Machine learning in trading: theory, models, practice and algo-trading - page 3287
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
😀
I don't even know what the squares mean. Probably what the apple guy writes :)
If "something" is found, what is its lifespan outside of teaching usually?
For me once a week is acceptable in terms of speed - 5 years for 260 retraining approximately passes.
We can tentatively conclude that indeed, the success of training depends on the sample size. However, I note that the results of the "-1p1-2" sample are comparable, and even better by some criteria, with the "-1p2-3" sample, while for the "0p1-2" sample the results are twice as bad in terms of the number of models meeting the given criterion.
Now I have run a sample with inverted chronology, where the train sample consists of the initial exam+test+train_p3 sample, and the test sample is train_p2, and exam is train_p1. The goal is to see if it is possible to build a successful model on more recent data that would have worked 10 years ago.
What do you think will be the result?
I didn't have to wait long - the result is in the last column of the table
I'll try to comment in a non-biased way. On the one hand, we can say that changing the chronology of the sample significantly worsened the results by the main criterion - the number of suitable models, but on the other hand, the very fact that such models were found says that there are some stable patterns in the data. Or is it random? Yes, of course, we need to take other samples and conduct similar studies, and only then we can draw conclusions with more confidence. For now, this is just information to ponder.
Objectively, there is a lot of data - I usually use multiples of smaller samples for training, albeit comparable to the chronological course. The sadder the Recall rate looks in all the experiments. I'm even surprised that nobody has paid attention to it. In general, we can repeatedly say that the standard metrics give a poor indication of the financial result if fixed takeouts and stops are not used.
If you have any ideas/wishes what else to combine here (sample plots) with something - tell me - I'll try to check what will happen.
I just retrain the model once a week. It may live longer, I haven't researched it.... but maybe less and it is necessary to retrain like SanSanych at each bar (if H1, then in principle it is possible).
For me once a week is acceptable in terms of speed - 5 years for 260 retraining approximately passes.
What is the result of such retraining in aggregate?
What is the result of such retraining in the aggregate?
Max, I don't understand why you're making fun of me.
If there are no assumptions - don't say anything, if there are - say it, like "the result will suck".
If it's funny, you can laugh, and if it's sad, you can cry.
What Aleksey Vyazmikin is discussing here cannot cause suggestions and it is impossible to evaluate "shitty - not shitty".
Example, a man comes up and says:
- let's send an iron to the moon.
We look with surprise.
And the man says:
- and we'll change the temperature of the iron and put different amounts of water in it.
Will you react " if there are no assumptions - keep silent, if there are - speak out, like "the result will be shitty"?
What leksey Vyazmikin does has nothing to do with the problems of the MoD. He pulls from one opera and tries to get an answer from another opera - all empty chatter of a man with a mess in his head.
All the graphs I've shown for the last year are obtained this way. Only OOS by Valking Forward.
Judging by the pictures, the Recall is also low, i.e. the model has little confidence in anything and is very cautious in forecasts.
I just retrain the model once a week. It may live longer, I haven't researched it.... but maybe less and it is necessary to retrain like SanSanych at each bar (if H1, then in principle it is possible).
For me once a week is acceptable in terms of speed - 5 years for 260 retraining approximately passes.
I have found out one fundamental problem here: looking ahead. It manifests itself in the following way: we take pieces of one large file, teach them, then test them, check them - everything is normal, the error is about the same. But as soon as we run outside these three files, which are pieces of one large file, the result is fundamentally different, usually catastrophic.
If we retrain at every step, the "looking ahead" problem is eliminated, because prediction is done on the same predictor values as training.
And if you don't teach at every step, then all predictors, including the training section, are taught on some values and then predicted on them. And here is the question: will the new values of predictors coincide with the values of predictors in the learning plot or not?
Here I found out one fundamental problem: looking ahead. It manifests itself in the following way: we take pieces of one large file, study them, then test them, check them - everything is normal, the error is approximately the same. But as soon as we run outside these three files, which are pieces of one large file, the result is fundamentally different, usually catastrophic.
If we retrain at each step, the problem of "looking ahead" is eliminated, because prediction is performed on the same predictor values as training.
And if you don't teach at every step, then all predictors, including the training section, are taught on some values and then predicted on them. And here is the question: will the new values of predictors coincide with the values of predictors in the learning plot or not?
Where is this peek you have?
I read it a couple of times and didn't see the logic behind the words.
They invent problems - then heroically solve them - one is looking into the future, the other is looking for the perfect markup....
What leksey Vyazmikin is doing has nothing to do with the problems of the Ministry of Defence. He pulls from one opera and tries to get an answer from another opera - all empty chatter of a man with a mess in his head.
If you don't understand what I'm doing, ask. Yes, often my experiments go beyond academic knowledge.