Machine learning in trading: theory, models, practice and algo-trading - page 3682
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
Purely theoretical question. I am thinking about the application of ensemble ideas from MO to a set of TCs.
The simplest, perhaps, is bagging - just a portfolio of independent TSs. The total position is defined as the sum or average of the positions of individual TSs.
Staking - a system of several interacting TSs, where the work of some TSs somehow depends on the work of others. The idea is more or less clear, although it is difficult to describe it unambiguously - there can be a huge number of variants.
It is not clear with bousting. Perhaps something like a sequence of TSs, each of which improves the performance of a portfolio of previous TSs (reduces drawdown, etc.).
https://www.mql5.com/ru/articles/4227/116679#!tab=article
https://www.mql5.com/ru/articles/4228/116680#!tab=article
https://www.mql5.com/ru/articles/4722/116681#!tab=article
There may be useful information here.
Thanks, I'll have a read.
The links only work after cutting them:
https://www.mql5.com/ru/articles/4227
https://www.mql5.com/ru/articles/4228
https://www.mql5.com/ru/articles/4722
When are we going to register a hedge fund in Dubai?
Imho, we should start with a trading platform for trading. Then add a trading forum and supplement it with the creation of a thematic cryptocurrency.
If it suddenly takes off, then adding a hedge fund will be optional, and if you start with it, you can immediately run into bad news (which is obviously coming) and spend the rest of your life hiding from investors).
Vladimir Perervenko #:
https://www.mql5.com/en/articles/4227
I had a quick look at your articles: as always, the highest professionalism.
But only cursorily, because I was trying to see an important point for myself.
I myself have tried different variants of unification, not at such a level as yours, nevertheless. Apart from combining predictions, I have tried a lot of other things.
If I take a ready file and divide it into chunks, I got just wonderful results: classification error less than 10% and accuracy over 90% - no problem.
BUT.
If you run it as in real life - step by step, all my exercises have the same dismal result: classification error about 50% and profit factor about 1. You can practice and achieve 40%.
Therefore, the evaluation of any trading systems ONLY in the tester, you can write a kind of tester in R, but necessarily - a new bar - predict and this new bar in no way included in the previous exercises on fitting and evaluation of models. And so every step.
If you train the whole steamer on the arrival of a new bar, you can get a profit factor of about 1.5 (classification error below 40%).
As of today I don't see any refutation: ALL MO is designed ONLY for analysis and prediction is impossible.
I would be grateful if you could refute this statement by posting the balance obtained by a step-by-step run.
Imho, we should start with a trading platform for trading. Then add a trading forum and supplement it with the creation of a thematic cryptocurrency.
If it suddenly takes off, adding a hedge fund will be optional, and if you start with it, you can immediately run into bad news (which is obviously coming) and hide from investors for the rest of your life).
If the news is too lousy, there will be no one to hide from :)
Purely theoretical question. I am thinking about the application of ensemble ideas from MO to a set of TCs.
The simplest, probably, is bagging - just a portfolio of independent TSs. The total position is defined as the sum or average of the positions of individual TSs.
Staking - a system of several interacting TSs, where the work of some TSs somehow depends on the work of others. The idea is more or less clear, although it is difficult to describe it unambiguously - there can be a huge number of variants.
It is not clear with bousting. Perhaps something like a sequence of TSs, each of which improves the performance of a portfolio of previous TSs (reduces drawdown, etc.).
Response from Dipsic:
Applying bousting to an ensemble of trading systems (TSs) requires an adaptation of its key principles - sequential improvement with a focus on error correction of previous models. Here are the main ideas and approaches:
1. basic principles of bousting in the context of a TS
Consistent correction: Each new TS focuses on the "weaknesses" of the ensemble (e.g., losing trades, drawdowns, excessive risk).
Dynamic weighting: The contribution of each TS to the overall position depends on its current performance (analogue of weights in gradient bousting).
Adaptation to residual errors: Subsequent TSs are trained on data where previous systems have shown poor performance.
2. Specific implementations
(a) Sag correction
Example:
TS-1 generates basic signals (e.g. trending strategy).
TS-2 analyses the drawdown periods of TS-1 and activates hedging positions (e.g. counter-trend trades) at times of high volatility.
TS-3 optimises position size by reducing risk in areas where TS-1 and TS-2 conflict.
(b) Meta risk management
Each subsequent TS introduces additional rules:
Dynamic stop loss at the ensemble level.
Adjustment of lot size depending on "confidence" in previous TSs (e.g., reducing position when signals diverge).
c) Gradient bousting for trading strategies
Formalisation of the problem as minimisation of the portfolio "loss function":
Goal: Each new TS reduces the cumulative drawdown or increases the Sharpe Ratio.
Method: Training the TS on historical data with a focus on periods where the current ensemble performed worse (analogue to residuals in a regression).
3. Example Architecture
Basic level:
Ensemble of independent TCs (bagging).
Common position = weighted average of signals.
Boosting layer:
TC-A: Analyses losing trades of the base ensemble and generates opposite signals to partially compensate for them.
TC-B: Optimises capital allocation between TCs by increasing the weighting of systems with better近期 performance (sliding window).
Meta-strategy:
Uses an ML model (e.g., gradient bousting) to predict the optimal combination of signals of all TCs based on current market conditions.
4. Advantages and Risks
Benefits:
Reduction of drawdowns by compensating for the weaknesses of individual TSs.
Adaptability to changing market conditions.
Risks:
Retraining on historical data (especially when weights are frequently rebalanced).
Increased complexity of infrastructure and higher transaction costs.
Possible signal interference (e.g., conflict between trend and counter-trend strategies).
5. Practical steps for implementation
Define a "loss function" for the ensemble (e.g., maximum drawdown, volatility).
Implement a mechanism to sequentially add TCs, each of which minimises this function.
Introduce dynamic weighting of signals (e.g., based on moving Sharpe Ratio).
Test on a long history taking into account different market regimes (trend, flat, shocks).
Key Thought
Busting in trading is not just a "sequence of TSs" but a feedback system where each new strategy purposefully improves the ensemble metrics by focusing on its current weaknesses. This requires deep integration between components and rigorous control over re-learning.
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?