Testing real-time forecasting systems - page 36
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
Hmmm, too lazy to translate to MT for clarity, but it seems to be working, I mean, it has rallied strongly (from 12:00 MSC), it has already gone over the upper limit, but it should come back
I would have to convert this model to MT anyway (and the others wouldn't hurt :o/), it's not very convenient.
to Figar0
Рынок не может себе такую роскошь как быть всегда прогнозируемым, но для нас грустно не это, а то, что эти непрогнозируемые участки обычно наиболее волатильны...
Not really, volatile areas can be well predicted and vice versa - a lot depends on the model used. Also, the concept of volatility is very vague. I define it as aspiration to the maximum, {max(y)-min(y)}->max and in this sense the trend can only be in a volatile market
Of course the drawings are good but does anybody have any trading (auto, manual or tester) following their prognosis that they would not mind showing? How do you do it?
The idea of the theme was different - the focus is on models and forecasts, not accounts. I don't think it worked here, but I've seen it in the next thread :o). Yes, it is clear that the forecast should eventually give instructions on how to manage the order. That will be the next stage. Take your time - everything will be there, you will see :o)
Oh man, by the time you calculate, by the time you post it, time will have passed. 24 hour forecast refinement (15 minute countdown) from now, or rather from now minus twenty minutes.
The level of 1.5 is unlikely, but it's coming, let's see... Some strong activity ...
not really, volatile areas can be well predicted and exactly the opposite - very much depends on the model used. Also, the concept of volatility is very vague. I define it as aspiration to the maximum, {max(y)-min(y)}->max and in this sense a trend can only be in a volatile market
You're probably right, if we call things by their proper names, I meant just sharp strong movements...
The idea of the topic was different - the focus was on models and the forecast itself, not the accounts. It didn't come out here, but there's a lot of it in the next thread :o). Yes, it is clear that eventually the forecast should give instructions of order management. That will be the next stage. Take your time and you will see.)
I've understood the idea of the subject, it's just how else to evaluate forecasts quantitatively). I've got the Expert Advisor, ran it in the Strategy Tester and now I understand the price of forecasts. In my case it's vice versa, I don't have any pictures and just a few figures used by the Expert Advisor in trading. I wanted to transform it into images too, but I couldn't use them except for this branch; that's why I gave up. I've got a lot of nice images here, it won't work for me, the principle of forecasting is somewhat different. I do not have "dogma", the forecast will be redrawn all the time...
You're probably right, if you call things by their proper names, I meant just sudden strong movements...
I have three models in operation. One of them can theoretically see big gaps.
I got the idea of the topic, just how else to quantify the forecasts?) I have made some Expert Advisor, ran it in the tester and understood the price of forecasts. I wanted to change it to images too, but cannot find any use for it other than in this branch. I've got a lot of nice images here, it won't work for me, the principle of forecasting is somewhat different. My forecast is not a "dogma", it will be redrawn all the time...
I don't think that it's a dogma, it will be constantly re-drawn. You can only check it by testing trades on history and "pilot" exploitation, which is called "classic". There is no other way, you're absolutely right. I personally have two stages:
1. testing in MathCAD (prediction and trade on every bar)
2. testing in MT (but it's still necessary to translate the system under MT, and it's not so easy)
The matter is that I am not able to create fully automatic system (roughly speaking - machine) yet, so point 1. In MathCAD I evaluate deals by their worst variant (Open/Close, but it does not matter, the number of target points is large). I'm preparing to next iteration (gathering statistics and testing). But the biggest problem is calculation time - very long. There is a lot to solve, several problems need to be solved, including increasing resolution and clarifying pivot zones (just for the order)
Besides, it is useful to keep yourself busy by publishing and testing forecasts in real time. It's also very useful to communicate during testing, gpwr for example is a very useful tool.
Pictures are good of course, but does anyone do any trading (auto, manual or tester) on their forecasts that they are not "sorry" to show? How do you do it?
First of all, as you said yourself, it is impossible to predict everything 100%. Secondly, long-term forecasts are less accurate. That's why I prefer to distill my predictor on every bar to allow new information to refine the forecast. Here's what my predictor showed before the EURUSD spike, for example:
If a SELL SHORT trade had opened on the previous prediction, it would have closed with a small loss before the jump on the changed prediction.
And here is the new prediction:
I will try to code this prediction as a trading system.
Sounds about right. But in my opinion, the prediction algorithm itself has to play some role here (and it's not without it anyway!).
2 grasn: mmM seems to be going the same way. The dyfurc should work roughly like this: a predicted plot (response to a jump in Greek parameters Α, Β, Γ and Ψ), then a change in Greeks again, and again a predicted instrument adaptation to the new Greeks being at a constant level.
αβγδσλ is just for fun, a game with esc-sequences...I don't know, I've got a "hurry", the second prediction was obviously wrong from the beginning, the start was at 1.45, while this level wasn't present at the moment of prediction. (it happens sometimes, because it's considered a kind of "grid" and sometimes, when there is a significant roughening, the forecast "shifts" in some direction). But otherwise it's pretty normal, I haven't even recalculated anything since then.
to Mathemat
2 grasn: mmM, похоже, идет к тому же. Дифурка должна работать примерно так: прогнозируемый участок (отклик на скачок греческих параметров Α, Β, Γ и Ψ), затем - снова смена греков, и снова прогнозируемая адаптация инструмента к новым грекам, находящимся на постоянном уровне.
αβγδσλ is just for fun, a game with esc sequences...
Yes, yes, I know your weakness for esc - sequences. :о) Quite possibly. I'll tighten my knowledge in some areas a bit - and then Belinsky will make me jealous :o)
PS: By the way, take a closer look at the model discussed earlier, I specifically "rolled back" to history to check the detection of this spike, geez, only slightly underestimated, but the "classical model" didn't detect a damn thing. The model based on statistics (I've been showing its results here lately) - only detected it on a small period. This is the only way to detect spikes, there are no others. Or there are and I don't know anything about them. :о)
to Figar0
One of the ways of testing that I use is very simple. On the entire history, I select the most "extreme points", spikes, changes of global and significant local trends, and perform tests of the model for "lousiness". One of the models often sees such "catastrophes" but does not see any other peculiarities. But it has a serious problem - from 3-9 hours of calculation. Others do not see such "surprises" at all. Such is the case ...
to gpwr
I think I will soon be able to formulate the first questions on RBF networks. How do you prefer, ask here (if the author is against it), by e-mail or in a separate branch for general cognitive purposes.
4480 is a very strong target and may go lower