Forecasting currency exchange rates using a neural network - page 3
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
Then, your coursework should consist of one sentence: "Forecasting the rate with a neural network is not possible". ))) That's the whole coursework ))))
Why is it impossible? Quite possible :). Let's take our favorite Expert Advisor, with which we make our living (everyone has such a "workhorse"). We find a linear regression there, or something like that. Replace the linear regression with neuronics... Voila :)
Then, your coursework should consist of one sentence: "Forecasting the rate with a neural network is not possible". ))) That's the whole coursework ))))
Does it mean that you are disappointed in neural networks and market forecasting in general? Just knowing how many years you've been studying the field it's strange to hear you say that.
I do not study this direction from the point of view of prediction, but from the point of view of finding patterns that can be used with preferably more than 70% probability to make a profit at a certain point in time, i.e. not on every new bar that comes along.
You must agree that these are fundamentally different things.
You do not need to forecast, i.e. continue the instrument line in the future on every new bar in order to profit.
I do not study this direction from the point of view of prediction, but from the point of view of finding patterns that can be used with preferably more than 70% probability to make a profit at a certain point in time, i.e. not on every new bar that comes along.
You must agree that these are fundamentally different things.
You do not need to predict i.e. continue the instrument line in the future on every new bar in order to profit.
But still, at the moment of entering the market, we expect that in the future the price will change in the direction of our entry, and this is the forecast.
Does this mean that you are disappointed in neural networks and market forecasting in general? Just knowing how many years you have been studying this field, it is quite strange to hear you say that.
This means that multilayer neural networks are not suitable for extrapolation (prediction). They are fine only within interpolation.
Simply put, the method is this:
Who doesn't believe it, let him try a simple task, namely to teach a two-way multilayer to a multiplication table in integers from 0 to 9. Then test it by feeding values for interpolation, i.e. any real numbers from 0 to 9 to the inputs. Then for extrapolation, i.e. to ensure that at least one input has values either less than 0 or greater than 9. The interpolation is handled by the grid, the extrapolation is screwed up.
The reason is elementary: sigmoid neuron outputs cut off all values that go beyond the range slightly greater than the range of the training sample. I.e. sigmoids have a limit of values at the output, which cannot be exceeded no matter what value is at its input. See the graph below: The sigmoid input can be any value, the sigmoid output cannot be more than 1 and not less than 0, i.e. outside the limited output range will be a bummer.
It turns out that on the one hand, thanks to sigmoids the speed of training of multilayer networks increases, because without them the algorithms of back propagation of errors lose their convergence. But on the other hand, such meshes become interpolators and cannot work adequately outside the ranges of the training sample.
Still, at the time of entering the market, we expect the price to change in the direction of our entry in the future, and this is the forecast.
Isn't regression also suitable for interpolation? Why build a grid?
Isn't regression also suitable for interpolation? Why bother with a grid?
Try multiple regression and multilayer neural network and compare the RMS for interpolation. You will also learn that all tools have functional limits in which they are most effective.
Isn't regression also suitable for interpolation? Why build a grid?
A multilayer neural network is one of the non-parametric methods of nonlinear regression. In some cases, neural networks can be abandoned in favour of linear/linear/nonlinear parametric methods.