MAchine learning based market prediction

[piyushranjan]

Hello all
I have read about Neural network(or other such stochastic methods) based market indices prediction. It basically involves predicting market trends based on historical data of other indices, fundamentals of the market, macroeconomics factors etc. I would like to know if anybody has experience in such field. Also if such approach is employed by any broker firm or bank on NSE/BSE ?

 

[vipin.kumar05]

Well this is the topic of my final year thesis in my engineering degree:

"Forecasting Sensex using neural network"

I did that 3 years ago.

Result: It doesn't work.

 

[superprogrammer]

Would you be kind enough to tell why it doesnt works?

 

[andygp]

It is interesting to note that a person having same name as one of the reasonable researchers in this field does not believe the capability of the area :lol:
http://www-users.cs.umn.edu/~kumar/

 

[bunny]

If you want to learn prediction, forget technical indicators, trading setups, costly technical analysis sofware, etc.

All you have to do is identify the hotspots of supply and demand, and co-relate it with volume. Volume is an essential technical indicator, and the importance of it is often understated, sometimes even ignored totally. It is possible to predict price movements with historical price and volume data. But it is easier to do it visually than write an algorithm(may be because I am not a engineer?).

 

[vipin.kumar05]

Would you be kind enough to tell why it doesnt works?

All neural networks depend on learning. Learning of the data pattern. The problem in my case was the simple 5-4-1 neural network either overlearnt the things or underlearnt. Finding the optimal layers of network, and then optimizing it for correct error tolerance and then predicting it for optimal period is like finding the Holy grailed trading system.

But the thing was, i was a newbie back then, for me it was my first foray in to the financial market. I was happy to see the OHLC of the SENSEX on my MATLAB .

 

[piyushranjan]

Vipin
Or may be the model itself is not wrong but the data is not sufficient. I don't know. I have never deployed NN for this purpose. I think we mentally do market prediction almost everyday. Things like "oh Dow was up overnight and Shanghai is also up this morning definitely it will go up"(please forgive me if I am wrong and if world has started following us(?)). Point is that there are definite patterns which people can learn and give certain prediction

@bunny
Thanks for the insight. Very intriguing indeed. Well the idea to deploy a NN to do market prediction is to eventually allow machines to trade. World over people are doing this. Companies like tower research are doing it quiet successfully. Even the larger banks. My interests are purely academic/exploratory in nature though

 

[piyushranjan]

@superprogrammer
If he knew that, he would have cracked it!

 

[vipin.kumar05]

@superprogrammer
If he knew that, he would have cracked it!

Heheh

You know that was the standard answer in all of my job interviews:

Interviewers: So, finally, what happened to your sensex forecasting algorithm.
Me: Sir, it didn't work otherwise i would not sitting here.

 

[piyushranjan]

Thats true ..isn't it?
After all you are trying to model a "black box" function. If you knew the correct inputs you would have cracked it
Any more efforts in that area ?