Machine learning in trading: theory, models, practice and algo-trading - page 152
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
Please stop with the accusations.
Every language has its place. R is great for interactive research. This is my second day exploring it (I read the book before) and it really looks like a powerful debugger with visualization of the guts.
Working with R has immediately revealed our weaknesses:
We have released the first algorithmic trading platform with the MQL language back in 2001. Each time we increased its possibilities, but the mathematical algorithm left much to be desired. We were developing the analysis, data access, tester, distributed calculations, and then we started to sell our products.
And then it became clear that most of the solutions were stuck in a vicious circle of theanalysis, indicators and fitting. We need to let the developers get to the next level of mathematical capability.
That is why we started extending mathematical libraries in MQL5 some time ago and also released in beta Alglib, Fuzzy and Stat. They will make it easier to transfer worked-out models from other systems to MQL5 and raise the class of created analytical solutions for the Metatrader 5 platform.
In the next 2 months you will see the progress we will make in developing the mathematical environment.
We welcome and welcome discussion of complex mathematical packages as well as articles on them. Write and send requests for papers to Rashid Umarov (Rosh). Our objective is to encourage and educate traders in more sophisticated methods, not to fence them in our own MQL5 world.
Of course, we are and will be defending our language and platform from attacks, but we are also working to develop them. So everything will be fine.
I'm glad that the long discussions and your brief acquaintance with the language are bearing fruit.
Your direction of development is clear.
Time will show how successful it is.
The main thing today (at least for me) is that MT4 works like clockwork. The rest will be done.
I am preparing articles. True, sometimes I am haunted by doubts, whether someone needs them ...
Good luck
Convergent NS are in trend now as comparatively recently were RF, boosting etc., but context of their application is limited by pictures, in fact only CNN trick is hierarchical selection of filters when there are obvious relationships between adjacent components of input vector, it can be done another way without backprops, for example by clustering. But the main question is still another one: what to feed to the input.
I will add: DNN, RNN, CNN, LSTM and of course "reinforcement learning" LR are promising.
I think I'm not discovering America especially at the end of 2016 that tn. "TA patterns" i.e. price "figures" of one series, now have no connection at all with the direction of the next tick or their average direction for some time horizon. It is like predicting the ball at a soccer field and using it as a predictor for the scoreboard, the correlation is weak, dozens of times less to recoup transaction costs. I think that the market is decentralized, unfortunately there is neither a standard feed nor a normal one, the ones that are available are fake. In general, to trade on the Forex ... Well, you know what I mean)))
This is the common opinion of traders on the fund. The market is fine. But you're wrong about predictors.
I'm working at the quant-fund and I can only talk about obvious things, it would be stupid and illegal to even hint at how to look for effective predictors, but it's pity to see how intelligent people go through such "patterns" as head-shoulders and various candlestick shifts, setting ML on it, The dependence in the price is minimal and decreases exponentially, the past price history of one series of a given length is connected with the future price history of the same length by the value beyond 3 sigmas, the deeper history makes no sense at all. Look for PRINCIPLES and the fastest possible sources of data on anything that might in any way affect the behavior of the masses of people all over the planet and in that whole primary broth, look for the precious ALPH.
Thanks for the advice.
Good luck
It's nice that long discussions and your brief acquaintance with the language are bearing fruit.
Perhaps we will make a splendid gift to the R community.
Time will tell.
It is possible that we will make a splendid gift for the R community.
Time will tell.
It is possible that we will make a gorgeous gift for the R community.
Time will tell.
We look forward to working and not giving up hope.
Good luck
I prepare articles. True, sometimes doubts overpower me as to whether someone needs them...
Good luck
They do...
I learned from your article.
Maybe we'll make an awesome gift for the R community.
Time will show.
We are waiting :)
1) in fact, the only trick of CNN is hierarchical selection of filters when there are obvious correlations between neighboring components of the input vector, it can be done in other ways without backprop, for example by clustering.
2) But the main question is still about something else: what to feed to the input.
1) A narrow understanding of the question. A convolution network filter can perform many operations on the input vector, and form, for example, n moving averages, partially overlapping or not, or take several sums. Which is already a claim for normal feature engineering.
2) Again, understatement or misunderstanding. For convolutional networks, it is enough to feed "raw" data (in a pseudo-stationary form), such as price returns with a lag of 1. The rest is done by the network itself.
1) I don't quite agree yet, I think it's quite possible to predict the price by searching for similarities in the past but it's not "TA patterns" unequivocally.... research is still underway but very time consuming in terms of mach. time
2) I agree, the model itself has little effect
For example if we take the Ribbon, separately divide it in buys and sales, then build their cumulative sum and draw the difference between these sums - you will get the same price, the sliver is a ticker, on the contrary, it predicts the price, at the highly liquid markets, the price always follows the higher liquidity, if there are more bids to buy than to sell in the sliver, then the market almost always goes down, but the difference between changes in the sliver and the price change is measured in millisecondsIf there are more bids than sales, the market almost always goes down...
I once checked the full orderlog of a symbol, many people here don't even know what it is.There is also indexation of a concrete participant in the deal and if there are 100 sell contracts in the market you can check if the only one person put this order of 100k or if this is the number of people who placed one price and it made 100k in total, you can also check if this person was taken or only 20k were poured for 100k and he cancelled the rest 80k etc.д..
So what am I saying all this, returning to the cup and the speed, there was a time that I calculated a guy who for a few ms managed to put up and cancel his request 4 times, just manipulated the price does not let it drop or rise, I have the same from the terminal only my order to poke or sell goes 0.8 sec.
4) Yes you are engaged in arbitrage in its many different forms what is there to hide :) from hft to long seasonal...
1) That's right, there are weak dependencies, I gave an example about soccer and scoreboard as a chip.
3) Well, about the "niches are busy", supercomputers, geniuses PhD and cables across the ocean for $ 300M, if it scares you, you should not do algorithmic trading, at least in hft on Si\RI more than 30% of elimination, either singles, or small groups of 2-3 people and make even a small but small profit, I mean not all hft occupied by large quantfond that can not squeeze in, and generally it will never happen, the more fund the harder it is with hft.
Yes, you have to work with a lot of order-logs, every order and moreover a deal on each instrument has an impact on the probability of the next orders and deals of all the other instruments.
4) No comment)
1)Non-disclosure signature in the quantum fund?
2)Should not be popularized?
1) Exactly
2) So it is, well, what do you think, if for example you were lucky by many years of work, for example, to find a way to quickly find places where at depth a lot of oil, it would be worth it to popularize this knowledge? Even if you are a philanthropist at heart, it is still more beneficial for you to invest or donate your hard earned money to the needy than for some other people.
So, according to you it doesn't make sense to trade on large TF (day, week)? But large banks successfully trade on forex exactly for long term.
Well, banks not only speculate, but if about long-term speculation in the style of Buffett, then of course it makes sense, if you know what others do not know))) The longer the horizon the worse the statistics work, in such cases you need to have connections in the government, a team of high-class macroeconomists and so on, it's another game, it's closer to politics than to algotrading, it's just like predicting the weather for the year ahead without being God.