Machine learning in trading: theory, models, practice and algo-trading - page 3071
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
classically, a set of signs at closing prices
15 years of OOS
15 years OOS
The approach turned out to be curious, but sensitive to traits all the same. It just doesn't work that way on returnees.
Can you share the catboost parameters? All the ones you changed.
Can you share the catboost parameters? All the ones you've changed.
1 depth,
Stumps?)
And the chips? They said no returnees.
Selection of best models by train or oos?
Stumps, huh?)
And the chips? Said not returnees.
Selection of best models by train or by oos?
Yeah, weeding out patterns
If we are talking about the final selection - then by overall R2 and that the oos does not differ from the traintest. It's a matter of taste and colour.
I've classified the signs so far, because I've been choosing for a long time. I can give a hint that it is better not to use bidirectional signs, which clearly separate buy and sell (like returns), because they lead to a large bias. When the global trend changes, for example. You need signs with a stationary average, preferably.And thefact that on the SB with any distribution should not find any patterns, on rows longer than 200k akurasi +-0.1%around 50%, the correlation of future retourn with predicted less than +-0.01
But if you forecast all kinds of smoothed and cunningly smeared to the right and left indicators (ZZ and so on), then on the SB forecast will be off the scale in inadequacy, as if you can predict the SB with an accuracy of 60 or even 95% depending on the classifier settings. This is nonsense as well as on price series the situation isnot very different.
Even correctly prepared signs and targets, which giveas if honest 1-2% alpha and this is usually looked at with suspicion, there are many more subtle moments (starting from banal persistent autocorrections of trades within the spread ,moving to minutes, to stupid dividing trends, etc. of random nature) .Making even these 1-2% not tradable. But what to do, this is the reality, we need to think in the direction of data and not packages in R.
For years I sometimes come here and see discussions about forecasts with "8% error" and SB on backtestequity ,"it's a shame". It is obvious that such researchers have no prospects, if they did not pay attention to it at once, and then, without prompting, did not figure out how to deal with it themselves.
Ifyou want to predict ZZ, or any retourns or anything else, you are welcome to do so, but use ONLY the right side of the series for the calculation of targets, just as you use ONLY the left side for signs. This is the law, otherwise you get only self-deception. Otherwise, the future is mixed with the past and the past is predicted solely on the basis of the past, which is why the numbers are so grail.
PS. I'm not Egyptian and I'm not going anywhere. Besides, I value relationships above money.
Sanych, you're like a little boy. All my bots are free and are described in the articles. Completely, from start to finish, all sources. Someone does not like free and wanted to pay. Sometimes they don't work. Some algorithms do. Why are you yelling now? This time, a few pages ago, I offered free again. No one contacted personally, so they want to buy?
Send it to me.)
What are you trading? What kind of instrument to fix it under.