Machine learning in trading: theory, models, practice and algo-trading - page 1161
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
Well done, Maximka! Finally started thinking about predictors, not about the model, and about adaptive predictors, for which I give you double kudos.
And yes, I would listen toIgor Makanu and did the test.
About a week ago I did something very similar, only using "ssa" spectral analysis method, the same cycles, same cutoffs, it worked by eye, but then I looked at the history and .... :(...
By the way it is interesting to hear how the network finds cycles , how does the algorithm work in general ?
i was remaking the Fourier extropolator from KB to work online in the tester, so it wouldn't fudge the old predictions values - you can see it's foolproof, that's why i suggested to try it in the tester
SZS: here I've found my remake, unfortunately I try to do it under MT4 for 20-30 minutes, but under MT5 it spends a couple of hours ((((
"Profitability" is a bullshit goal. Make up another one and you'll be fine.
I don't think it makes any sense to make a forecast for the rest of the time.
I have been searching for patterns on history, all I can say is "First sacred cow: " If the trend started, it will continue" - this is all that can be found with the help of indicators-"predictors" (I've even searched on Renko charts and took apart the SSA method by bones - the result is the same), if@Maxim Dmitrievsky will search for the Grail further, but with a tester, it should turn out that approximately 55 predictions out of 100 will be correct, and 45 will not be correct - that's how I did it.
SZY: as a variant of search direction - what else I have not checked and maybe@Wizard2018 wanted to say: you should look for a forecast at certain moments (market conditions), ie, you need a filter, maybe even the classic indicators will do for these purposes: if RSI>70 or RSI<30 then make a forecast (or the opposite if RSI<70 && RSI>30), but at other times, there is probably no sense to predict
You can input sets of correlations of moving averages and/or prices, so that the network would find a connection between these patterns and profitability of trades during training and give an indicator on the output, by which they could then be sorted.
I have muwings on the input.
I get signals based on these moving averages. The trades appear.
Now you are suggesting to use moving averages and deals.
No, that's not it. I tried it.
Here's an example. There are about 5 sells. What algorithm can discard 4 of them and keep 1 sell?
Classification or something?
I want to feed exactly the deals to the input of the algorithm.
What algorithm can handle this task?
I have muwings on my input.
I get signals based on those muwings. I get trades.
Now you propose to take the muwings and deals again.
No, that's not it. I tried it.
Here's an example. There are about 5 sells. What algorithm can discard 4 of them and keep 1 sell?
Classification or something?
I want to feed exactly the deals to the input of the algorithm.
What algorithm is able to do it?
I do not know what algorithm you use to calculate signals, there are many variations with muwings, but I think you can always extend the pattern.
For example increase the number of bars, initial price, type, etc., on which the neural network can find patterns of erroneous signals to try to sort out trades.
I have been searching for patterns on history, all I can say is "First sacred cow: " If the trend started, it will continue" - it's all that can be found with the help of indicators-"predictors" (I've even searched on Renko charts and took apart the SSA method by bones - the result is the same), if@Maxim Dmitrievsky will search the Grail further, but with a tester, it should turn out about 55 predictions out of 100 will be correct, and 45 will be wrong - I had it so.
SZY: as a variant of search direction, what else I have not checked and maybe@Wizard2018 wanted to say: you should look for a forecast at certain moments (market conditions), ie, you need a filter, maybe even the classic indicators will do for these purposes: if RSI>70 or RSI<30 then make a forecast (or the opposite if RSI<70 && RSI>30), but at other times, there is probably no sense to predict
I agree, you can make predictions about trends, levels, prices etc., but all these things are intermediate goals.
It's like in deep learning, by layers of neural network the level of abstraction changes, and each of them may be considered as an intermediate goal, but the final goal on the output layer remains the same, in our case it's profitability.
"Profitability" is a bullshit goal. Make up another one and you'll be fine. (No thanks.)
Profitability is really the only goal. But designing or teaching a TS based on profit maximization is indeed a bullshit goal. I may have already outlined this approach in more detail in this thread.
.... even searched and took the SSA method to pieces ......
Have you ever tried it?
We take one of the main components, the second or the third one, if it has a pronounced periodicity.
(Everything that is after the vertical black line in the picture is the future that is not known to us, we, like, cannot see prices, and all lines drawn in the picture are predictions).
Like there is no periodicity, then we take another component faster (high frequency) 8 ... 15 in order
We connect them and get another series
The second "fast" component needs to be matched so that the major extremums of the purple band more or less coincide with the extremums of the slow (blue) component, and something like heads and shoulders appears at the extremums
After we have selected everything correctly we obtain an interesting feature of those "heads with shoulders" - price almost always breaks away from one of the three shoulders (purple line) if that figure is formed under a large reversal (extremum of the blue component)
So it turns out that the market may be cyclical, it is just necessary to learn to make accurate forecasts with faster components. For example, the main component shows that the trend is changing from descending to ascending, but some faster component has not yet completed its cycle downwards and we obtain an allegedly wrong forecast of the trend upwards
I did it all by SSA method, but I think any spectral method will do: PCA, Fourier, wavelets
Have you ever tried this?
We should take one of the main components, the second or the third one, if it has a pronounced periodicity.
(Everything that is after the vertical black line in the chart is the future that is not known to us, we, like, cannot see prices, and all lines that are plotted are predictions).
Like there is no periodicity, then we take another component faster (high frequency) 8 ... 15 in order
We connect them and get another series
The second "fast" component needs to be matched so that the major extremums of the purple band more or less coincide with the extremums of the slow (blue) component, and something like heads and shoulders appears at the extremums
After we have selected everything correctly we obtain an interesting feature of those "heads with shoulders" - price almost always breaks away from one of the three shoulders (purple line) if that figure is formed under a large reversal (extremum of the blue component)
So it turns out that the market may be cyclical, it is just necessary to learn to make accurate forecasts with faster components. For example, the main component shows that the trend is changing from descending to ascending, but some faster component has not yet completed its cycle downwards and we obtain an allegedly wrong forecast of the trend upwards
I did it with SSA method but I think I can use any spectral PCA, Fourier, wavelets
Why do I need so many charts, if it is for conviction it is more important to publish at least one, but for a particular trading instrument with a forecast for the real future.)
I wrote, especially for those who are in the tank.
(Everything that's after the vertical black line is in the future, we don't know what's ahead of us, we don't see the prices, and all lines are forecasts)