Machine learning in trading: theory, models, practice and algo-trading - page 744
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
Remember I said I got a model that has been dialing in since 01.31.2018, and here's how this model has been working out these two weeks since 03.05.2018 to the present day. Tester result.
Pretty good for an old lady trained on 40 points and has been working for like 1.5 months on OOS.
And this is her full OOS from 01.31.2018
And you still think it's a fit???? Let me remind you that in the screenshots the OOS section
Although these are pictures from the tester, I never saw the monitoring. But I believe you. I have to admit that your approach works. Therefore I apologize.
Apology accepted!
I'm just a practitioner, and most people here are theorists and researchers...
And it's impossible to find these connections mathematically, so you have to use dumb fitting or market research :)
Stupid fitting is also a cool thing, actually, if you use generalization
Max, I wonder what a neural network or a Random Forest does. In the future, when such a "pattern" appears, the machine will easily recognize it. The fact that this is a 50/50 forecast is true for everything in the market. Take the example of a breakout of a triangle, for example, in the classical version take is greater than foreclosure. Now let's multiply this case by 50/50 and we have a profit as a result. This is the simplest way to explain how to make a profitable system using machine learning.
Max, I wonder what a neural network or a Random Forest does, ... Then, when such a "pattern" appears in the future, the machine easily recognizes it. The fact that this is a 50/50 forecast is true for everything in the market. Take the example of a breakout of a triangle, for example, in the classical version take is greater than foreclosure. Now let's multiply this case by 50/50 and we end up with a profit. This is the simplest explanation of how to make a profitable system based on machine learning.
The main problem is that training with a tutor does not find anything by itself and the ratio of features to the target is not always chosen optimally by us, hence the big mistakes in classification and retraining and a lot of topics on how to optimize this process. If we talk about a full-fledged NS bot, it should mark tags in an optimal way without expert (human) participation. How it is realized today - I sent you some links, for example, through reinforced learning, but there are some difficulties, such as exploration and exploitation problem, i.e. find balance between studying the environment and application of obtained knowledge, in fact, it's the equivalent of the dilemma how often you should retrain the NS, but in automatic mode
The main problem is that learning with the teacher does not find anything by itself, and the ratio of features to target is not always chosen by us in the optimal way, hence the big mistakes in classification and retraining and a lot of topics on how to optimize this process. If we talk about a full-fledged NS bot, it should mark tags in an optimal way without expert (human) participation. How it is realized today - I sent you some links, for example, through training with reinforcement, but there are some difficulties, such as exploration and exploitation problem, ie find a balance between learning and application of the knowledge, in fact, the equivalent of the dilemma how often you need to retrain the NS, but in automatic mode
I will not be too clever, because I have very little theoretical knowledge. I can only give an opinion of the observer and the practitioner. In fact you can do with even 2 masks, but then the result is corresponding, I won't tell you about the way these masks can be taught, it's not the main point. So as an observer a lot of tests I can say that the frequency of retraining the NS is far from stationary thing, sometimes it turns out one set of enough for example once a week, and other times it happens that once a month. for different sets of data different frequency retraining. but in the end we still get the fit, but the fit is not parameters of the same machine and adjust the frequency of the signal from the wagon specified period. How long can such a fitting hold up? It's like in a swamp, you never know when you enter the swamp.
I will not be too clever, because I have little theoretical knowledge. I can only express the opinion of an observer and a practitioner. The selection of predictors is a very tedious process, a lot of big is not necessary. actually even 2 masks can be managed, but then the result is appropriate, I will not tell about how these masks can be taught not the essence is not important. So as an observer a lot of tests I can say that the frequency of retraining the NS is far from stationary thing, sometimes it turns out one set of enough for example once a week, and other times it happens that once a month. for different sets of data different frequency retraining. but in the end we still get the fit, but the fit is not parameters of the same machine and adjust the frequency of the signal from the wagon specified period. How long can such a fitting last? Here it is like in a swamp, you do not know when you enter the swamp.
Well, you need enough descriptions of the environment and the right switches, roughly speaking, from mode to mode... because the patterns change, yes
Some people solve this problem by switching different TS, some are trying to make one but adaptive, and some are trying to fit everything to a single distribution, like Alexander
Mishan has made a profit on the growing market and while it is growing he is rejoicing, but as soon as the turbulence starts he will cry.
Well, you need a sufficient number of state descriptions of the environment and the right switches, roughly speaking, from mode to mode... because the patterns change, yes
Some people solve this problem by switching different TS and others try to make one adaptive TS, and others try to fit everything to a single distribution, like Alexander did
Mishan has got the profits on the growing market and while it is growing he is rejoicing, but as soon as the turbulence starts he will cry.
I hope he doesn't cry but reconstructs in time... It's not like we're here to argue...
God forbid that he didn't cry and rebuilt in time) we're not here to argue, as it were...
well, the coin game without a normal backtest, the outcome is just obvious
Good day, everyone.
I wanted to summarize a little bit... What do we know about the future candlestick, for example? We know the opening time and the closing time. We know that it may have 3 states: a white candle in the up direction, a black candle in the down direction and a doji. We know that the probability of a "long" or "big candle" you know)) - is small compared to the "average" candle or doji. We can find a channel, or call it a range, in which the price moves. That's it? We don't know anything else? It's too small to make a forecast even for a simple classification like a down candle or an up candle... If you don't try to predict directions... you can't enter a trade without predicting the direction... it won't work, so you have to predict it anyway. What else can we say about a future candlestick that would allow us to classify it? After all all the forecasts based on past data give signs of past candlesticks. And the prediction on this data is presented in the form of "today will be like yesterday" - this is not good....