Machine learning in trading: theory, models, practice and algo-trading - page 3683
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
You can memorise history almost perfectly with a single tree. Not a word about success on new data.
I have the same situation as SanSanych.
You don't need to memorise perfectly, you need to find regularities and generalisation. But how to get generalisation and not just underlearning?
It's not the first time I've tried to raise this question: everything is fine at the learning stage and random result at the step-by-step run-through.
And this suggests that there is no point of any tricks at the learning stage - something primitive is enough.
And if you go deeper into it, the random result on a step-by-step run means that the patterns that the MO has found on the history do NOT exist in the future. And discuss on the stationarity of the patterns.
You can memorise history almost perfectly with a single tree. Not a word about success on new data.
I have the same situation as SanSanych.
You don't need to memorise perfectly, you need to find regularities and generalisation. But how to get generalisation and not just under-learning?
You can memorise history almost perfectly with a single tree. Not a word about success on new data.
I have the same situation as SanSanych.
You don't need to memorise perfectly, you need to find regularities and generalisation. But how to get generalisation and not just underlearning?
1) Each pattern tends to be unique, if we compare prices among themselves. With each new price in the comparison window (the first price with the second, the first with the third, the second with the third, etc.), the "uniqueness" increases many times.
The same is true when formatting the price space into simplified patterns: with each new comparison, the "uniqueness" increases, and a figure with a couple of dozens of reference points compared will meet in the future in 120 years.
2) Chess has a conditionally infinite number of patterns. But the machine learns to play infinitely correctly and infinitely strongly.
3D-space has an infinite number of states of the agent, starting from the initial point, ending with the ratio of the angles of bending of the limbs of its dummy, multiplied by each pixel-voxel, etc. But the machine learns to walk, play, and win without learning all these states.
Is it possible to somehow apply methods "from there" to let the agent into the graph.
The agent will see constantly an infinite picture of patterns "a lot of input data (in fact - a bare graph), but somehow learn to "win" in this environment.
And this suggests that there is no point in any tricks at the learning stage - something primitive will suffice.
It happens that 1 tree divided into 2-3 splits shows better results than deeper trees/forests.
But this is already too primitive. You want something more accurate. And with such a rough model there will be a lot of unsuccessful trades, which lowers the profit factor almost to 1 and lower.
It is the same when formatting the price space into simplified models: with each new comparison the "uniqueness" increases and the figure, which has a couple of dozens of reference points compared, will meet in the future in 120 years.
That's probably why 3 splits show good results. Leaving the model very rough and not memorising unique features
The problem is both with the regularities (whether they exist, and how strong they are) and with the methods for finding them.
Well, in general, ensemble in MO makes sense if there are weak algorithms that at least a little, but still work.
Accordingly, ensembled TSs should give at least a small profit. It is generally accepted to collect such models into portfolios - in essence, bagging.
The (theoretical) question was, in fact, whether it is possible to invent some more complicated procedure like bousting instead of portfolioisation of these weak TSs.
But a suitable platform for MO trading would be nice. With custom events and normal synchronisation of events on history.
during meditation, sometimes even cognac.
Which one do you prefer?
Well, generally speaking, ensembling in MO makes sense if there are weak algorithms that are at least somewhat, but still work.
Accordingly, ensembled TSs should give at least a small profit. It is generally accepted to collect such models into portfolios - in essence, bagging.
The (theoretical) question was, in fact, whether it is possible to invent some more complicated procedure like bousting instead of portfolioisation of these weak TSs.
Maybe yes. I build a TS on the basis of an approach to trading, but not a platform. Then I do the tying in the form of a connector or transfer to an existing platform. In general, this is the smallest part of the work. The hardest part is to come up with the approach itself, unfortunately :) a new approach is invented only in the process of meditation, sometimes even cognac meditation.
Yes, sniffing quality cognac from a snifter warmed by the palm of your hand :).
It's just sometimes annoying not to have a standard feature for portfolio OnTick() or to have to shamanise to synchronise news with prices when testing.
There are also times when 1 tree split into 2-3 splits performs better than deeper trees/forests.
But this is already too primitive. I want something more accurate. And with such a rough model there will be a lot of failed trades, which lowers the profit factor to almost 1 and below.
I guess that's why 3 splits show good results. Leaving the model very coarse and not memorising the unique features
I observe with simple constructions: rough profit on the forward not only on the trained one, but also on other pairs. You can put it in a portfolio.
And vice versa: a good forward of a "heavy" model is absolutely random on all other pairs.