Econometrics: one step ahead forecast - page 26
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
In what other sense can we talk about cyclicality? Do you think cyclicality is when "the market goes up, then down, then back up again"?
The point of trading is the same - to make money, you need to conform to other people's mass trading methods. You need to predict when there will be massive selling or buying, opening positions or taking profits or losses. Other people's buying or selling moves the market, so they are the target of prediction, not deterministic components, etc. It is likely that mass trading techniques can be described within an econometric model.
I agree completely, but as soon as we saw it, the reaping ended and we got leftovers again at best, and at worst, everything had already gone stale.
No one is stopping us from making the model continuously adaptable. But adaptability must also be logical. Identify which methods are gaining popularity and their most popular average parameters. I.e. again, as you say, so that the model is adequate to the quotation today.
How does HP even know when people will drive a cotier and where? :)
Based on 4-5 forecasts with huge errors ....
Forecast error = 59 pips. Is this a lot or a little?
1.4513 - 1.3190 = 1323 pips
Standard deviation = 344 pips.
Is 59 pips of forecast error a lot or a little?
Let's take the series increments and break them into groups. We get:
We see that increments less than 100 pips = 47 out of a sample of 77 bars and 13 bars had increments over 200 pips. And I have an error of 59 pips. At the same time I am predicting the direction and not the target.
But the most interesting question is whether it is possible to estimate the minimum error of a given section of the quotire. If we know it and have found such a model, then end of modelling.
No one is stopping a model from being continuously adaptable. But adaptability must also be logical. Identify which methods are gaining popularity and the most popular average of their parameters. I.e. again, as you say, so that the model is adequate to the quote today.
How does HP even know when people will drive a cotier and where? :)
In reality you and I do not disagree on anything. I took the price and you can take (add) the activity details.
We don't know the future, but we have a past. If the past has a trend, then prediction is possible, if it is pure random walk (no drift), then prediction is not possible. and it doesn't matter what data is used in the model.
C-4 in this thread posted the link and I'm advertising. The author has found cyclicality in quotes in addition to price trends and used it, and his cyclicality allows him to predict future market crashes.
OK, let's say 59. Here is your table:
Such an error, almost invariably, corresponds to different movements, from 30 to 102 pips. I don't care at all about your er-square, which is very high all the time. Its informativeness is not very high and you know that too.
That's not what I'm talking about. You are grasping at the autocorrelation check with a dead hand to remove the dependence in the residuals. And I'm saying that's not enough, because Pearson autocorrelation only explains linear dependencies, not all of them. In the branch about feature selection alexeymosc already gave an example when dependences calculated by information theory (not only linear, but all in a row!) were extremely high even at very large lags. The vast majority of participants in that thread, including its author, agreed that it was all about volatility. (In this model, by the way, there are no notions of trend/flat, although they can be drawn there as well).
I still don't see sufficient grounds to confidently say that volatility is to blame. Perhaps on the daily, yes, but I've said several times that the mutual information on the daily is significantly less than on the watch or 4H. Almost nobody was interested in it, and all results were still only posted for days. So the conclusions were accordingly, i.e. incomplete.
Give me an example of "content" and "physical/economic sense"
Now I focus on another model, because the "information" model is not so easily interpreted in the light of existing and mind-dominating hypotheses of market efficiency.
It is the self-organisation of the whole market under a strong attack on a given currency. And it is in this model that I see the economic sense so far, and it is very simple: if there is strong news for chif, then all chifocrosses will also move in a more or less coherent manner according to the "vector" of the news.
The meaning highlighted in blue is revealed by combinatorial methods, and it can be modelled by physical analogies (e.g. thermodynamic). That is why it turns out to be "physico-economic". I won't give any details here, as there are many of them and there are enough nuances.
Mathemat:
That's not what I'm talking about. You are grasping at the autocorrelation check to remove the dependence in the residuals. And I say that this is not enough, because Pearson autocorrelation explains only linear dependencies, not all of them. In the branch about feature selection alexeymosc already gave an example when dependences calculated by information theory (not only linear, but all in a row!) were extremely high even at very large lags. The vast majority of participants in that thread, including its author, agreed that it was all about volatility. (In this model, by the way, there are no notions of trend/flat, although they can be drawn there as well).
I still don't see sufficient grounds to confidently say that volatility is to blame. Perhaps on the daily, yes, but I've said several times that the mutual information on the daily is significantly less than on the watch or 4H. Almost nobody was interested in it, and all results were still only posted for days. And the conclusions were accordingly, i.e. incomplete.
Now I concentrated on another model, because the "information" model is not so easily interpreted in the light of the existing and mind-dominating hypotheses of market efficiency.
It is the self-organisation of the whole market under a strong attack on a given currency. And it is in this model that I see the economic sense so far, and it is very simple: if there is strong news on chif, then all chifocrosses will also move more or less coherently according to the "vector" of the news. But this economic sense can be modelled by physical analogies (e.g. thermodynamic). That is why it turns out to be "physico-economic". I will not set out the details here, because there are many of them, and there are enough nuances.
It's not so obvious about adaptation. Why do you think that the "information" model discussed in the feature selection thread has no adaptation? Yes, and a lot of it. You don't have the faintest idea about the prediction algorithm (it wasn't even mentioned in the thread).
SunSunich, you have no idea what anguish I go through trying to justify/refute the robustness of the relevant system. This is not a standard econometric model, for which all the necessary statistical tests are already more or less known (although they are of no use). But I will have to do it anyway.
Detrending in econometrics has a clear goal - to obtain a series as similar to I(0) as possible, and without dependencies (you like to say "autocorrelations"). But detrending in the "informational" model is already done, because we analyze not the levels of the course, but the returns. As for the returns series, it is hard to say that it is I(0), as there are a whole carload of dependencies there. And I have no idea to what extent they are "stationary", because stationarity for them is not so easily defined.
In short, there are a lot of problems. But success only comes to those who do not follow the beaten path. So I'm trying, relying on my own strength.
You will be raking in the money, because you are the only one who will be using this theory.
What's wrong with that?
It's not so obvious about adaptation. Why do you think that the 'information' model discussed in the feature selection thread has no adaptation?
I'm not talking about model adaptation, I'm talking about market adaptation. There is an opinion that if some unusually winning trading strategy emerges, as soon as a large enough number of traders start using it, that system will cease to be a winning one.
SunSunich, you have no idea what anguish I go through trying to justify/contradict the robustness of an appropriate system. This is not a standard econometric model, for which all the necessary statistical tests are already more or less known (although they are of no use). But I have to do it anyway.
That's what I meant in the posts in your thread. All this may be a thesis, but it is not possible to write a thesis communicating on a forum - you need a qualified team of opponents. To make a trading system based on new principles alone? The days of loners are long gone.
In short, there are many problems. But success comes only to those who do not follow the beaten path. So I'm trying, relying on my own strength.
It's a very expensive idea. You can buy a bicycle and ride it for your business, or you can invent a bicycle, but you won't have time to ride it and it won't be fun.