Designing a System from scratch

U

uasish

Guest
#83
Ajay,
It has already taken more than 3Hrs 45 mins for only 1 col & 473 bars,still 4lack 24k epochs crossed.Not even 50 % finished.

Asish
 

oxusmorouz

Well-Known Member
#84
Ajay,
It has already taken more than 3Hrs 45 mins for only 1 col & 473 bars,still 4lack 24k epochs crossed.Not even 50 % finished.

Asish
One thing I have noticed is the quality of input fed to form the predictive model. To bring the error ratio down to 1% with flawed inputs is quite a task indeed (I've tried it). There are two ways to deal with this:
1) Increase the acceptable error ratio to some 4%.
2) Increase the quality of inputs.

Anyway, can't comment much till CV responds. I might be moving with the wrong premise.
 
C

CreditViolet

Guest
#85
One thing I have noticed is the quality of input fed to form the predictive model. To bring the error ratio down to 1% with flawed inputs is quite a task indeed (I've tried it). There are two ways to deal with this:
1) Increase the acceptable error ratio to some 4%.
2) Increase the quality of inputs.

Anyway, can't comment much till CV responds. I might be moving with the wrong premise.
Hi Oxy
Sorry about the late reply, the volatility clustering effects kept me busy.About your XL sheet, looks interesting, however two points.
EOD data - You will have very hard time walking it forward, you should work on other TFs

Work on derivatives of Price as well along with working on Raw data itself.

Also, I didnt get your PM as Amit was saying, in couple of days ( hope so if Wifey doesnt throw weekend tantrums ) will PM you with few more inputs.

Laters
 
U

uasish

Guest
#87
I believe Ajay must have taken the Price Derivative in his Module,well Jesse is the best person to judge that.
 

oxusmorouz

Well-Known Member
#88
Hi Oxy
Sorry about the late reply, the volatility clustering effects kept me busy.About your XL sheet, looks interesting, however two points.
EOD data - You will have very hard time walking it forward, you should work on other TFs

Work on derivatives of Price as well along with working on Raw data itself.
Hi CV,
Thanks for your reply.

I've used EoD data due to availability of fundamental information. Learning a bit about predictive models, I've read that more related information available as inputs to train NN, better the reliability of the model. I've been collecting anything which affects close, which can be quantified, and is available or is decipherable. Thus, if the time frame was to be changed, I felt it would lead to fewer inputs to train NN.

Regarding using raw data as inputs, I've read that raw data (especially high and low) contain a significant amount of noise in them and can show extreme blips which, if used, could be mislead the predictive model, and it is better to reduce the clutter using a moving average etc and use them as inputs instead...

Shall be waiting for your PM.

Sincerely yours.
 
Last edited:
U

uasish

Guest
#89
http://www.traderji.com/107159-post86.html

Ajay,
Your preference to Lagrangian method in optimization reflects your Focus to the Core area of System's Performance & Fund Allocation,hence this post & subsequent queries & replies to this post should be in this thread.
However pardon my lack of knowledge on the relevant subject,still as an uninitiated
may i put forward that if the Module has a Benchmark to compare then Viterbi Algorithm which uses Dynamic Programing will find the best path faster .

Asish
 
#90
can anybody give me the metastock formula for achieving the following:-

formula for 1. If the c and ref(c,-1) cross 10 DMA the color of moving
average to become Blue.


2. If 10 DMA cross above c and ref(c,-1) the color of moving
average to become red.
 

Similar threads