; the "human learning" part of the problem is more important to success.

Also success functions are also "soft" in the sense of somewhat better some of the time is OK.

That all make sense to me, some of my friends are trying to work on this particular issue. According to them, the main idea here is to teach machine to think like a human, when analyzing the statistical data. Generally, this project is sciencefictional, but I like to belive that someone will eventually succeed and AI statistical model would be build.
For now, let me call this Model "the mysterious one"