Statistics as a way of looking into the future! - page 18
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
No, you misunderstand me. Large pictures uploaded to this forum are compressed to fit on this page. But if you click on them, they will open in all their glory :)
I did it!
I wish the author would tell me how he got this beauty... - In some places the mouving does not lag behind the cotier at all.
I apologise for not being able to maintain the discussion promptly, absolutely no time to look at the forum. Although delayed, I will try to answer questions.
Neutron - "For me, for example, a one-step-ahead prediction with a prediction at each step is relevant. In this formulation, NS is probably out of competition."
I too subscribe to this concept, but I don't take the position that NS gives the best results, e.g. prediction by Kohonen maps (when they are not used to render pictures, but models are built from them) are much more accurate and smoother than that of NS. Linear regression sometimes gives good results too. As an example I can give some more pictures of the joint use of LR and NS: the blue line is the initial signal plotted by the indicator, the red and yellow lines are forecasts of that signal with different horizon. In the archive there is a file corresponding to the figure for H4, the first column contains data corresponding to the blue signal, the second one contains yellow data, the third one contains red data and the fourth one contains Close. Unfortunately, I have no time to evaluate the profitability in the cloud and calculate it, in case you are interested you can do it with the help of this file.
I use my own indicators as a source of signals (I've never used any standard ones or other ones). However, to call what I use a simple indicator is a bit simplistic. It is rather a system for modeling signals based on quotes that are processed in a block with feedbacks and self-tuning, as well as adjustment of some parameters according to the equalizer type allowing a user to form the necessary signal form, of course with some limitations, because the basis is a real quotes flow and the task is to increase the information value of data and filtering with minimal lag using predictive models.
Modeling block uses the principle of group consideration of arguments, i.e. like GA it uses not the best model, but a group, even if not the best one, because the market is volatile and with time some not best signals become the best and vice versa. Besides, I try to get the maximum diversity of signals covering the entire variation range of the target function both by phase and amplitude relative to which the models are trained. In general, the system has a hierarchical tree structure with LR- and NS-based models in the nodes of the branches. As an example of the spectrum used for modeling as input signals for LR and NS training I give a fragment of the figure, black colour (though hardly visible) shows the target signal, all others are derived from the quotes passed through various models. Calculation of models is performed at the zero bar by ticks, but due to the complexity of models not all ticks have time to be processed, but it is not essential - upon coming of a new bar the values are fixed and do not change further. The scaling factors introduced into models and corresponding to timeframes allow switching from one timeframe to another preserving the constant scale, phase and amplitude characteristics of signals without any adjustments.
I'm encouraged by the results presented in the first example, I've never done such tests before and did not think that regression models, or NSs, could be so stable during long-term work without retraining. In my estimation, the dollar was artificially supported and strengthened for the US election. Now the resources to support it have run out, plus the crisis is preventing further strengthening. So, I don't think thatEURUSDwill fall much further , after elections, which are coming soon, dollar will start to fall a bit, though not very much, as production and consumption is decreasing because of the crisis, oil prices are dropping. Further significant fall of the dollar will begin when the world's financial system will recover from the crisis, and it will not be soon, but meanwhile the fluctuations will be in the range of 1.3 - 1.5, and I find it encouraging, because I trained all the models of LR and NS in this system based on H4 data, I took 5000 bars from July 18, 2005.This means that all my models will work stably without retraining until the price significantly deviates from this range, and the LR, as shown by the example, can work well with a significant deviation from the training range. Although training was conducted on H4 models work adequately on all timeframes So the system built on this basis will be stable without retraining for many years, this is encouraging.
I apologise for not being able to maintain the discussion promptly, absolutely no time to look at the forum. Although with delays, I will try to answer questions.
Piligrimm, thanks for the informative post and especially for the data file. I will think over it and analyze. I think there will be questions soon.
So, Piligrimm, we have the initial time series (TP) - Close H4 price (black points), a muwwing smoothing the initial TP according to some algorithm (blue line), and a series of forecast values built by analyzing the muwwing for initial TP one step ahead for each bar H4 with different parameters of NS settings (red and purple lines).
So, looking at it, nothing bad can be said about the algorithm so far...
Let's build a TS that will open and close a position on each bar H4, in the direction of the prediction, which is set by presented predicates (or predicates?). It is clear that the task includes prediction accuracy and BP volatility on the selected TF. Then having plotted the increment of price in pips on the abscissa axis and the prediction of this increment on the ordinate axis we will get the prediction cloud and using its slope tangent we will evaluate the TS yield in pips per transaction.
Assuming the instrument volatility of 30 pips, the return for the regression line is 1.4 pips/transaction, Predict1 is 6.6 pips/transaction and Predict2 is 10.7 pips/transaction.
If the author is not mistaken in preparing the data, the TS which is based on this NS-algorithm, will bring up to 8 pips of average profit every 4 hours for EURUSD, taking into account the spread, with a risk of +-30 pips during the same time. I.e., the balance line will grow at a rate of 40 pips a day and hover around this line with an amplitude of +-100 pips. The overall view of the balance curve found from the estimated integral characteristics is shown by the red line in the figure below. For comparison, the blue line shows the balance curve plotted by the "fair" TS trading according to the data provided by Piligrimm.
The results coincide well, indicating the adequacy of the suggested integral method of evaluating the profitability of TS by the slope angle of the predictive cloud.
Generally speaking, the profitability can still be considerably increased by requiring from the TS not to close an open position if the Predict of the next price increment coincides with the direction of an already open position.
The algorithm implemented by Piligrimm is very good! There is a lot to strive for.
But whatever it is, it works!
Nothing prevents the author from using this algorithm as I suggested and everything is fine:-)
The problem may lie in something else, namely: the author could implicitly use Open, High, Low or Close to construct a Predict of the same bar... then it's all for nothing! I.e. to build a prediction with a "peek", e.g. to use High or Low of an already formed bar. But I think the author will soon dispel our fears.
The results coincided well, which shows the adequacy of the proposed integral method for estimating TC returns by the slope of the predictive cloud.
Agreed. This is a self-evaluating result. Neutron, it would be good to formalise the method in the form of an article detailing the methodology of practical application. This could become a standard as it is spread "among the masses". At the same time, the TC position opening can be considered as the prediction of an average profit trade on the next bar (the interval equal to the average order lifetime). Then the method can be made universal. We are obviously missing such indicator for evaluating TS today and the development of your idea seems to be very versatile in this sense.
P.S. As an option, on the fuzzy evaluation scale the right side could appear "for real!" and on the left side "ftopkus!" :-)