Machine learning in trading: theory, models, practice and algo-trading - page 3473
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 explain it by the fact that on small TFs small average rate of profit per each trade - a few pips. And NS averages (approximates) everything, i.e. trades do not cover spread/commission/slippage after training. You can try to calibrate the model and use only high probabilities.
But there is a better way - train NS on hours, take signals from them, and trade on small TFs, add trades artificially on each bar, average positions. There will be a lot of trades too.Explanation
For watchmakers, the quantile of increments is 20,000 bars,
Half of it is less than 5 pips, and only 20% of the increments are of some interest independent of the usual spread and commission.
In explanation
For watchmakers, the quantile of increments is 20,000 bar,
Half are less than 5 pips, and only 20% of the increments are of some interest independent of the normal spread and commission.
This is if you mark up the increments, but nobody in their right mind does that, right? Usually the forecasting horizon is further than one increment.
For example, there on a zigzag, or I just put an arbitrary number of bars forward.That's if you use increments, but nobody does that in their right mind, do they? Usually the forecast horizon is further than one increment.
For example, there on a zigzag, or I just put any number of bars forward.Doesn't matter: profits are increments.
For example, if we use ZZs for training, there are very gentle and long ZZs. In the end it all comes down to what we teach. If we teach profits above 10 pips, we can ignore the spread.
It doesn't matter: profits are increments.
For example, if we use ZZs for teaching, there are very gentle and long ZZs. In the end, it all comes down to what we teach. If we teach profits above 10 pips, we can neglect the spread.
Then it stops generalising well ) That's what the story is about.
And again, the old song about the main thing. I continued testing the method of quantisation through ZZ, I used the already worked out scheme:
1. Decomposition of predictors into quantum segments on the sample train.
2. Evaluation of each quantum segment on the train sample for selection into the pool. 3.
3. Removing/filtering the part of the sample described by a quantum segment with a target "0" bias according to a given criterion. This time looked to see if there was any effect of the quantum cutoff on the test sample for confirmation.
I made 100 iterations.
This is the result for the samples train, test, exam. The indicator is the probability of occurrence of the target unit (percentage of units in the sample).
The positive dynamics on the independent sample was observed up to 79 iterations, and then a couple of unsuccessful quantum segments reset the model to the starting point.
At 90 iteration, the drop was 0.6 points, which is the maximum. The chance of picking a good quantum cutoff was 33% out of 1329 choices.
As I said earlier, until a meaningful criterion for selecting quantum segments is found (in a simple tree estimation - split), we cannot talk about the stability of the trained models. And this is despite the fact that I already use a number of metrics to assess stability, but apparently they are not enough.
And so, of course, it is possible to randomise the choice of quantum segments, and there will certainly appear a variant of the model with good performance, which during construction accidentally collected positive quantum segments.
And, of course, all kinds of methods of model selection by delayed sampling or even stopping training may not work as expected, which is confirmed by the above shown graphs.
In continuation of the post.
I was wondering, what are the chances of the model to select the right quantum segments, so that they would show a stable result on three samples.
Quantum segments are selected for each target separately, although my ingenious algorithm used only negative target "0" to build the model, but the statistics for target "1" is also available - the red curve on the graph. The graph below shows the number of selected quantum segments for two targets at each iteration.
The shape of the graph shows that there is no linear dependence on the iteration number, but there seems to be a correlation between the number of targets at each iteration. If the graph for the target "1" is in the conditional range, then for the target "0" we see some explosive growth on 2/3 of the graph with a further downward trend.
Below is the graph showing the percentage of stable quantum segments for each target.
It looks like an inverse correlation - the reason is not clear. Also noteworthy is the strong drop in the chance of choosing a good quantum segment for target "0" up to about 40 iterations - at the bottom the chance is in the range of 5%, which is very small, at 2/3 there is a return to the original probability. This spread is surprising, as is the pronounced dynamics over many iterations.
Looks like an inverse correlation - not clear why.
If you have a buy and a sell, when there is a global upward price movement, the buys start winning more often than the sells.
If you have a buy and a sell, when there is a global upward price movement, the buys start winning more often than the sells.
Target "0" means do not trade, and "1" means trade. The direction is determined by the independent variable.
Even on a weekend labour? )
there's a Fallout TV series based on the iconic game.
Even on a weekend labour? )
there's a Fallout TV series based on the iconic game.