Where is the line between fitting and actual patterns? - page 21
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
For the particularly gifted:
1) in russkomazykenetlova osobogodnymi.
Well, you can once, but not five times in five consecutive pages.
..............
2) The harm of splitting periods.....
и
The harm of OOS tests....
agree these are different things.
Don't misrepresent reliable information. There isn't much of it already.... )
No, it wasn't, if the test on OOS was positive.
If there was a negative result on OOS, it does not necessarily mean a fit with Sample - the market can change. To be sure of the fit, it is necessary to conduct tests on OOS before Sample and after Sample. If the result for both of these forwards is negative, then we already deal with the bare fittings.
Successful forward tests do not guarantee the profitability of the TS in the future. Their purpose is to identify the fit.
In order to make sure that the TC is not bluntly adjusted for the optimization period, we conduct an OOS test.
All is correct. But. OOS must contain information that differs from that which is available in the optimization period, for us to be sure that TS is capable of generalization (some regularity is found and results will be similar on different data but containing this regularity). Otherwise, the test on OOS will have the same results as on sample, but it will not confirm that the TS learned a pattern, and in the future, the TS will not be able to work with the same results.
Now, pay attention! Question: Do you check if OOS contains data different from S? If so, how?
If anyone finds inaccuracies in the diagram itself, please let me know...
The dimensions are true, but it's still a blurry image. In the separate window it's clear.
You say that you are preparing data for training. Could you be more specific, how long have you been using such techniques? Something in your words is very familiar, I remember that I suggested, like in the branch about context, to prepare synthetic data with preliminarily required parameters for optimization, so you can change data parameters and see the response of the TS. I think in de like you were just agreeing with me, but suggesting a slightly different option to mine - prepare data from real pieces of history, is that right?
1) in russkomazykenetlova osogodarstvennyy.go.... )
For the particularly gifted:
see. Dictionaries and encyclopaedias on Academics: Specially Gifted
P.S. For especially gifted connoisseurs of the Russian language: there are spaces in the Russian language to separate some words from others.
Wai, wai, wai...
How complicated you are...:))
lasso, your best bet is to read something not too complicated on nerve grids. I think that's where all this terminology came from. Perhaps I'm inaccurate in terms, please excuse me, it's been a long time since I read it:
1. sample data: the training plot. This is the section from which we directly take the data and teach the network.
2. Verification data: verification section. We do not train at this part, but use it for error estimation and control when we should stop training. There is a well-known curve of verification error depending on the number of runs. This is a curve with a minimum. If we train too long, i.e. if we don't stop in time, the error in the training section will still decrease, but the error in the verification section will now increase. This is the fitting: we approximated the data in the training section quite well, but overdid it, as the verification error started growing. It is the second error that is the evaluation of the quality of training and ability of a neural network to generalize.
3. testing data. This is a true OOS, out of sample.
The second plot, the verification plot where the error is assessed, is not OOS, although we don't train on it. Nevertheless, the data from this section is used to train the data from the first. In order to properly and completely independently verify the quality of training (more precisely, generalization), we need to take data that we have not yet seen or used in training.
Here in the tester we do not have neural networks. An error is estimated directly at the Sample data section. So there is no way to directly transfer nerve-methods here. Although, maybe xeon invented something here too with its TestCommander...
For the particularly gifted:
see. Dictionaries and Encyclopaedias at the Academic: Specially Gifted
P.S. For especially gifted Russian language experts: there are gaps in the Russian language
Figar0:
You say that you are preparing data for training. Could you be more specific, how long have you been using such techniques? Something in your words is very familiar, I remember that I suggested, like in the branch about context, to prepare synthetic data with preliminarily required parameters for optimization, so you can change data parameters and see the response of the TS. I think in de like you were just agreeing with me, but suggesting a slightly different option to mine - prepare data from real pieces of history, is that right?
preparing the data for training on some rule is just introducing an extra filter into the system.
For those who are especially gifted: non-stationarity is the absence of statistical regularities, such as expected payoff and variance.
Put Bollinger envelopes on the chart and you will see what the "patterns" of nonstationarity are, because the centre of the indicator is expectation, and the distance from the centre to the envelopes is dispersion.