Neuro-forecasting of financial series (based on one article) - page 6
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
Actually, it's overtraining. I am surprised that you do not know this. The conventional wisdom is that a network is overtrained when it starts working the way it did yesterday. That is, it does not highlight key points in the input, but starts producing the same signal as yesterday.....
Nonsense, or what's another word for that kind of nonsense?
If the network was taught that 2*2=4, what should she answer if she is asked the question, "2*2=?
2*2=4 was yesterday and will be tomorrow. It would at least be strange if the network started giving different answers to the same questions.
In general, as I see it, many of those on this forum who have practised with networks think that neural networks are like a calculator - you enter data and the network produces a result. That's right - networks are calculators.
But a thorough knowledge of networks and their architectures does not and cannot provide insight into their properties. It is a "narrow" understanding of how ANNs work.
Thus, a brilliant neurosurgeon cannot know what the human brain is capable of. Only psychologists and psychotherapists (not to be confused with psychiatrists) together with philosophers and perhaps sociologists can know about it. These are not questions of the "physiology" of networks, but of Intellect - and this is a completely different level of understanding of ANN (and the human brain).
Only with such an approach is it possible to achieve something from ANN. Stupidly feeding quotes to networks has never given anything good before, and will not give anything in the future. But the author, in my "diagonal" opinion (I don't have time for a second detailed opinion) is looking at prediction using ANN from the right angle, from the viewpoint of Intellect - using indicators, some "derivatives" from quotes.
But, purely intuitively, I doubt it's possible to forecast at 98%. 80% probably yes, but not 98%. Maybe the author has dragged the research results a bit - I am quite willing to admit that. But to verify these results, you need to do the same experiments as the author - only then you can say that it's a "sure thing".
Integer:
I am also very surprised how you can have an opinion without having grounds for it.
I already gave my opinion about the term 'overtrained' somewhere here. The term does not reflect the essence of the phenomenon at all. What English word is so translated? Rather, the term "rote" (from "crammed") or "learned" is more appropriate. The phenomenon is similar to that of a crammer or a crammer with a blank head, who understands nothing but has rote the paragraph word for word. The phenomenon occurs when teaching a volumetric network with a small number of samples. The network reacts correctly to the training samples, but it is of no use, because it can accommodate the training of a much larger number of samples, i.e. it is just an empty head. The result it gives is whatever it is, not yesterday's. So, what you have such yesterday is not clear to me, some kind of miracles, some kind of magic overtraining.
It's like programming formulas, they say about memory leaks on every occasion. So it's when talking about networks - over-training, over-education, but few understand what it is.
You name it. It's just that in some textbooks I've seen the phenomenon that the network gives out signals "Just like yesterday" it doesn't mean that it gives out the same signal all the time. No, but the essence and meaning of the signals doesn't change. They do not work in the future. Generally speaking, finding the training parameters of a network is the most difficult task, which can be solved by long statistical research.
Suppose a network has given out 100 options for setting its parameters. And only a few of them are correct. The rest will leak out in the future. Let's take a network on the mashes for example. It is very good for re-training in any area. I wanted to try this theory myself, but it's not my fate. So, maybe someone will need it.
We have two sections of training
1. we train the network on the 2nd part. Receive a profit (the net on any part of the training will receive a profit).
2. Watch how the net was working on section 1.
Let's train the net on patch 1 so as to get the same result, as we had when training on patch 2.
That is the target function for training will not be maximum balance, but quite a different one. So, using statistics we may calculate what is the best.
I just raced nets in NS, and there with a choice of targets for training is somehow sparse.
As a result, having gathered statistics, we can come to a conclusion that the purpose of training maximum barans is not always good. But here is another question, how to find the very goal, so that the NS would work well in the future. There is only one option, as they say.
We start a network and monitor its trading and make a conclusion as to whether it works now or not. Like this .....
As the result, profitable settings of the net in training look like this So, you may train the net for this balance and it may gather in the future. But there is an achilles' heel, how do we know what the balance should be.... and you can't train the TC in NS like that. There is a difficulty with the target training functions. If only in another program to try....
Or like this, in both cases the baoan curve in the optimised area is extremely different, and certainly not in the plus. However, this has not stopped the network in the future. Bottom line plus....
It seems to me that if I collect statistics for sections of the network, I can get a general similarity in the balances we need to establish a certain pattern. Thus, we should not train the network for the maximum profit at a certain part of the net but find the hidden signals that will work in the future. But not at the moment. So, here is the trick. Forcing the network to find parameters that don't work now, but could have a useful impact on the network in the future..... There's a lot of statistics to gather, and they're far from trivial......
In any case, by solving this problem (selection of optimisation results) by at least 60%, you can get a tool for trading and it is not bad at all.
I do not know about other packages, but when optimizing in NS, I get several results, starting with a negative balance in the optimization area. If only it were possible to get all optimization results in NS. And then choose from them, purely visually by balance. Then you could filter out useless results for luck. Put the network in operation and watch which of them will start pumping. The one that starts leaking gets shut down... Or whatever..... It is a bummer that NS does not have this option. At least save indicator parameters to a file during optimisation....
Again, it doesn't mean that in the optimization of the 1st plot you will get the values you got in the optimization of the 2nd plot. So there are a lot of unresolved questions here......
I mean, here's the trick. Getting the network to find parameters that don't work now, but could have a useful impact on the network in the future..... There are a lot of statistics to gather, and they are far from trivial......
this is how she will remember after training that they don't work... in nhs in ts is not a network but a ga...
that's how she'll remember after training that they don't work... in nhs in ts not nets but ga...
I agree.... there seems to be a lot of freedom, but what you need can't be tweaked....
But the author of the article, in my first "diagonal" view (and I don't have time for a second detailed view yet) looks at prediction with ANN from the right angle, in terms of Intelligence - using indicators, some "derivatives" of the quotes.
But, purely intuitively, I doubt it's possible to forecast at 98%. 80% probably yes, but not 98%. Maybe the author has dragged the research results a bit - I am quite willing to admit that. But to verify these results, you need to do the same experiments as the author - only then you can say something "hundred percent".
this is an illusion... Integer was right when he said it - the plot is left for the forecast...
In reality, everything is verified by a primitive example... you may use any method - from the second or third article ))))
if we sample up to the first yellow vertical line, we get 90-95% the same... move to the next one... apply the same algorithm... we get less... the last section (growth) the network will never predict...
all this is for predicting several steps for 1 VP (entry from the same) of course... If you highlight trough peaks (cyclicality), you can still capture growth in principle...
Agreed.... there seems to be a lot of freedom, but what's needed can't be tweaked....
the black box is this nsh...
it's just that there is some cyclicality ( dynamics ) ... so if this dynamic gets into the model (ns, regression or others... basically it makes no difference) and then continues for a while... then fine... if it changes, then it fails... reasons for changing dynamics 2...market and DT filters...
Yes, I agree with you.
This is the second explanation for the performance: a period was deliberately chosen for testing in which the network was locally successful in showing a staggering result. By the way, it is strange that the length of the test period is 150 points and not 200 or, for example, 451. It turns out to be a latent fit.