Machine learning in trading: theory, models, practice and algo-trading - page 2465
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
How does it not remember? That's exactly what it does.
Have you ever come across the expression "neural network based databases"? I came across it once, and I think it's the best definition of what NS/trees are.
One tree can be trained to the last split and then it will remember absolutely all the history with absolute precision (get an over-trained model).
If you stop division not up to the last split, but a little earlier (for example, 10 examples in the list), then we get memory with generalization and with averaging results of these 10 most similar examples. There will be less overlearning. I.e., you have to stop division when underlearning begins to turn into overlearning. This is the main and most difficult task.
I think you should not mix neural networks and machine learning in trading - I liked the picture here - from which I see neural networks - as a possibility for Technoanalysis (where you need eyes, ears, etc., as noted above) or for recursive versions of them,
like when you have too much data on hand, when you don't have the formula you need to help you find a relationship between the inputs and outputs in your dataset, or when you need to make predictions rather than come up with explanations.
Or with weights - (but how to define them logically I don't see how (except for dumb 0 to 1), or if with training (0 or 1), so they don't go astray)
Neurons form layers through which the signal passes sequentially. All this is connected by neural connections - channels, through which data is transmitted. Each channel has its own "weight" - a parameter that affects the data it transmits
And machine learning in a more general form is a decision tree and even a forest of decisions, in which, I agree with you, the main thing is to stop in time, and simple genetic algorithms (in Excel), and statistics with errors and error back propagation for further learning... And probably even the same Monte-Carlo and Sharpe Ratio to evaluate portfolio risk and ways to diversify and hedge (that's where you can really download a lot of data for analysis)
p.s.
And some sort of overview of NN to predict price movements
There is no one correct network organization. Each network architecture has its own benefits and drawbacks. Backpropagation networks are common because they offer good performance, but are often difficult to train and configure. Recurrent networks offer some benefits over backpropagation networks because their "memory feature" can be used to extract time dependencies in the data, and thus enhance prediction. More complicated models may be useful to reduce error or network configuration problems, but are often more complex to train and analyze
- I.e., memory will only be available if you put it into the network architecture... But to put memory into a model of a variable (like price) seems to me very naive and reckless... Memory can be put in the analysis of demographic data, seasonal fluctuations, or anything else more systematically repetitive and constant... BUT NOT price movements for day-trader... At the very least, in the analysis of the losses... imho... or a little memory of recent K's only if you watch (and even Flat/Trend is only 0.5 probability)
(for the reasons described, I guess this argument is really pointless - when speaking without reference to a particular architecture of a particular network or other machine learning variant) ... but thanks for the warning
With memory you don't have to go far (or rather, without memory - run and forget, intermediate memory data are no longer important in Output):
Monte Carlo is a decision-making tool that assumes that every decision will have some impact on overall risk.
how the Monte Carlo method is used to optimize a portfolio -
First, random weights are given to stocks, and then the return and standard deviation are calculated. The obtained values are saved. The next step is to randomly change weights (the main thing is to remember that their sum should be equal to one) and everything repeats - the calculation and saving the obtained value. The number of iterations depends on the time, the capacity of the computer for the calculation and the risk that the investor is ready to take.
(Although in Excel with the Package of the Analysis and search of decisions - probably, even easier realization - the question again in the model remains, and the genetic and other algorithms are set there in the search of decisions on a choice - and here it is " a method of stupid search" is ready) ... but for risk management, and not yet prediction of price movement
After all the answers you have to make a Slicer of conclusions, because so far everyone is talking about his own... (the theme is - there will not be one conclusion - there are many methods - there is no common denominator)
As in the tree that elibrarius mentioned.
I.e. you have to stop the division when underlearning starts to turn into overlearning. This is the main and most difficult task
After all the answers you have to make a Slicer of conclusions, because so far everyone is talking about his own... (the theme is - there will not be one conclusion - there are many methods - no common denominator)
as in the tree that elibrarius brought up
This is all for the casino, it's different in Forex. What can you explain to the robot? When the blue one is higher than the red one, buy. That's all.)
Vladimir, I have a dozen unique positions/states
If you flip through stackoverflow, you can see a very simple logic (just like in the Find Solution add-in in Excel):
0/ be honest and use the mean and scaling inferred from the training set -
1/ Neural network remembers what its learned through its weights and biases.
2/ Randomly initialize weights - this shifts the responsibility to the machine - WHAT'S GOOD, not the developer
3/ train it with some patterns - on big data (and by the way big power of pc) - either regression or logistic model or other (worth thinking about and choosing - not much to choose, because by choosing a linear regression, we will take away the opportunity to work optimizer, so I think)
4/ ... enter input and get the result from a trained machine AND ALL
p.s. remember Continuous vs Discrete artificial neural networks -- the 1st is probably an integral function, the 2nd is for determining a flat/trend, for example...
and some advice about fashionable
Recurrent neural networks have at various times been a fashionable method for various financial prediction applications, for example
p.p.s. A bit of theory
In Bayesian networks the vertices and edges have meaning- The network structure itself gives you valuable information about conditional dependence between the variables. With Neural Networks the network structure does not tell you anything.
and python libraries (you can also use R), if excel is not enough
(while those who can neither read nor understand still dream of explaining something to their robot, because they haven't played enough at the casino)
... but the linear regression is confusing because, as I recall, prices are non-linear and return is linear (if not the other way around?)
NOT the other way around, since.
Although the prices are not normally distributed, the price return in many cases has a normal distribution.
NOT the other way around, because.
although logic to the contrary, but in conjunction with distant memory, is confirmed... (although I don't often deal with Quantitative Finance and Financial Researches)I made a working MT5 application for the end user, it is installed in two clicks.
Neural network is launched on user's computer in exe file (you can do it without exe, in this case Python script).
No external requests, API, dll, etc.
Forecasts in the form of standard indicator.
Click here for more info.