Machine learning in trading: theory, models, practice and algo-trading - page 2259
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
Maxim, have you tried the Neural Turing machine? In which framework and what were your successes?
Happy New Year and all your wishes!
Hi, Happy New Year. No. I'm more interested in generative models now, they are closer to Turing, if you cannot distinguish the artificial series from the real one. Actually, the right solution to apply MO to the market has already been found, there are still nuances. You just need to model the concept drift correctly and educate yourself
...the right solution for applying MO to the market has already been found... And what is that solution? I assume there are a number of competing solutions)
And what about modeling concept drift? Doesn't the feedback help?
And conceptually what is more assuming:
- Gradual change over time.
- Periodic or cyclical change
- Sudden or abrupt change
Or do we include everything at once?
...the right solution to the application of MO to the market has already been found... And what is that solution? I assume there are a number of competing solutions)
And what about modeling the concept drift? Doesn't the feedback help?
And conceptually what is more assuming:
- Gradual change over time.
- Periodic or cyclical change
- Sudden or abrupt change
or do we include everything at once?
You need to look at what exactly is changing and what the axis is designed to do. Model what changes, i.e. create artificial series. Look at the range of changes, the history. There is no single solution, but according to a situation it is possible to make it work for a long time. Inverse relations for modeling norms, for example recurrence gan's, but I haven't gotten to them yet. And the classifier for the model itself can be anything
Usually it all comes down to fairly trivial things like the average increment bias and variance, which need to be changed. And volatility clustering is perfectly modeledIt is necessary to look at what exactly is changing and what the TS is designed to do. Model what changes, i.e. create artificial series. To see in what range of changes, on the history. There is no single solution, but according to a situation it is possible to make it work for a long time. Inverse relations for modeling norms, for example recurrence gan's, but I haven't gotten to them yet. And the classifier for the model itself can be anything
Usually everything rests on fairly trivial things like average incremental shift, which needs to be changed. And volatility clustering is perfectly modeledWhat if the shift in mean (or maybe median) increments, assuming it's "Gradual change over time"/"Periodic or cyclic change" to introduce into the model as a control variable? Based on the concept of LifeLong Learning.
It is probably more complicated with a sudden or abrupt change, though it can be exactly the opposite).
How about a shift in the mean (or maybe median) of the increments, treating it as a "Gradual change over time"/"Periodic or cyclic change" to introduce into the model as a control variable? Based on the concept of LifeLong Learning.
With sudden or abrupt change, it's probably more difficult, although it may be exactly the opposite)
I'm not familiar with such concepts. I think it's enough to break the row into batches of n-bars each and you can shuffle if you want suddenness. I don't believe you can specifically identify anything, but through enumeration of variants to get a normal model is not a problem. Which did not see the new data, but was trained on something similar, generated. The main thing is the coverage of variants to be large, otherwise it is possible to randomly select for the history
For example, I obtain good models for all currency pairs with a horizon of 5 years by training for only a couple of months + artificial ones. I don't know what the global change is, but if you look at the seasonal patterns, the shift of the mean is different. Haven't modeled it yet.
For example, on all currency pairs I get good models with a horizon of 5 years, learning in just a couple of months + artificial. What is the global change there, I do not know, but if you look at the seasonal - the shift of the average is different. I haven't modeled it yet.
On stock/commodities get models of the same horizon? Is "a couple of months" a part of BP's history? If it is so, it's a Klondike!
Do stock/commodities get models of the same horizon? Is "a couple of months" a section of BP's history? If so, it's a Klondike!
Everything there is situational, somewhere it's 2 months, somewhere the market has changed a lot and this history is not enough. I don't know, but I will try other instruments. I haven't tried other instruments, I may try to use indices.
I've tried other tools for indexes, you can try them. It's just the approach itself - we need a lot of plausible examples, it works everywhere, not only for time series. There's no super science, just poke around and watch )
For example, teach on specific clocks (seasonal components), made a brutforce like this. Selection of models by hours. Each point is one model, 10 models trained for each hour. The denser and higher the points, the better
You can see on the chart that there are a lot of good models at the edges of the trading day. The strategy works worse (on average) in the middle part of the day when volatility is high. There are only a few outright garbage periods, the rest can be worked with.
Then I look for the 5th hour and get such a balance curve. All the patterns turn out good for him. Half test, half track (for 5 years). For the seasonal ones I need more than 2 months, because there are not enough examples.
And everything in this vein. I wanted to write an article, but it's too short in words.
It's GBPUSD, but it works on all currency pairs
Selection of models by hours. Each point is one model, with 10 models trained for each hour. The denser and higher the points, the better.
Then I look at the 5th hour and get the balance curve. All the models are good for him. Half test, half track (for 5 years). For the seasonal need more than 2 months, because there are few examples
And everything in this vein. I wanted to write an article, but it's too short in words.
This is GBPUSD, but it works on all currency pairs
How they ignored the obvious for so long...