Machine learning in trading: theory, models, practice and algo-trading - page 82
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
When training the model, the more trees the more the "0" class learns and as the "0" class learns better and stronger it begins to sort of (absorb - squeeze out) the "1" , "-1" classes That's why the more trees the less deals
Something is wrong, the forest is almost always trained to 100% accuracy with any skewness. On validation with such data there may be problems, but predicting results on training data the forest will always be accurate. I can assume that you are not using classification but regression, and when you test the model you get not crisp classes -1;0;1, but valid numbers in the range of -1 to 1. With that kind of skewed regression, no good will come of it.
A better way to do it is to say that you have a trainingData table, where the last column is the target variable and all the other columns are predictors:
not really...
What I'm writing is only applicable to my approach.
you know how i do my target, it is reversals
I have three classes of reversal "up", "down" and "no reversal" ( 1 , -1 , 0)
Also you know that the skew of the classes is enormous, the class "0" is dozens of times more than "-1" and "1"
and this means that the model is best trained on the "0" class because it has the most observations, and when you train the model, the more trees the more "0" class is trained and as the "0" class gets better and stronger it starts to "absorb" and "-1" classes. That's why the more trees the less deals
For me you write amazing things!
Very original approach for unbalanced classes.
Then my understanding is that your list of predictors not only contains no noise, but also has a very high predictive power.
If so, how did you achieve this?
Something is wrong, the forest is almost always trained at 100% accuracy with any skew. On validation with such data there may be problems, but predicting results on training data the forest will always be accurate. I can assume that you are not using classification but regression, and when you test the model you get not crisp classes -1;0;1, but valid numbers in the range -1 to 1. With a regression skewed like that, no good will come of it.
Better to do so, let's say you have a trainingData table where the last column is the target variable and all other columns are predictors:
100% for a tree is absolute nonsense!
If all your predictors are noise, it is very difficult to achieve such a result: there will still be a 3%-5% error. Noise always gives very good results with all the crossvalidation and other tricks.
Accuracy of 100% means the only thing: among predictors you have a duplicate of the target variable (some modification of it). That is, the model looks into the future.
The accuracy of 100% means the only thing: among the predictors you have a double of the target variable (some modification of it). That is, the model looks into the future.
What don't you like about this approach? How are you going to select the parameters?
I think the approach is great, it's hard to think of something better, the problem is in the market itself...
Remember, I told you about my experiments, correlations and search for patterns in history (not SSA, but before), what I did:
I took the current situation and looked for its analogues in the past and looked how they ended up, if I found a situation "X", when I found 20 analogues of which 17 ended with a fall and 3 with a rise, so the statistical advantage is obvious(we should sell),by the way, why is not crossvalidation? just by a single pattern, do you agree? We found out that this pattern will not work, it turned out that with a huge percentage of probability the market will go up, it turns out the market goes against its own statistics with a huge statistical probability.
Ifto make it simpler, if yesterday there was event "x" and after it there was a fall and the day before yesterday there was event"x" and after it everything fell down too and after yesterday there was a fall after event"x"- then if today will come"x"- everything will grow, what crossvalidation will help here?Nothing, never.
Let me explain the most important thing, the understanding of the process.
The market is a cruel business, where some people who have built this business legally take money from others, the crowd, the majority are always obliged to lose, everyone knows about it, everyone writes about it, no one hides it, in short, it is an axiom - a pattern of 95% of traders lose money - the market moves against the majority with the probability that traders lose money with that 95%.
And what does the crowd use when trading? Well, the crowd basically has nothing besides one thing that makes them make deals.
Any actions starting from visual observation of charts and searching for patterns in the breakdown and finishing with neural network training is nothing else but trading by statistics, the very statistics that doesn't work in the market, do you understand what I'm talking about?
the market moves against the crowd's trades ----- the crowd acts on statistics ------ all you need is to predict the crowd's action in the future and do the opposite, the only way to predict is statistics
)))) start with one and end with the third :) Well, that's how it is, at least I got it out of my system.)
p.s. all that i said here is my opinion only, i don't impose anything, i can argue and prove it, but i'm not in the mood for it
Something is wrong, the forest is almost always trained at 100% accuracy with any skew. On validation with such data there may be problems, but predicting results on training data the forest will always be accurate. I can assume that you are not using classification but regression, and when you test the model you get not crisp classes -1;0;1, but valid numbers in the range -1 to 1. With a regression skewed like that, no good will come of it.
Better to do it this way, let's say you have a trainingData table where the last column is the target variable and all other columns are predictors:
No, classification, you have something wrong...
I had two models, separately buy and sell classes (1, 0) and (-1, 0)
looks ugly ) agree
I think the approach is great, it's hard to think of something better, the problem is in the market itself...
If you remember, I told you about my experiments, correlations and search for patterns in history (not SSA, but before), what I did:
I took the current situation and looked for its analogues in the past and looked how they ended up, if I found a situation "X", when I found 20 analogues of which 17 ended with a fall and 3 with a rise, so the statistical advantage is obvious(we should sell),by the way, why is not crossvalidation? just by a single pattern, do you agree? We found out that this pattern will not work, we found out that with a huge probability the market will go up, i.e. the market goes against its own statistics with a huge statistical probability.
Ifto make it simpler, if yesterday there was event "x" and after it there was a fall and the day before yesterday there was event"x" and after it everything fell down too and after yesterday there was a fall after event"x"- then if today will come"x"- everything will grow, what crossvalidation will help here?Nothing, never.
Let me explain the most important thing, the understanding of the process.
The market is a cruel business, where some people who have built this business legally take money from others, the crowd, the majority are always obliged to lose, everyone knows about it, everyone writes about it, no one hides it, in short, it is an axiom - a pattern of 95% of traders lose money - the market moves against the majority with the probability that traders lose money with that 95%.
And what does the crowd use when trading? Well, the crowd basically has nothing besides one thing that makes them make deals.
Any actions starting from visual observation of charts and searching for patterns in the breakdown and finishing with neural network training is nothing but trading according to statistics, the very statistics that doesn't work in the market, do you understand what I'm talking about?
the market moves against the crowd's trades ----- the crowd acts on statistics ------ all you need is to predict the crowd's action in the future and do the opposite, the only way to predict is statistics
)))) start with one and end with the third :) Well, that's how it is, at least I got it out of my system.)
p.s. all that i said here is my opinion only, no one is not imposing anything, i can argue and prove it, but not the mood
Ok.
For me Forex is a stupid signal with noises. If I find an addiction I take my money, if I re-train, the broker takes my money. I have no information about the crowds, etc. And how to get enough of it I do not know.
After I finish this experiment, I want to start trading on the stock quotes. There will be real volumes, which is also a plus.
OK.
For me forex is a stupid signal with noise. If I find an addiction, I take the money, if I retrain, the broker takes it away. I have no information about the crowds, etc. And how to get enough of it I do not know.
After I finish this experiment, I want to start trading on the stock quotes. I've got real volumes, which is also a plus.
After completing this experiment, I want to take up stock quotes. There will be real volumes there, which is also a plus.
believe me, volumes won't help you much, you can replace them with regular volatility, I've almost always traded only in the real market, I know what I'm talking about
ok.
I'll try to make input fiches out of the volume and then we'll see.