Massively Parallel Processing

[TheNegotiator 13]

Does anyone know CUDA or CUDA.net? Is anyone integrating it in any way? Are there any good alternatives to it? I think there is ATI Stream maybe.

[Tams 20]

Does anyone know CUDA or CUDA.net? Is anyone integrating it in any way? Are there any good alternatives to it? I think there is ATI Stream maybe.

 

but the market is a sequential event...

[TheNegotiator 13]

A single market in real-time is a sequential event yes. What about if you were interested in processing real-time data of all the stocks in the S&P 500 for example? How much quicker might you be able to run backtesting and optimisations if you are clever about how you use the programming? I know that cuda is used already by various financial companies. Why should it be their edge though when all we have to do is pop a couple of gpu's in sli into our own pc's? Not that you could compete at the same level, just on the same playing field.

[Tams 20]

A single market in real-time is a sequential event yes. What about if you were interested in processing real-time data of all the stocks in the S&P 500 for example? How much quicker might you be able to run backtesting and optimisations if you are clever about how you use the programming? I know that cuda is used already by various financial companies. Why should it be their edge though when all we have to do is pop a couple of gpu's in sli into our own pc's? Not that you could compete at the same level, just on the same playing field.

 

backtest is still a sequential event

 

you can use genetic algorithm to simulate optimization,

that speeds up processing without added hardware.

 

 

ps. CUDA is good for vector processing (Look up Google)... ok to make 3D financial models, but not sequential events like trading.

 

One famous charting software (which should remain nameless) had to turn off multi-core CPU and multi-thread processing, because of problem with data integrity. This should give you an idea the practical application of massively parallel processing.

[TheNegotiator 13]

Hmm. Interesting. It seems to not be of huge benefit right now. If you were backtesting/optimising, could the time period you were analysing not be broken up into many constituent parts, tested and then reformed to give an output?

 

I am sure there are applications for this, perhaps at a later date as the technology evolves.

[Tams 20]

Hmm. Interesting. It seems to not be of huge benefit right now. If you were backtesting/optimising, could the time period you were analysing not be broken up into many constituent parts, tested and then reformed to give an output?

 

I am sure there are applications for this, perhaps at a later date as the technology evolves.

 

the basic fact is.... you have to buy before you can exit long -- that is an sequential event.

You cannot chop up a sequential event and process the exit long before a buy is executed.

 

 

furthermore, most of the trading logics require something like this:

if price is larger than 50 SMA ... then buy...

if number of loss trades today is smaller than "max_allowed_loss_trade"... then proceed...

 

 

that 50 SMA is a sequential event that must go back 50 bars... you cannot chop up this event for parallel processing.

 

same goes for any conditional decisions that looks back in time.

 

 

Hope these examples helps.

[Eubie 10]

Hello guys,

 

I was considering learning CUDA for making a backtester of my own making faster, but the problem with GPGPU is that these technologies are greatly hindered by the RAM -> GPU bandwidth. Hence it is benefitial for backtests where a large part of RAM is moved onto the graphics card once, from where the GPUs can perform the whole backtest. However, with today's cards having at max 3GBs of memory, you can't possibly hold more than 1.5 year of tick data for some less liquid instruments (say TF).

 

Backtesting is a sequential event, however not in the sense discussed here. The fact that SMA50 needs to look back 50 periods is valid, however once you precompute the the indicator for the whole investigated period, you can have each core investigate a different period of the market (searching for entries). The same can be applied for exits. However, it is the synchronization of entries and exits on patterns/SL/PT that cannot be done in parallel. So, in my view, GPGPU in backtesting if worth it only if you have enormously complicated entry patterns that take a lot of time to evaluate, where having 3000 cores evaluating entries at every tick is what makes the backtest fast. In any other setup, the power of GPGPU can't be used for quicker backtesting in my view.

[mbbcat 10]

some links;

 

CUDA Zone

Forums - Nvidia CUDA GPU technology.

Nvidia - Cuda Toolkit for options pricing | Page 2 | Quantnet - Community for Quants

Computational Finance

CUDA applications for Tesla

these guys have a cuda plugin Trading Software, Neural Network, AI, Technical Analysis, Market Timing, Stocks, Futures, FOREX

Build Neural Network Indicator in MT4 using Neuroshell - Page 29 | Trade2Win Forums

AMAX Engineering Corporation: Private Company Information - BusinessWeek

Wilmott Forums - sample cuda problems in finance

QuantHouse, low latency data and algo trading solutions!

[Tams 20]

Hello guys,

.... However, it is the synchronization of entries and exits on patterns/SL/PT that cannot be done in parallel. ....

 

.......... LOL ..........

 

there is a genius in every clown.

[andro 10]

the basic fact is.... you have to buy before you can exit long -- that is an sequential event.

You cannot chop up a sequential event and process the exit long before a buy is executed.

 

 

furthermore, most of the trading logics require something like this:

if price is larger than 50 SMA ... then buy...

if number of loss trades today is smaller than "max_allowed_loss_trade"... then proceed...

 

 

that 50 SMA is a sequential event that must go back 50 bars... you cannot chop up this event for parallel processing.

 

same goes for any conditional decisions that looks back in time.

 

 

Hope these examples helps.

 

Unless you are executing the same process for many tickers ;-) From simple view of the world and our actions to it - yes it is not really possible to do stuff in parallel. But when you move a bit into what is really going on - you'll find a lot of things for parallel computing

 

Anyway back to topic - CUDA is usually about heavy C programs. If your are able to do that - than try first to write backtest for your data on your CPU in C... maybe you will find the performance sufficient.... / like me, I was planning to the same. Then I found that by using optimized libraries and 4 CPU cores my speed up was :haha: 250x

[Eubie 10]

...Then I found that by using optimized libraries and 4 CPU cores my speed up was :haha: 250x

 

Hello andro, just ouf ot curiosity, what libraries do you use?

[andro 10]

Hello andro, just ouf ot curiosity, what libraries do you use?

 

Python and Numpy, for code overfilled with "if"s Cython.

 

And just to be precise I'm talking about code "vectorization" and "parallelization". But the same applies to CUDA.