Neural Network Approach for Market Prediction : Inputs

I've been working on this topic for quite some time.
A simple network designed to evaluate the possibility ended up with mixed results
Before going any further I need some help to sort out few problems with input variables

I used following data as inputs to the network.
Please points out if I'm missing anything important or if I have included something totally irrelevant.
-------------------------
1. Previous closing price
2. "" High
3. "" Low
4. Trade Volume
5. Exchange Rate(Mean)
6. Interest Rate(Lending)
7. ASI
8. MPI
9. OIL price
10. Gold price
-------------------------
Also I thought about adding Sector P/E, Price Index and Trade Volume
but historical data is nowhere to be find

PS:
After training the network with two years of past data (above inputs),
it was able to predict daily and weekly forecast with following accuracies.
-------------------
Next Day : 80% ++
5 Day : 60% +
-------------------
Green line of the attached image is the predicted out put while
Magenta line represents the actual data of NTB.
#Will post in detail figures in next weekend.

Thanks in advance for your valuable comments.

Great work, It this a software you created?
would like to hear more details about it

Little explanation about the chart,
-->axis X is the number of data records(i.e Date)
-->axis Y is the closing price
#Test data starts from 2nd Jan 2004 to 20th Dec 2005 (443 data records in total)
(that detail was there at first, but had to crop the image due to upload limit)

xhora wrote:Great work, It this a software you created?
would like to hear more details about it

Not exactly a software at this stage, but a kind of.
What I did was modeling the neural network using a modeling software.
I used a trial version of NeuroSolutions to model this.
There are lot of modeling software available in the market but very expensive
MATLAB is the most powerful one(i believe).

Also the network can be programed using a language like C/Java.
It is time consuming but totally free

Dileepa wrote:I've been working on this topic for quite some time.
A simple network designed to evaluate the possibility ended up with mixed results
Before going any further I need some help to sort out few problems with input variables

I used following data as inputs to the network.
Please points out if I'm missing anything important or if I have included something totally irrelevant.
-------------------------
1. Previous closing price
2. "" High
3. "" Low
4. Trade Volume
5. Exchange Rate(Mean)
6. Interest Rate(Lending)
7. ASI
8. MPI
9. OIL price
10. Gold price
-------------------------
Also I thought about adding Sector P/E, Price Index and Trade Volume
but historical data is nowhere to be find

PS:
After training the network with two years of past data (above inputs),
it was able to predict daily and weekly forecast with following accuracies.
-------------------
Next Day : 80% ++
5 Day : 60% +
-------------------
Green line of the attached image is the predicted out put while
Magenta line represents the actual data of NTB.
#Will post in detail figures in next weekend.

Thanks in advance for your valuable comments.

@ delipa + for your hadwork.

I thought about slightly different input.

1. What about entering Volume Weighted Average closing price. Than entering high and low prices.
2. would the input of a unit trust price be relavent. I was just thinking. Not sure about this. Ceybank Century Growth Fund , the investments are based on CSE. Think about this.

I could remember your previous post on the same subject.
Good Luck Bro.

good work.rep from me.

http://forum.srilankaequity.com/t4125-closing-price-of-a-share#28014

Re. Volume weghted Avg. details.

I simply do not understand the concept and the charts. What are you trying to do? I only can see a JPG file, No program or a software. Can any one of the expert help me to understand this!

Dileepa wrote:I've been working on this topic for quite some time.
A simple network designed to evaluate the possibility ended up with mixed results
Before going any further I need some help to sort out few problems with input variables

I used following data as inputs to the network.
Please points out if I'm missing anything important or if I have included something totally irrelevant.
-------------------------
1. Previous closing price
2. "" High
3. "" Low
4. Trade Volume
5. Exchange Rate(Mean)
6. Interest Rate(Lending)
7. ASI
8. MPI
9. OIL price
10. Gold price
-------------------------
Also I thought about adding Sector P/E, Price Index and Trade Volume
but historical data is nowhere to be find

PS:
After training the network with two years of past data (above inputs),
it was able to predict daily and weekly forecast with following accuracies.
-------------------
Next Day : 80% ++
5 Day : 60% +
-------------------
Green line of the attached image is the predicted out put while
Magenta line represents the actual data of NTB.
#Will post in detail figures in next weekend.

Thanks in advance for your valuable comments.

Dileepa seems you are making significant progress. Great work.

How about adding announcements ( dividend, stock splits etc) which could affect price.

If you do not have latest data ( I see oyu are using 2004-2005 am I correct) try www.slstock.com
user and password is guest

milanka wrote:http://forum.srilankaequity.com/t4125-closing-price-of-a-share#28014

Re. Volume weghted Avg. details.

Thanks Milanka,
Does that mean closing price published by CSE is in fact VWAP or is it published separately
Could not find historical VWAP.
Regarding Unit Trust Price(didn't really understand that ),
I checked at ceybank but again could not find historical data.

Quibit wrote:I simply do not understand the concept and the charts. What are you trying to do? I only can see a JPG file, No program or a software. Can any one of the expert help me to understand this!

Neural Networks are basics of Artificial Intelligence. The programmers try to mimic actual world relations as far as possible in the model developed inside the computer.

More relations and interactions you can think and include in the system will bring it closer to functioning of our brain and produce better results. I think you can write the program to do self adjustments or fine tune when the system is fed with more and more actual data comparing its own predictions.

