Please tell us your opinion. - page 5
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
The result is nice, but only the market knows what will happen next. For two years I have been writing EAs based on the same principle and I have noticed one peculiarity - abrupt alternation of profitable and failing sections. Sometimes it seems to be a real grail, half of a year of forward deals is so beautiful, and the next six months... My range with 50 profitable trades may be changed by another one with losing trades. Why is it so? I don't understand it completely, but I suspect that from time to time the market acquires a kind of "spontaneous nature", when its behavior slightly depends on the previous period, which is used by NS for learning/learning/adaptation (depending on the realization).
In any case the result looks good, and chances of robustness are there, don't listen to anyone :) (IMHO).
So, in your opinion, by no criteria is it possible to determine the future workability of the TS?
Well, I do not believe in the existence of eternal TS, with or without neural networks, it is not important. Any TS has a lifespan, like everything in this world. So I wrote that we had trained the TS and got good equity during the training period and out-of-sample period, but what next? After all, this all does not guarantee any work in the future. For example, I don't have a problem making a TS and getting good equity on OOS and training. Even very good on OOS. But this TS does not always work well on the real market. It may be better or worse. It is not clear what it depends on. So I'm struggling with the question - how to evaluate and calculate whether this TS will work in the future?
Неее, матожидание это не то.....
The expectation of winning is the mathematical expectation of winning. This statistically calculated indicator reflects the average profit/loss of one trade. It can also be thought of as reflecting the expected profitability/loss of the next trade.
Right. The average result of a trade (its "expectation") is the total result on the testing interval divided by the number of trades (i.e. net profit with a sign divided by the number of trades). Frankly speaking, I do not understand what is the problem and how what you are saying is different from what I am saying. The expectation is about the same as the average, and it has no direct bearing on the future.
Metaquotes has a slight terminological inaccuracy here: it's not the m.o. of a win, but an estimate of the m.o. of a trade (since the average trade can also be a loss). Just take and divide the figures in the report by each other. The result is exactly "m.o.s. winnings".
In general, we can only talk about m.o. when we already have a probabilistic model of the phenomenon, i.e. it is a purely Terwerian notion. We do not yet have a model, but only a sample of the mythical population, so we have - statistics. So, we can only speak correctly about expectation estimation.
So, in your opinion, it is not possible to determine the future performance of the TS by any criteria?
Well, I do not believe in the existence of eternal TS, with or without neural networks, it does not matter. Any TS has a lifespan, like everything in this world. So I wrote that we had trained the TS and got good equity during the training period and out-of-sample period, but what next? After all, all this does not guarantee any work in the future. For example, I don't have a problem making a TS and getting good equity on OOS and training. Even very good on OOS. But this TS does not always work well on the real market. It may be better or worse. It is not clear what it depends on. So I'm struggling with the question - how to understand, evaluate or calculate whether the TS will work in the future?
Definitely it is impossible, it is possible with some degree of probability to hope for a favorable result, but this probability in most cases tends to 50%) Probably with the rare exception of pipswise systems, but so they work with noise, not with the movement, although it sounds paradoxical, and from a practical point of view, they do not stand criticism.
But I'm counting on an eternal system). One of directions for further work I've chosen refusing from all trainings, relying on selflearning and full autonomy of the trading system. I hope to increase the likelihood of the system working in the future by using this method. Time will show how it will work out)
And again, your result is not bad anyway, and it is worth trying it in trading (I would try it for sure). If you discard such results, what is there to trade?
I am counting on a perpetual system). One of the directions of further work, I have chosen to refuse all training, all bets on self-learning and full autonomy of the trading system. I hope to increase the likelihood of the system working in the future by using this method. Time will show how it will work out)
Well, self-training is also a training. Only from one side only. Anyway, we must choose the criterion of self-learning stopping. When it (the system) must complete its self-learning in order to work profitably in the future?
I am counting on a perpetual system). One of the directions of further work, I have chosen to refuse all training, all bets on self-learning and full autonomy of the trading system. I hope to increase the likelihood of the system working in the future by using this method. Time will show how it will work out)
Well, self-training is also a training. Only from one side only. Anyway, we must choose the criterion of self-learning. When it (the system) must complete its self-learning to work profitably in the future?
Also true of course... But in my opinion, the "fitting" component of such systems is introduced exactly during the initial training/optimization, when the system parameters are selected that remain unchanged, be it the period of the input indicator or even the network topology itself... Here you have training late 2007, test 2008 -> uptodate. And even if in the process of work adaptive capabilities of the system appear and some parameters are changed (for example, weights of synapses are adjusted, probabilities are changed, and something obtained during the initial training in 2007 remains unchanged, right? Otherwise there would be no point in this training. That's exactly what I'm trying to get rid of with eternal learning.)
Except the systems are heavy, but that's probably out of the scope of this thread....
Also true of course... But in my opinion, the "fitting" component of such systems is introduced precisely at the initial training/optimisation, when the system parameters are chosen to remain unchanged, be it the period of the input indicator, or even the network topology itself... Here you have training late 2007, test 2008 -> uptodate. And even if in the process of work adaptive capabilities of the system appear and some parameters are changed (for example, weights of synapses are adjusted, probabilities are changed, and something obtained during the initial training in 2007 remains unchanged, right? Otherwise there would be no point in this training. That's what I'm trying to get rid of with eternal learning)
I agree. I too believe that only adaptive TCs can hold up in this market for a longer period of time. But still the question arises - how to set the criterion to stop training or self-learning(whatever you want to call it). After all, if this criterion is unclear, it will be impossible to program it. After all, the longer the longer the TS trains or self-learns, the more likely it is to fit the historical data and the more likely it is to fail in real trading.
Adjusting for historical data and adapting are different things. The Russian language is rich and it is not for nothing that it is called different words.
Yes, thanks, I know it's different......The question is different. I seem to have written without mistakes)))))))))))))))
It is difficult to tell anything about the robustness of the system from the shape of the curve. Robustness is in the idea. For NS, the idea is the correct preprocessing of the input data, with topology etc. taking second place. The input data should most accurately and unambiguously describe the market process being used, and for that you need to have an idea of it. And the right criteria for system rejection rule.
It is difficult to tell anything about the robustness of the system from the shape of the curve. Robustness is in the idea. For NS, the idea is the correct preprocessing of the input data, with topology etc. taking second place. The input data should most accurately and unambiguously describe the market process being used, and for that you need to have an idea of it. And the right criteria for system rejection rule.
You have written it all correctly. I agree. But it's all too general to find the right solution. As they say "about everything and nothing". I want to be specific.
In addition to the curve shape, you can see a lot of other useful information. Profit, number of trades, mathematical expectation, etc...... It seems to me this information may be useful when estimating working capacity of my TS in the future. Or am I mistaken?
Although the last sentence is not clear.......