Machine learning in trading: theory, models, practice and algo-trading - page 387
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
than a way to make money.
If you want to make money, the numerae kind of released their own cryptocurrency. Its holders kind of have the opportunity to participate indirectly in investing.
I cannot say more details, I want to do it, but I cannot find it in my hands.
It is interesting that at logloss 0.690 - 0.691 on the validation data almost all showed a good result and the new data, I do not even know what it is related.
Most likely not greedy, not overcomplicated the model and therefore did not overfit, although there was like a guy with ~0.65 on some tour, which in live ~0.68 +.
I am more confused in their data mixed up samples, "era" with unknown id can not reproduce their calculations on the past tours. That is, the hell knows what's in the training set, it would be nice if from the past rounds LIVE samples were calculated by the trash logos then laid out and you could train a model for this test dataset, at least to understand why so divergent. IMHO it's like they accumulate points from a long time ago, and then give out a random subset of them as a train, and the id-schemes do not reflect the chronology of dataset accumulation, and then the live points, which should have been vanguaged, are not given separately to adjust the models and get a "finger in the sky" poking at the chance.
I hope they will fix it :)
And is it possible to withdraw that money?
Yes, instant withdrawal in bitcoins at the current rate.
If you want to make money, numerae kind of released their own cryptocurrency. Its holders kind of have the opportunity to participate indirectly in investing.
I will not say more details, I want to find it out, but I can't find it.
Yes, now they give their crypto to winners together with dollars. For example I've got 300 NMR (Numeraire), but now they can't withdraw them or do something with them. And in general they don't really have this crypto yet, they only give it away.
NMR itself is ethereum crypto token(https://github.com/ethereum/EIPs/issues/20), and investing that they offer is based on ethereum possibilities as well. More precisely, it is not even an investment, but an opportunity to put money on their predictions. You upload your predictions, make a bet in NMR, and then algorithm in ethereum after a while defines the winners and gives out the prizes(https://numer.ai/whitepaper.pdf). Casino, frankly speaking.
Can someone confirm that it's better to use random forests for binary classification? In general case or can I pick up neuronet in special case that will give a little bit less error?
Yes, instant withdrawal in bitcoins at the current rate.
Hmmm... strange.
There's a question why American hedge-fund, with Renaissance uncle patronage, where like the coolest quants on Earth, with $200-300k salary and six-figure bonuses, these prediches from ML fans world (Russia, India, China...), for the price of one average Moscow programmer salary for the whole crowd (>300 people) ?????
As if they don't have their own quants for this? Or Harvard quants don't want to do it, they only communicate with investors and outsource models....
Hmmm....how can they not pay for such lack of foresight...
I've read in articles posted here that the classification world is ruled by gradient boosting (a special kind of forest), in R package gbm for example.They say they overfit too... but I'll read more... Anyway, better than MLP. By the way, boosted decision trees give slightly worse error than simple decision trees in studio
They say they're overfishing too... but I'll read more... At least better than MLP. By the way, boosted decision trees produce a bit worse error than simple decision trees in the studio
Ada is a bit better than randomforest. But caret has some issues with ada (I don't remember which ones), so it's not worth the trouble.
The most promising thing is the selection of predictors. It's all ours.
A large number of predictors were used in this thread, but all derived from one currency pair.
Why one pair and not many?
Why a currency pair and not the predictors?
And where did the macroeconomic data go?
As of today my error rate on the new file is a little less than 30%. 20% is not achievable by any tricks, replacing the models does not lead to anything.
But I don't have the strength to try and answer the above questions.
ada is slightly better than randomforest. But in caret there are some funny things with ada (I don't remember what they are), so it's not worth the trouble.
The most promising thing is the selection of predictors. It's all ours.
A large number of predictors were used in this thread, but all derived from one currency pair.
Why one pair and not many?
Why a currency pair and not the predictors?
And where did the macroeconomic data go?
As of today my error rate on the new file is a little less than 30%. 20% is not achievable by any tricks, replacing the models does not lead to anything.
But to try and answer the above questions I do not have the strength.
I already have predictors, oddly enough. I have a ready bot, which stands on the real, I wrote it in less than a month. The most important thing is the predictors, it's out of question. For example, with my feverish fantasy predictors are chosen at once, I've been working as an analyst for 5 years :) I think selection of predictors is not such a difficult task as studying the architecture of the NS, the main thing is to sit down and pick up, take 2-3 weeks :)
In particular, the latest popular now is LSTM, they are very demanding on computing power, but give an awesome result. I'm thinking now about buying a powerful gaming computer at least for calculations on the video card, at most - NVidia tesla.
From my experience with ns small - they need to be retrained on the machine or retrained further, this is a mandatory process. Using GA for selection of parameters of the same predictors for NS is also a must. All this requires power, but it is really worth it. We already have some semblance of AI, taking into account these 3 components. Overtraining+genetics+powers
SanSanych Fomenko:
As of today, my error on the new file is just under 30%. 20% is not achievable by any tricks, replacing models does not lead to anything.
Very high result even for HFT, could you please check dataset in csv, which gave such results, I would like to make sure that this is not overfitting.
From high and far away sometimes echoed that even the Renaissance minute predicts accuracy no more than 65-70%, given that they are trained by thousands of unthinkable features, including features from satellite imagery processing and data activity of urban infrastructure of cities around the world, all that does nature and the crowds of people on the planet - the data.