Machine learning in trading: theory, models, practice and algo-trading - page 1106
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
The small ones we call a flat, the large ones a trend.
But here again the question is what are small and large movements, relative to what they are small or large?
In fact these are serous questions that give an answer to the question why parametric systems and "mo" as well will never work on unprocessed market data
Well there are no clear definitions - where is the flat, where is the trend, small or large. For some applications are small, for others - big.
Define what exactly is small for you (your system) and what is big in any particular units, and everything will immediately fall into place.
Anyway, here's the deal. I started to use boosting, because in addition to accuracy and high generalizability in this algorithm model building is more unambiguous. The method is also easier to set up because of a small number of specific external parameters. The only disadvantage is the loading of RAM when calculating, and accordingly the model size increases in tens and hundreds of megabytes depending on the number of iterations. As the result of comparison with random forest and shallow neural network methods I came to conclusion that boosting is more preferable for classification tasks.
I tested a lot of predictors. Mainly they are sequential time series formed from the most various indicators and their combinations. Testing was performed in multicurrency mode (27 currencies) by the program method taking into account the real spread (2 points). Timeframe - an hour. In the output - a binary class, calculated using a zigzag signal with a step depth of 100 points. Almost all results are negative. If you exclude the spread, the plus can be substantial. As an option, you can try to take a higher timeframe.
I have in mind how to further develop the study:
1. to try another type of output zigzag or other parameters.
2. to use cyclic component signals selected using the Fourier method or wavelet filters.
3. use real values of indicators on the output (regression). For example, difference of close and open prices of candlesticks, price change for several bars ahead.
4. use inconsistent data as predictors, for example, key points or levels
5. filter the initial sampling on different indicators (Volumе or ATR indicators), i.e. to teach working only on certain parts of the market.
I will be glad to listen to your opinions and advices.
I am waiting that someone was able to achieve a result from the many clever words expressed in this thread, I personally do not need either source code or algorithms, but the result of the MO in the form of trapped signal or screenshots for a couple of days. And so far all we have is talk and irrelevant.
As for the flat and MO, in fact AI will find the necessary probabilistic behavior at the current moment in the process of learning whether it will be a flat or a pulse. So I don't see any sense in writing separate algorithms for determining the flat, it's useless.
This article will show you a real trading robot, that is, in its sound mind, no signals or financial reports will ever show. Do not even hope. There is and will be nothing but talk.
As for MO, these methods have nothing to do with AI. And MO methods will not find anything unless you show and tell them what to look for and where to look. Otherwise it will be like in a hackneyed phrase: garbage in - garbage out, and nothing more. No matter what homegrown gurus say, one of the main problems of using MO is the preparation of representative data. And all sort of trend-flat splits may be necessary to prepare such data, instead of indiscriminately feeding everything to the input of the MO.
Waiting for someone to be able to from the many clever
Waiting, waiting....
Open your eyes already))
Here's an example of a neural network on levels.
Red oversold, green overbought...
Look at the picture before, the eu was overbought and here's the reaction
(Forecast is on the air)
Waiting, waiting....
Open your eyes already))
Here's an example of a neural network on levels.
Red oversold, green overbought...
Look at the picture before, the eu was overbought and here is the reaction
(It's a live forecast).
Now let's draw a regression line and establish a channel, and all these levels will coincide with the channel's borders. And overbought/oversold will no longer be needed.
No real working trader, in his right mind, will ever show you any signals or financial reports. Don't even get your hopes up. There is and will be only talk.
As for MO, these methods have nothing to do with AI. And MO methods will not find anything unless you show and tell them what to look for and where to look. Otherwise it will be like in a hackneyed phrase: garbage in - garbage out, and nothing more. No matter what homegrown gurus say, one of the main problems of using MO is the preparation of representative data. And all sorts of trend-flat splits may be needed to prepare such data, instead of indiscriminately feeding everything to the input of the MO.
exactly
Now let's draw a regression line through all this and build a channel, and all these levels will coincide with the borders of the channel. And overbought/oversold will no longer be needed.
Explain
explain
I'll try your picture.
I'll try it on your picture.
Please, it is the best.
Please, that's the best way to do it.
This is roughly what it would look like. The regression line can be replaced by a long EMA
.