Neuromongers, don't pass by :) need advice - page 11
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Hello all.
As far as I understand the author deals with recurrent networks, but classics warned us about problems with application of such networks in financial markets long time ago.
I quote: The standard connectionist approach to incorporate time history uses recurrent networks. How-
ever, recurrent networks have not seen much success on financial problems where the small amount
of signal contained in a typical training set seems to be insuficient to determine the structure and
parameters of this unconstrained architecture. Some constraints need to be imposed. We propose
the model class of hidden Markov experts (Shi and Weigend 1997). This class strikes a balance
between time-ignoring regression models and fully recurrent architectures. Hidden Markov experts do
take time into account explicitly, yet avoid the dificulties of fully recurrent architectures by imposing
stringent constraints on the way time enters.
Source: Density Forecasting Using Hidden Markov Expert (Shanming Shi & Andreas S. Weigend)
Feel free to comment.
It is my understanding that the author is dealing with a recurrent network
Because the next step was normalization by volatility. And maybe move to a pseudo price series, but the volatility is not jumping so much...
You just take the first differences. It begs the question why not the second or third differences.
By taking the first differences you start to depend a lot on the TF. This begs the question, why go to the returns?
You just take the first differences. It begs the question why not the second or third differences.
By taking the first differences you start to depend a lot on the TF. That begs the question, why go to returns?
What are your suggestions other than returns?
Nah, I think it's nonsense that this needs proof. Or at least examples of divergence in quotes confirming that the network learns admirably on some and badly on others.
And "looking into the future" in quotes implies that history is rewritten on every tick - well that's bullshit.
Bullshit is more your theories about the Forex market.
Nobody said that history is rewritten on every tick in real time. Just to reduce "fuzziness" and gaps in historical quotes, they are smoothed by algorithms that look beyond the zero bar (in MT terms).
As proof, you can run the forward using historical data from other sources. Illusions will disappear.
Is it so difficult to do?)
But if you like building air castles... I won't get in the way)
You'd better tell me where to get a better story.
A matter of taste... Reuters, Bloomberg, ... there's a lot of paid ones.
Of the free ones, you could try Dukas, and a couple of European brokers.
What are your suggestions, other than returns?
1) Bullshit is more your theories about the Forex market.
2) No one said that history is rewritten on every tick in real time. Just to reduce the "fluffiness" of historical quotes, they were smoothed by algorithms looking beyond the zero bar (in MT terms).
3) As proof, you can run the forward on historical data from other sources. The illusions will disappear.
4) Isn't it that hard to do?)
5) But if you like building air castles... I will not interfere)
1) Bullshit is not supported by experimental results. My theories are confirmed.
2) If the history is not overwritten, then there is no problem. Where is the information about the filtering algorithms used by brokers/dts? Can MQs confirm this information? Can you provide tests to prove these claims?
3) It's not clear. What kind of forward? Train on data from one source and forward on another? Why?
4) I presume you have already done this? Share the results then.
5) Which locks are we talking about?
In general, the system in question is not a pipsator, and it is not sensitive to fuzzy quotes. It will equally well operate on data from one source, and trained on the other (do not take into account the sources with outright "broken" history).
1) Delusion is not supported by experimental results. My theories are confirmed.
Blah, blah, blah
2) If history is not being rewritten, then there is no problem.
The problems will appear when you go from tester to real...
3) It's not clear. What's the forward? Train on data from one source and forward on another? What for?
Train and forward on data from a decent vendor.
The system is not a pipsator, so it is not sensitive to fuzzy quotes. Equally well it will work with data from one source, and trained on the other (not taking into account the sources with outright "broken" history).
In general, the system in question is not a pipsator, and it is not sensitive to fuzzy quotes. The system will work equally well on data from one source, and trained on other sources (not taking into account sources with outright "broken" history).
From the 15M timeframe the history from different sources differs very little, and starting from 1H the OHLC prices almost always coincide. Of course, if you do not use history found somewhere in the dumpster. I do not understand what some people are talking about here either....