Matlab

[nitro]

Quote from prophet:
I’ve done some simple backpropagation NNs, both with the Matlab toolkit and with my own code,

AFAIK, this will not give you much, at least it results in no better models than taking a system, making it adaptive in a more traditional way, and doing some kind of walkforward optimiation at first.

My current field of interest is in using Time Recurrent NNs, sometimes generalized as Back Progragation Through Time NNs. I see a great deal of promise in this kind of network, but my overall opinion of networks for _the_system_themselves_ is very low.

I have a great idea where I think a NN would shine though...

...along with some exotic home-brew networks based on analyzing the singular value decomposition (SVD) products.

I am surprised this would work. SVD I guess is OK for data preparation if you map the to the significant Eigenvector/Eigenvalues to act as a filter, but IMO, the data you are throwing away is just as important.

Any kind of adaptive model for forecasting of prices is difficult, even when the model is fit across thousands of stocks and thousands of days.

Ja

I never base my network models on price alone. I usually rely on system performance statistics or inter-stock money flows. I keep the networks very simple, constraining or structuring them as to specify a hypothesis. The network then figures out the details [/B]

Only the toy ones rely on price alone. AFAIK, no one has ever written a serious package that is turnkey with trading as it's target from the beginning. My guess is that the ones that have gone all the way are proprietary implementations.

nitro

 

[thetraderprofit]

You guys are making my head spin.

 

[prophet]

Quote from nitro:

AFAIK, this will not give you much, at least it results in no better models than taking a system, making it adaptive in a more traditional way, and doing some kind of walkforward optimiation at first.

I didn’t work on the NNs long enough to get any success. I agree that traditional models are often superior, certainly for the processing speed advantage.

My current field of interest is in using Time Recurrent NNs, sometimes generalized as Back Progragation Through Time NNs. I see a great deal of promise in this kind of network, but my overall opinion of networks for _the_system_themselves_ is very low.

Should be very successful.

I have a great idea where I think a NN would shine though...

Option pricing models? Or is that already overdone?

I am surprised this would work. SVD I guess is OK for data preparation if you map the to the significant Eigenvector/Eigenvalues to act as a filter, but IMO, the data you are throwing away is just as important.

The SVD is very nice because once you have the decomposition, you can do many tests with it very rapidly, such as creating various filtered pseudo-inverses.

Only the toy ones rely on price alone. AFAIK, no one has ever written a serious package that is turnkey with trading as it's target from the beginning. My guess is that the ones that have gone all the way are proprietary implementations.

What about targetting short term price forecasts? Good forecasts make for good trading. My systems target outright price forecasts. Trades are secondary to that.

 

[hmerkinger]

Dear flakac,

today I have posted a message on EliteTrader (Software forum) about the release of Trading with Matlab 1p2.

TWM does exactly what you need, it provides all IB TWS API functions in Matlab.

You can find a demonstration of its capabilities on

www.gloriosia.com


Best wishes,
H.M. Erkinger