Building a trading system using digital low-pass filters - page 17
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
I need one that allows me, after testing two systems A and B on H4, to firmly and confidently say: "System A with 73% probability at any time interval of 1 year will show drawdown of more than 30%", or "System B with 61% probability at any time interval of 1 year will not exceed 6%, with 94% probability - not more than 18%, and with 99.9% - not more than 37%". I would bet on the second one...
That there are no systems capable of generating money forever, I probably agree with you, bstone. But the fact that systems that statistically guarantee limited drawdowns over a certain set period of time are possible in a non-viner process, I'm somehow convinced. But in a Wiener process you can't give such a guarantee either...
The only stationarity test I know of is the Dickey-Fuller test. But it assumes some model of the process (in this case, a 1st order autoregression). But what if the model is unknown to us beforehand?
...
Hmm, but a model for making money would be simpler :). Because it's not intended to reproduce all characteristics of the market (like the model designed to test an arbitrary TS), but only those essential to make money.
On point 2: yes, but that already implies some kind of algorithm designed to reveal these invariants. In the case of two dummies these are some invariants, when using additional indulators they are others. And if it is ZZ+Fibo, then these invariants are very complex and testing by this idea is very difficult.
For now, it is enough for the model to reproduce the stationary process in the broad sense (MO, RMS, ACF).
Well the one that will claim stationarity. At the moment it is (doubtful, but still hopeful).
By the way. Here I remembered, poked around... S.V. Bulashev in his book "Statistics for Traders" gives an example of proof (using Pearson's criterion) that logarithm distribution of price ratios follows the exponential distribution law.
However, the word "non-stationary" appears once in his book - he admits that "the dynamics of exchange prices of assets can be represented as a stochastic and non-stationary process".
So two fresh thoughts:
Well, here is the simplest linear transformation, which transforms non-stationary into stationary: Wiener with independent increments is non-stationary, while the series of its first differences (returns) is the real stationary, Gaussian noise.
to Northwind
There is a predict (,,,,) function in Matkadec, yes you probably know it works based on Berg's (or Burg's, generally called Burg ) method.
to Northwind
There is a predict (,,,,) function in Matkadec, and I'm sure you know it works based on Burg's method (or Burg, generally speaking).
A little clarification, I suggested this to Prival (out of spite and will remind him on occasion):
Try to collect statistics by this method (it's based on autocorrelation), but you should input the signal itself (filtered). I warn you, it will lie, but once again it depends on what to consider as a correct result. But if we pass from the forecast series (the forecast horizon is given) to some generalized characteristics of the signal, and from generalized characteristics pass to levels (no method will ever predict the price series precisely), then it may work out fine. Quite nothing. This was entertained as an elective, in my opinion - not a bad approach, quite scientific. Along the way collect statistics, maybe it will become clear when it makes sense to make a prediction, when it does not.
PS (addendum):
Or try predicting with this method MA, but obtained on signal after filtering. Also nothing. Knowing the "accurate" MA prediction - you will "accurately" recover the future BP (within accuracy limits)
I didn't make MT-prediction based on Burg's method, though I got satisfactory results in MathCAD. Proposed "strategies" I tested on hours at ( H + L )/2 (as a rule calculated levels are very far from "the current price" and generally there's no need to simulate the process directly (ticks, minutes ...). The forecast level is tested on every bar (counting) (I always test it that way and I recommend you). I'm confused only by a huge number of input parameters, starting from VLF and finishing with the input parameters for the forecast. I was interested to see how it works (I've seen other methods too), in general it seems OK, but for me it's optional (it's not difficult to write in MatCAD about ten lines, equivalent to MT "sheet", sorry - I'm praising MatCAD again). But once again I'd like to remind, it's useless to forecast series, no forecasting method can do it - we must necessarily and very skillfully go to generalized characteristics, in other words - to some price level.
I only have a general idea how to implement this method, for now it's enough.
Here's a small example: we are forecasting MA with a 90-sample window, a series of 500 samples is taken as initial BP. We can get some input characteristics for the forecast based on ACF of the initial one. As a result we get:
The predicted MA is the red curve after the vertical line (like the current count). It can be compared to the grey (true) MA after the current count - we can see that there are very good coincidences. And knowing the "exact" predicted MA and the original (current) series - you will "exactly" reconstruct the future BP. So, I recommend to take a closer look in this direction, but preferably with the limitations described.
Addendum: I have decided to increase the carina:
PS : Returning to strategy based on ELF (what the current topic is about), here's where I once shared some meager thoughts https://www.mql5.com/ru/forum/51428 may be of interest to anyone