Machine learning in trading: theory, models, practice and algo-trading - page 3288
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
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 predictor values be the same as the predictor values in the learning plot or not?
Judging by the pictures, the Recall is also low here, i.e. the model is not confident in anything and is very cautious in forecasts.
There are also imprudent models - just take less clean leaves (not 0.9, but 0.7 or 0.5), but they drain (like the one above in that picture, the one below uses cleaner leaves and successfully skips drain periods/months and years). I.e. the pattern in the chart is the same - the differences are in the cleanliness of the leaves used.
Overall nothing worthy that I would put money on I haven't found.
There are also careless models - just take less clean leaves (not 0.9, but 0.7 or 0.5), but they drain (like the one above in that picture, the one below uses cleaner leaves and successfully skips drain periods/months and years). I.e. the pattern in the chart is the same - the differences are in the cleanliness of the leaves used.
Overall nothing worthy that I would put money on I haven't found.
I see often beautiful pictures in my research, but the confidence is not enough that tomorrow it will be like that.... so far also not very satisfied with the results. I'm changing the learning process.
Now I started to write an article, but I realised that I do not want to describe my method in details, so I am looking for interesting results in experiments with CatBoost.
Why don't you want to switch to partitioning the sample not in a continuous way, but by patterns, as I do?
I see often beautiful pictures in my research, but I am not sure that tomorrow it will be like that.... so far I'm not very happy with the results either. I'm changing the learning process.
Now I started to write an article, but I realised that I do not want to describe my method in details, so I am looking for interesting results in experiments with CatBoost.
Why don't you want to switch to partitioning the sample not in a continuous way, but by patterns, as I do?
I didn't do MO all summer, only here I could discuss something. There will be nothing to do in winter - maybe I will experiment some more.
I envy all of you, your enthusiasm and non-stop testing and experimenting....
Yes - in the solid disappointed. The growth that is on the charts is obtained with a long wait for the completion of the deal - up to 10 days. I reduced it to 3-5 hours and it turns out to be only random on OOS. I came to the conclusion that this profit from night entries is obtained and it takes more than 3 hours to complete the deal (i.e. they are completed mostly in the daytime). Daily entries 300-500 pts usually in 1-3 hours passes and trades are completed - that's what they show random.
I made a trading system with a good profit on the Moscow Exchange a long time ago, but only in the tester - alas, it turned out that I did not take into account the morning price rush - which in a minute could fly by 1000 points on Si and the terminal at these moments liked to move.... now I make closing for the night during training. I don't really support the idea of closing after a set time, I still think that it is more correct to close either by market event or by timing of the beginning of a new bar (for example, we opened at the minute and closed at the new hourly). I use trawl or levels from ATR more myself.
I have not been working on MO all summer, only here I could discuss something. There will be nothing to do in winter - maybe I will experiment more.
I envy all of you, your enthusiasm and non-stop testing and experimenting....
On the contrary, your behaviour is more peculiar to a normal person. I'm the one who wants to drop everything at times, but constantly coming up with ideas that I want to test..... that's excessive viscosity.
The right thing to do is to do it all for a salary or as a hobby. Enjoy the process.
I don't really support the idea of closing after a set time, I still think it is more correct to close either by market event or by timing the start of a new bar (for example, we opened at the minute and closed at the new hourly). I use trawl or levels from ATR more myself.
I don't time my closes. I just check the reached TP or SL, looking ahead 1-3-10-100 hours. The basis is this indicator https://www.mql5.com/ru/code/903 - parameter bars_future = 10 (I made it 10000 to make sure all bars are marked)
For the case of 3 hours, if TP/SL is large, 300-500 pts per night does not always pass, so such trades are left as "I don't know". When there was a peek at 10000 bars, they were marked and participated in training and I think that it was due to these overnight entries that the profit was made.
I tried Traal - I didn't like it. It catches trends, but rarely with pauses of 1-2 years.
Well, rather not as a horse among people, but as a guiggnm among ehu).
The issue of markup (teacher construction) for price data is very important, but I doubt it will be possible to discuss it in a meaningful way. There are too many different ways to do it, so everyone speaks in languages incomprehensible to others. And fear of blabbing about my know-how, of course) That's why we get something obscene in the thread.
take and take
what's not clear?