Machine learning in trading: theory, models, practice and algo-trading - page 3527

New comment
 
Aleksey Vyazmikin #:

All I fancy is my method - where predictor ranges are essentially zeroed out if a probability bias is detected for the null class, thus avoiding inconsistencies.

What is the point of your method?

If in your terminology, 0 and 1 are spread across different predictor ranges.

 
Maxim Dmitrievsky #:

If in your terminology, 0 and 1 are spread across different predictor ranges.

Did I understand correctly that conditional cleaning is happening - replacing 1 with 0 after randomisation?

Or sort of picking predictors to do this?

But, how do you relate this to the stability of the probability bias, because you can do anything on training, but what happens afterwards?

 
Aleksey Vyazmikin #:

Did I understand correctly that conditional cleanup is happening - replacing 1 with 0 after random markup?

Or sort of picking predictors for that?

But, how do you relate this to the stability of the probability bias, because you can do anything on training, but what happens afterwards?

I'm removing examples that are on the conditional class separation boundary. It's all according to the canons of MoD. The improvement of stability is due to simpler boundaries.

Gradually reduce the number of attributes still, there are models that work well on 2 mashas. So it is possible to reduce the number of features to 2.

Bousting is trained at most 100 iterations later. It must be something from Active Learning.
 
Aleksey Nikolayev #:

Obviously a price reversal timetable)

Too bad it turned out not to exist)

First you have to answer the question of what a price reversal is.

and then look for the timetable.

 
Maxim Kuznetsov #:

First, we need to answer the question of what a price reversal is.

and then look for a timetable.

At least two subsequent increments (before and after the reversal) should have different signs. This leads to a negative correlation between these increments. Conversely, where there are no reversals, the subsequent increments have the same sign, which leads to a positive correlation.
 
Aleksey Nikolayev #:
Naturally, we failed to see any "turnaround according to the schedule".

But what can be seen on the charts that can be taken as a "schedule"?

Obviously, these are effects related to daily volatility fluctuations.

Suppose we have two realisations of the SB, which differ only in the dispersion of increments. No one in their right mind would say that one is more predictable than the other. In fact, they differ only in that the more volatile realisation makes time seem to speed up. BUT, if these realisations are sliced into pieces, mixed and combined into one, the eye can start to see some patterns. For example, it is obvious that in a faster implementation the zigzag tops will appear more often, which of course does not make them more predictable.

Conclusion: you can start writing a fanfic based on Taleb's "Fooled by Randomness")

 
Aleksey Nikolayev #:

But what is it then that is visible in the graphs that can be taken as a "schedule"?

Obviously, these are effects related to daily volatility fluctuations.

Suppose we have two realisations of the SB, which differ only in the variance of increments. No one in their right mind would say that one is more predictable than the other. In fact, they differ only in that the more volatile realisation makes time seem to speed up. BUT, if these realisations are sliced into pieces, mixed and combined into one, the eye can start to see some patterns. For example, it is obvious that in a faster realisation the zigzag tops will appear more often, which of course does not make them more predictable.

Conclusion: you can start writing a fanfic based on Taleb's "Fooled by Randomness")

Well, that's if you take SB. And you can chop it up into chunks and then do tests to see if it's different from SB.
Somewhere I have seen or invented (I don't remember) that a chart in different periods, for example trading sessions, are different markets, which are piecewise combined into one. You can find justifications and refutations for this, but it makes sense to try to analyse them separately. My cluster algos confirm rather than refute.
Inefficiencies, aka they may not yet be determined by tests. That is, TC development basically consists of "improve a little there, tweak a little here", and so on round and round :)
 
Maxim Dmitrievsky #:
Well, that's if you take the SB. But it is possible to cut into pieces and then make tests for differences from the SB.
I saw somewhere, or I invented it myself (I don't remember now), that a chart in different periods, for example, trading sessions, are different markets, which are piecewise combined into one. You can find justifications and refutations for this, but it makes sense to try to analyse them separately. My cluster algos confirm rather than refute.

It also seems obvious that there is some kind of temporal structure related to the sessionality of the market. Another thing is that it can hardly be described by simple methods like "scheduled reversals".

Imho, there should be some dependence on the size of price movement. Large movements (from the daily range and above) are hardly determined by daily rhythms.

 

Here's a squeeze on cluster analysis from gpt-omni

**K-Means Clustering-Based Trading Strategies**

**Principle:**

* K-Means clustering is an unsupervised machine learning algorithm that groups data points into a specified number (k) of clusters based on their similarity.
* In trading, it can be used to identify different market states or patterns.

**Strategies:**

**1. Cluster-Based Market Segmentation:**

* Cluster historical market data (e.g., prices, volumes, indicators) into k clusters.
* Each cluster represents a distinct market state, such as uptrends, downtrends, or consolidation.
* Trade according to the prevailing market state identified by the cluster.

**2. Cluster-Based Trend Identification:**

* Group price data into k clusters based on their underlying trend.
* Trade in the direction of the strongest trend identified by the largest cluster.
* This strategy aims to capture momentum and avoid trading against the trend.

**3. Cluster-Based Price Breakout Detection:**

* Cluster price data into k clusters based on their volatility.
* Identify clusters with high volatility, which may indicate potential price breakouts.
* Trade in the direction of the breakout once it occurs.

**4. Cluster-Based Support and Resistance Identification:**

* Cluster price data into k clusters based on their distribution.
* Identify clusters with high concentration, which may indicate support or resistance levels.
* Trade by buying at support and selling at resistance.

**5. Cluster-Based Pair Trading:**

* Cluster two or more instruments into k clusters based on their correlation.
* Trade long on the instrument in the cluster with the highest expected return and short on the instrument in the cluster with the lowest expected return.
* This strategy aims to exploit relative value opportunities.

**Implementation:**

* Determine the appropriate number of clusters (k) based on the data and desired specificity.
* Choose suitable data features for clustering, such as prices, volumes, or technical indicators.
* Train the K-Means model on historical data.
* Use the model to classify new data points into clusters.
* Define trading rules based on the identified clusters.

**Advantages:**

* Can identify hidden patterns and market states.
* Reduces subjectivity in trading decisions.
* Can improve risk management by clustering similar market conditions.

**Limitations:**

* Clustering results can be sensitive to the choice of parameters and data features.
* May not perform well in rapidly changing market conditions.
* Requires historical data for model training.
 
Aleksey Nikolayev #:

It also seems obvious that there is some time structure related to market sessionality. Another thing is that it is hardly described by simple methods like "scheduled reversals".

Imho, there should be some dependence on the size of price movement. Large movements (from the daily range and above) are hardly determined by daily rhythms.

Well, yes, situationally.

Imho, the development of TS is more like shamanism than some scientific work :)

1...352035213522352335243525352635273528352935303531353235333534...3574
New comment