![MQL5 - Language of trade strategies built-in the MetaTrader 5 client terminal](https://c.mql5.com/i/registerlandings/logo-2.png)
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
here it is at the bottom http://www.wikimapia.org/#lat=31.3536369&lon=-24.3786621&z=8&l=1&m=a&v=2
nothing to be sorry about with neural network indices
You better invent a math function for forecasting.
No, I'm done with forecasting, only novels now.
Please tell us the intervals for the variables in Kristograf.
What kind of intervals do you mean?
Piligrimm, just one question. In your opinion, is it possible to predict a LUCKY value?
Answer options:
1. Yes.
2. No.
3. For who is casual and who is not!
Perhaps my conclusions are hasty, but I had time to play with PolyAnalyst and got two results that stunned me. A simple CSV-export of quotes from the terminal for 30-minutes was fed and a polynomial was constructed using their OHLC that calculates closing price of the next bar by the data of the current one. The function is (surprisingly) not very complex and I immediately pasted it into EXCEL. To get the polynomial I took only the first half of the data from the CSV-file, in order to check the correctness of the prediction in the second half.
The first baffle was that the polynomial gave the following Close price which in the vast majority of cases (over 90%) gave an error of only a few points!!! I could not believe it. I struggled for a week with different times and pairs and the result was just as impressive. At the weekend my brain relaxed a bit, and on Monday it suddenly dawned on me: I give real quotes to my function and it gives me the close price of the next bar, but it's a real peep into the future :) I updated the model a bit, got polynomials for all four components of OHLC and a new calculation of the function and did not input data from quotes into it, but data that it had calculated itself at the previous step. Here I was in for a second shock - I was not getting anything close to a stock chart. The forecast was a slowly falling down very smooth and fluent curve. Unfortunately, I can't remember where this result is now in order to attach screenshots. After that there were a lot of attempts to make it work properly. If I find it, I will definitely add to the post.
Maybe I did something wrong, I spent a couple of weeks to get acquainted with PolyAnalyst. But I have a feeling Pilligrim is stuck in the (my) first phase :(
So, a question to the outgoing Piligrim: how correct are your models in light of what I managed to obtain? Maybe you also have every time the data known "in the future", so to speak "correct" data and they correct your forecast all the time? Or am I wrong about something?
I can share a similar story that happened to me while studying the Caterpillar time series analysis package.
The first impression I had when I looked at the results of the "forecast" was Euphoria! I could see myself in the Canary Islands... Even though I knew what I was doing, the package was persistently (as it turned out during the detailed analysis) peeking into the "future" inside itself, although in the program settings I had strictly limited the analysis and comparison areas. Further, I learned how to prepare food for the caterpillar correctly and the miracles disappeared. The prediction accuracy is 50/50.
I think the creators of the package intentionally turned a blind eye to this "little" bug, which when you first encounter this miracle of mathematics inevitably leads to a desire to give $100 and quickly return it in Forex. I remember showing all my acquaintances beautiful drawings and proving the obviousness of imminent riches for myself and for them, if only they are friendly with their heads :-)
I agree with ForexTools 100% that the reason for so many of our "woes" in numerical modelling, is the hidden peek.
This is one of the last variants where the MA was "predicted"
Here is the truth of life :((((
To anyone wondering - the formula in Excel was like this:
= (1.00002*IF(AND(0.00026974 <= 1*G2,1*G2 < 0.00026974 + 141.828),1/G2,0.121066)*G2*G2*G2-1.25485*IF(AND(0.00026974 <= 1*G2, 1*G2 < 0.00026974 + 141.828),1/G2,0.121066)*IF(AND(0.00026974 <= 1*G2, 1*G2 < 0.00026974 + 141.828),1/G2,0.121066)*G2*G2*G2)/(IF(AND(0.00026974 <= 1*G2, 1*G2 < 0.00026974 + 141.828),1/G2,0.121066)*G2*G2-1.25521*IF(AND(0.00026974 <= 1*G2, 1*G2 < 0.00026974 + 141.828),1/G2,0.121066)*G2+0.000389753-0.00000823394*IF(AND(0.00026974 <= 1*G2, 1*G2 < 0.00026974 + 141.828),1/G2,0.121066)*IF(AND(0.00026974 <= 1*G2, 1*G2 < 0.00026974 + 141.828),1/G2,0.121066))
P.S. Or am I still very much mistaken?
Piligrimm, just one question. In your opinion, is it possible to predict a LUCKY value?
