"Forecast Modell Financial Markets"

Indexer

Hi,
If got my one time series which come from my forecast modell. They should predict the stock market trend, currencies and commodities. I do that on excel but I think rapid miner would be a better way. I look for somebody who could develop a forecast modell on rapid miner. I would deliver my time series (my modell) and as well as the time series for the financial instruments which I want to predict (like S&P 500 aso. ). The aim is that the modell tells me if I must be long or short the financial instrument. So the only thing I need to know is buy or sell. I hope that there is anybody out there who could develop me that on rapid-I and give me also the instruction (written or also online).

For sure I'm willing to pay for that.

Kind regards

Marin

Dear Indexer,
    as you can see community is developing models (along with instructions) all the time. I would suggest learning how to catch fish rather than getting the fish with the cooking instructions. Especially as in stock, Forex and commodities markets, yesterday's recipes are like yesterday's fish.
    It does take some time, but in a long run that is the only way to get through financial markets. I can suggest the following (it all depends on what you seek to accomplish):
-look video tutorials here: http://rapid-i.com/content/view/189/198/lang,en/ Thomas Ott gave examples of financial time-series trend prediction
-read the manual Manual RapidMiner 5.0 (English) http://rapid-i.com/content/view/26/84/lang,en/
-start reading this forum regarding operators which you expect to use for your solution
-try Online tutorial in Rapidminer
-try the templates in Rapidminer
-check out different RM products, e.g. RapidMiner Sentilyzer might be a good extension to your model http://rapid-i.com/content/view/184/194/lang,en/

And most certainly make sure not to miss meeting us all at RCOMM 2010 in September ;-)
Thomas Ott has an invited talk on the similar financial market problem. More about the conference: http://rapid-i.com/rcomm/index.php?option=com_frontpage&Itemid=28
I suggest you don't miss the opportunity to talk with all these people on the subject of your interest.

Last certainly not the least, check out SERVICES http://rapid-i.com/content/view/60/200/lang,en/ for professional help with guaranteed support and response times, webinars and training which you might find more appealing to your needs.

Yours truly,

Marin

Indexer

Dear Marin
I think you understood me wrong. Sorry that my english is not the best. If developed my own modell, that means that I have time series which if think can predict the market. For example my time series goes up than I long the stock market, if the time series crosses the 10 day average of my time series than I sell the stock market. Now I have about 500 times series for nearly every instrument in the market. If proofed  in on excel that this modell predicts the market, but now I want to do that with rapid-miner.

IngoRM

Dear Indexer, dear Marin,

I think Marin got the right idea: you have your own model, built in Excel, and now you want support to use RapidMiner for the calculations, right? Well, in my opinion Marin's answer perfectly fits this generic request as far as I can see. Without any concrete details probably no one here is able to help you better.

And by the way, Marin: Two thumbs up for this great overview about all introduction material and services we have  :-*

I look for somebody who could develop a forecast modell on rapid miner.

So what's wrong with the guys of Rapid-I? Of course I am bit biased but I think they would do a great job on this 

Cheers,
Ingo

gero_schwenk

Dear Indexer,
you're trying to solve though puzzle. I'm tempted to state that even if you are getting better than random guess, you won't generate much gain after transaction costs. (At least in the low frequency domain and with sole prediction of stock prices.)

Furthermore, it takes much time to get familiar with the peculiarites of financial data. Stochastic processes are martingales and don't carry much autocorrelation information after necessary differencing. Furthermore, performance of predictive models is fluctuating over time - given that you have found predictors which have some out-of-sample- resp. backtesting-performance.

Bluntly spoken: Working part-time on the topic, it took me about a half year to establish a modeling framework including the performance metrics which allowed me to judge the trustworthiness of a prediction. That's a whole lot work to pay for - and you would get only moderate and fluctuating performance in return.

Here you can find a link to an example output which predicts the N225 based on predictors found by hierarchical clustering and ranking of lagged correlations. This one used a Epsilon-SVM / Epsilon-SVR, but MARS (and probably many other machine learning schemes) provides comparable results. Inclusion of more predictors does not improve performance beyond overfitting.

http://www.aq-me.de/aq.pred.N225.300310.pred.pdf

Cheers,
Gero

PS: I used the R-framework for building this.

wessel

What is this pdf?

I don't understand it!

gero_schwenk

its an output, which is mostly about model performance in roling data windows.

the first page is organized as follows:

1) prediction 2) history of predictions
3) rolling classification error / out of sample 4) rolling classification error / crossvalidation
5) rolling classification error / out of sample based on SVR 4) rolling classification error / crossvalidation based on SVR

the second page shows target and predictor variables

wessel

A good way to visualize error is to plot "predicted" against "real".
If all data points in your sliding window validation (aka prequential validation) are correct, all points lie on the diagonal.
Over estimates of the real value lie above the diagonal, under estimates of the real value lie under the diagonal.
Total random prediction will just be a cloud of points.

An example of such plot:

gero_schwenk

Ups, sorry for the confusion. I just wanted to give an example for the hardships of prediction of the sign of stock returns...

Regarding the error time series in my output: these report the error proportion (out-of-sample and cross-validation) for a window that rolls / slides over the original time series. So every point in the plots is a summary error measure for a specific dataset - namely the data window which is rolled over the original time series.

Cheers and thanks for your interest!
gero