In Dileepa's program Green line is the predicted out put while
Magenta line represents the actual data of NTB.
With the time, he can fine tune the program to make the two lines closer and closer.

slstock wrote:

Dileepa seems you are making significant progress. Great work.

How about adding announcements ( dividend, stock splits etc) which could affect price.

If you do not have latest data ( I see oyu are using 2004-2005 am I correct) try www.slstock.com
user and password is guest

Thanks, actually I already knew about slstock from a previous post of yours
I downloaded the full set of data, but arranging them seems to be taking some time.

Also thought about adding announcements but do not know how or on what base.
problem is with the arrangement of data records,
1. each record has to be complete with data
2. if I add dividend/split info in dayx, what I am going to feed on other days

Dileepa wrote:

milanka wrote:http://forum.srilankaequity.com/t4125-closing-price-of-a-share#28014

Re. Volume weghted Avg. details.

Thanks Milanka,
Does that mean closing price published by CSE is in fact VWAP or is it published separately
Could not find historical VWAP.
Regarding Unit Trust Price(didn't really understand that ),
I checked at ceybank but again could not find historical data.

On the CSE web site the page http://www.cse.lk/listedcompanies/overview.htm give s you the VWAP once the market is closed Else its the last traded price . Its given ther with an ** .

The a unit trust they invest money in different sectors. The specific one I mentioned invests money in CSE shares. They sell units to investors but finally money is invsted in shares. it you visit www.ceybank.com you may have and idea. I was wondering that the price of a unit could related to a ASPI or MPI.

Ther are two pricws for a unit. the buying and selling price. Buying is < selling. Thast the Ceybank proft or commission thay say. For your purpose either is okay.

To get the data , you have to contact them. In every mian BOc ther is a person who do the unit trust sales. First try them. Tell that you want to invest in unit trusts and you wanted past data.

Hope this helps.

Dileepa - good work mate. What about adding inflation rate as a variable. This could be as published by CB time to time?

Or interest rate (lending) itself is enough? as interest rate and the inflation are related.

Dileepa wrote:

slstock wrote:

Dileepa seems you are making significant progress. Great work.

How about adding announcements ( dividend, stock splits etc) which could affect price.

If you do not have latest data ( I see oyu are using 2004-2005 am I correct) try www.slstock.com
user and password is guest

Thanks, actually I already knew about slstock from a previous post of yours
I downloaded the full set of data, but arranging them seems to be taking some time.

Also thought about adding announcements but do not know how or on what base.
problem is with the arrangement of data records,
1. each record has to be complete with data
2. if I add dividend/split info in dayx, what I am going to feed on other days

Are you training on overall CSE data a set or just on a single company ?

If one company just an example:
To test you could add
a) +5% increase on a share split annoucement day
b) -4% drop on 2 days afterwards
b) + 5% on the day the shares have been split
c) - 4% on day follow the split

Again these numbers can vary but giving you an idea.

For Dividend you could have 2% share price increase as an example on the annoucement date , then a 1% drop the next few days etc..You get the idea


You could test and see.

Country's GDP growth rate can also be considered, isn't it?

Good work keep it up.. there may be buyers if you can perfect this
If you want mathlab, pm me..

Hey Dileepa, Try to use some adaptive filtering as well. it would reduce the error between actual and forecasted.

smallville wrote:Good work keep it up.. there may be buyers if you can perfect this
If you want mathlab, pm me..

What's mathlab? Or is it Matlab? I did my final year project in Matlab!

StocksWatch wrote:Country's GDP growth rate can also be considered, isn't it?

About your previous post,
I could not find inflation rates, so I used interest rate.
As you pointed out I think it is pretty much the same.
GDP, Yes there might be a connection.
By the way GDP is calculated monthly rite

slstock wrote:
Are you training on overall CSE data a set or just on a single company ?

If one company just an example:
To test you could add
a) +5% increase on a share split annoucement day
b) -4% drop on 2 days afterwards
b) + 5% on the day the shares have been split
c) - 4% on day follow the split

Again these numbers can vary but giving you an idea.

For Dividend you could have 2% share price increase as an example on the annoucement date , then a 1% drop the next few days etc..You get the idea

You could test and see.

For the time being thought of concentrating on BFI sector (to be specifically NTB).
Attached file is a screen shot of sample test data.
(sorry about the colors, had to make it mono due to upload limit)
And thanks for the suggestion, will post some more results again

Good work Dileepa. + Rep from me. Keep up ur good work.

keep the H, L, O, C, volume & ASI + maybe MPI

Everything else is useless.

good work deelipa ,eventhough i coudnt understand much of it it seems your are trying to do a marvelous work it fully succes...wish you all very best +rep from me to your hard work

Chinwi wrote:

Quibit wrote:I simply do not understand the concept and the charts. What are you trying to do? I only can see a JPG file, No program or a software. Can any one of the expert help me to understand this!

Neural Networks are basics of Artificial Intelligence. The programmers try to mimic actual world relations as far as possible in the model developed inside the computer.

More relations and interactions you can think and include in the system will bring it closer to functioning of our brain and produce better results. I think you can write the program to do self adjustments or fine tune when the system is fed with more and more actual data comparing its own predictions.

In Dileepa's program Green line is the predicted out put while
Magenta line represents the actual data of NTB.
With the time, he can fine tune the program to make the two lines closer and closer.

thanks for your nice explanation ..now i got understood +rep from me