Machine learning in trading: theory, models, practice and algo-trading - page 2356
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 only found a reference that he worked as a master machinist for six months at AQR (about 150 billion under management).
Anyway, the book is only useful as an outline of the process as a whole.
I have a great-grandfather who was a chief machinist, had a nice estate and a house in the center of town. Probably genetic memory.
Has anyone already experimented with non-price data, but with fundamental indicators, for example?
Approximately so - the input is all possible fundamental indicators of all major countries and the world economy at the beginning of the day (hour), a classifier or regressor of price movements for the day (hour).
Has anyone already experimented with non-price data, but with fundamental indicators, for example?
Approximately so - the input is all possible fundamental indicators of all major countries and the world economy at the beginning of the day (hour), the classifier or regressor of the price movement for the day (hour).
We've been studying the works of Prado lately). Therefore, I will answer with a quote from his book: "Fundamental data is extremely regularized and low-frequency. Given the public availability in the marketplace, it is unlikely that there is any value in them for exploitation."
Long fiddled with the models with the markup by TP and SL.
For the basics I took TP=SL=50. This way you can immediately see the success of the training just by mistake. At first it was about 45% on the OOS. By a lot of tricks I have brought it up to 40% that could be used in real trading. From 13000 trades 5300 lost in 8 months, i.e. about 2500 trades net profit of 50 pts. Drawdowns are about 100 trades in a row, that is, no more than 1% of the deposit can be spent on a deal, and the better to spend 0.5%.
But as it turned out experimented with a successful section of February-September 2017, when there was a constant growth of the euro.
Shifting six months forward or backward or increasing the OOS plot to 1-2 years leads to an error of 49% ... 55% on the OOS.
Trace with 0 errors and regularized to 10, 20, 30, 40% all tried. Always 50% (except for lucky 2017).
I.e. it is simply impossible to predict whether the price goes up or down by 50 pts better than by 50%. I trained on price deltas. ZigZags. Adding volumes from CME adds 2-3% (in 2017), but on other periods CME volumes do nothing.
Variants with TP! =SL when recalculated also lead to zero profit.
In general, I think that the work with TP and SL markup is unpromising.
I now plan to go to training without a teacher (with a random teacher markup). But apparently there is almost the same success with drawdowns in six months to a year (judging by recent articles).
Now I plan to go to training without a teacher (with the occasional teacher markup). But apparently there almost the same success rate with drawdowns in six months to a year (judging by recent articles).
It's no use either... It is even worse than with the teacher (more "expensive") with the same or worse cuts
In general, I have another look at finding patterns and ways to extract them, but to acquire such a vision must be trained with 1000 + models ...
I think (I'm sure) the secret is to expand the space of features, models have "information hunger", they are too simple, too primitive ... They(models) are a reflection of us, the more primitive we are about the market, the more primitive our models are (how else could it be?)...
A foolish person says "what about those signs, we have returnees, everything else is derivative".
The thinking person understands that even locally they need considerable power ... I'm not talking about some 10-layer neuronics (it's too popular for fools), I mean common sense ...
I came up with a pattern finder (only in my head so far) with interactive visualization and sliders. I.e. I can very quickly find patterns (if there are any) like seasonal patterns, but an extended version (seasonal patterns with conditions).
i found a cool library for visualization of all this (python), i will have to learn
you can immediately generate bots on the fly, if something is foundWhat is the meaning / logic of the search?
Well, this is before the search. search for small forces and if something is found pull up the forces.I want to make a tester, which from OHLC will calculate profit of trades inside MO model (without output to tester).
I keep thinking about my idea to add spread and commission to price, if there is one, at moments of buying when opening a trade or at closing if opened to sell.
The standard field stores the minimum spread per bar. The only option is to run the test on all real ticks and find this real spread. But it's long and will have to be stored in files, which will complicate the overall system.
Maybe just use spread + commission + another 10 pts. The number of successful trades in the tester will decrease, up to the fact that there will be no successful model at all.
On the other hand, the 10 pt will cover slippage at the native brokerage company and worse spread conditions at another brokerage company.
What do you advise?
I want to make a tester that will calculate profit of trades inside MO model (without output to tester) from OHLC.
I keep thinking about my idea to add spread and commission to price, if there is one, at moments of buying when opening a trade or at closing if opened to sell.
The standard field stores the minimum spread per bar. The only option is to run the test on all real ticks and find this real spread. But it's long and will have to be stored in files, which will complicate the overall system.
Maybe just use spread + commission + another 10 pts. The number of successful trades in the tester will decrease, up to the fact that there will be no successful model at all.
On the other hand, the 10 pt will cover slippage at the native brokerage company and worse spread conditions at another brokerage company.
What would you advise?
Yes, but what bar do you mean? Hourly, daily or tick bar?
You can advise, but which bar do you have in mind? Hourly, daily, or maybe a tick bar?
M1-M5