Answer options:
1. Yes.
2. No.
3. For whom is random, and for whom not!
The answer is number 2, BUT, first, purely random processes used in practical life, not so much; secondly, even if you face one, sometimes it is possible, by applying special processing methods, to make the process quasi-random - then there is a chance to predict it with higher or lower probability. For instance, when I started to predict lotto sports, and the process itself is random, I didn't do well for a long time. Eventually I realised that I couldn't solve the problem with lobbying. Then I started looking for workarounds. Having statistics of draws I've started to model the process of working of the barrel, and I used this model as a modulating signal that processes history data; moreover I've introduced several deterministic signals, that can be easily predicted, and that give a quasi-random process in covariance with a random statistics signal. And using the Method of Grouped Argument Counting (i.e. genetic algorithm), I succeeded to reveal some regularities in the work of the system as a whole, which allowed selecting the group of numbers, which could appear in the next draw, with high probability. As a result of the forecast, of course, I did not get the value of 5 numbers of the next draw, I got about 10 most likely numbers in the output, of which it was necessary to make all possible combinations of 5 numbers to fill the tickets, and as a result only one of the tickets of this combination was a winner with an accuracy of 3-4 numbers out of a possible 5. But nevertheless, it worked.
And when it comes to the Forex market, I don't consider it a random process at all. It's very complicated, multi-factorial, multi-dimensional, there's a lot of distortions and interference, we don't have the fullness of information, and often it is very delayed, but with all that - Forex market is not random!
Answer number 2, BUT, first, purely random processes used in practical life are not so much; secondly, even if you face one, sometimes it's possible, using special processing methods, to make the process quasi-random - then you have a chance to predict it with higher or lower probability. For instance, when I started to predict lotto sports, and the process itself is random, I didn't do well for a long time. Eventually I realised that I couldn't solve the problem with lobbying. Then I started looking for workarounds. Having statistics of draws I've started to model the process of working of the barrel, and I used this model as a modulating signal that processes history data; moreover I've introduced several deterministic signals, that can be easily predicted, and that give a quasi-random process in covariance with a random statistics signal. And using the Method of Grouped Argument Counting (i.e. genetic algorithm), I succeeded to reveal some regularities in the work of the system as a whole, which allowed selecting the group of numbers, which could appear in the next draw, with high probability. As a result of the forecast, of course, I did not get the value of 5 numbers of the next draw, I got the order of 10 most probable numbers at the output, of which it was necessary to make all possible combinations of 5 numbers to fill the tickets, and as a result only one of the tickets of this combination was a winner with an accuracy of 3-4 numbers out of a possible 5. But nevertheless, it worked.
The number of combinations of 10 to 5 = 252. How much was a Sportlotto ticket worth? 50 kopecks, I think. And the three-digit winnings? I think it was 3 rubles. So, to have a high probability of winning 3 roubles, you'd have to spend 126 roubles. It's clear that this works, as well as why you didn't have to become a millionaire.
My friend, by the way, bought just one ticket, guessed 5 out of 6 and got 2400 roubles, after which he never bought Sportlotto again ...
Perhaps my conclusions are hasty, but I had time to play with PolyAnalyst and got two results that stunned me. A simple CSV-export of quotes from the terminal for 30-minutes was fed and a polynomial was constructed using their OHLC that calculates closing price of the next bar by the data of the current one. The function is (surprisingly) not very complex and I immediately pasted it into EXCEL. To get the polynomial I took only the first half of the data from the CSV-file, in order to check the correctness of the prediction in the second half.
The first baffle was that the polynomial gave the following Close price which in the vast majority of cases (over 90%) gave an error of only a few points!!! I could not believe it. I struggled for a week with different times and pairs and the result was just as impressive. At the weekend my brain relaxed a bit, and on Monday it suddenly dawned on me: I give real quotes to my function and it gives me the close price of the next bar, but it's a real peep into the future :) I updated the model a bit, got polynomials for all four components of OHLC and a new calculation of the function and did not input data from quotes into it, but data that it had calculated at the previous step. Here I was in for a second shock - I was not getting anything close to a stock chart. The forecast was a slowly falling down very smooth and fluent curve. Unfortunately, I can't remember where this result is now in order to attach screenshots. After that there were a lot of attempts to make it work properly. If I find it, I will definitely add to the post.
Maybe I did something wrong, I spent a couple of weeks to get acquainted with PolyAnalyst. But I have a feeling Pilligrim is stuck in the (my) first phase :(
So, a question to the outgoing Piligrim: how correct are your models in light of what I managed to obtain? Maybe you also have every time the data known "in the future", so to speak "correct" data and they correct your forecast all the time? Or am I wrong about something?
If you only knew how many times I have been euphoric about the results over the past 10 years! But almost as many times I have been disappointed. I've learned a lot on the subject, but I haven't wanted to deceive myself and I have no reason to deceive you all. What you describe is not a problem with PolyAnalyst, it's a problem with any method of quote processing, if the input data is not prepared completely correctly, and the targets are not set correctly.
The main thing is what you are aiming at with the forecast. If your aim is to make a forecast for the next 10 or 20 bars using data on M30 timeframe with at least 60% probability, I think you will not succeed. The market is very unsteady, and most importantly - volumetric and complex, the players who are present now and determine the main trend, may leave in an hour, and new ones will appear with completely different characteristics. It is not possible to build an adequate market model with the resources at our disposal. Although it sounds like a contradiction to everything I have said before, nevertheless it is true and there is no contradiction.
I gave up forecasting the future values of quotes a long time ago. The purpose of my forecasts was not to obtain the results one step ahead, but to draw lagging signals, for example indicator readings to the null bar level, i.e. to forecast their behavior in the current moment without any lag, in the current moment we have a current market picture and do not go beyond the limits of information field, which hasn't been formed yet. At the end of the day, what matters in trading is not knowing the future, but reacting appropriately and timely to what is happening in the present moment. Look at Figures 1 and 2 on page 9, I did not give this example by accident. Fig. 1. is just an attempt to train a model to predict the future when the information field has not yet been formed, no data from this future is yet available, and in most cases the target signal is ahead of the available input data, it has nothing to latch on to and the trained model will have very low predictive ability, or if the input data is not formed correctly and there is a peek into the future, the model will give inadequate signals. You can see from fig. 2 that the target function in most cases does not go beyond the information field of input signals, the model has a chance to learn and reflect the trend accurately enough according to this set of input data and targets.
Speaking about how much I was able to pull lagged signals to the current time with my predictions, look at the first figure in this thread for an example. The red line is a filtered but very lagged indicator signal, and now imagine that this signal has moved 20 bars to the left on average, I said average because in some cases this shift may be 15 bars, in others - 25, this figure is not constant because the pulled signal cannot go beyond the current quotes, it is close to zero bar and keeps the characteristics of smoothness but does not go beyond it. That's how I was talking about predictability of Forex market, the more accurate the information we possess and the more powerful our computer resources are, the more accurate and less delayed we can build models of market tendencies, eventually we will be able to build forecasts to the right of the zero bar, But of course this will require not only quotes from one timeframe, like I did in this problem, but a lot of information about different symbols, timeframes and fundamental factors, but it is a completely different task that will require completely different resources and will be solved without me.
The number of combinations of 10 times 5 = 252. How much did a Sporting Lottery ticket cost? 50 kopecks, I think. And the three-digit winnings? I think it was 3 rubles. So, to have a high probability of winning 3 roubles, you'd have to spend 126 roubles. It's clear that this works, as well as why you didn't have to become a millionaire.
A friend of mine, by the way, bought just one ticket, guessed 5 out of 6 and got 2,400 roubles, after which he never bought Sportlotto again ...
I did not specify some details when writing the post.
1. I often got the result and less than 10 figures in the output, their number was not rigidly limited, there was a restriction that the probability of falling out the corresponding figure must not be below a certain point.
2. Along with digits in the output I also had probabilities, and when making combinations the weight of each digit was not equal, for instance if the probability of 3 figures out of 10 was accordingly 90 %, 85 % and 80 %, then these figures were taken as the base, and the rest were combined with additional ones to make the necessary 5 figures in the ticket, again taking into account their probability. So in the end all combinations were often made into 25 tickets, and often even less. But in general the dynamics of predictions was such, that profit from one winning with a correct prediction of 4 figures, overlapped a series of small losses from previous wrong decisions. (An analogy with forex.)
And I did not become a millionaire for another reason - I did not have my own computer, and on the computer, where I worked, was busy during 3 shifts with small production tasks, and even when I debugged small fragments of my task, it occupied CPU so, that other tasks could not go anymore, and I had frequent conflicts with other programmers, whose tasks, which required 2-3 minutes for calculation, were standing in line for hours because of my task. Therefore, for the full debugging and calculation, I came to the central computer on Saturday, enabled the computer, and quietly ran my tasks until Sunday evening.
And as far as I remember, the price for one ticket was 20 or 25 kopecks.