Hello dear readers of our blog.
In today's post I want to tell you about the test calculation function of the future price level.
It is based on a simple neural network that is trained on the pairs in the selected timeframe.
Next I will tell you how there is a computation and learning network, but first I want to tell you what it's done.
After all, to know the level of prices for the work of the strategy is important, but more important is the levels within which the price will move and the time to achieve these levels.
Understanding the range of price movement at time intervals, you can automatically assume with probability, where and when would be the price.
This is an important filter to screen out noise on the input and in conjunction with our trade strategy will help to increase the number of profitable trades to increase the profit on the course, eliminating unprofitable transactions.
After all, if lost - then earned.
Now about the mechanics of training the neural network for the prediction of the price levels and time to achieve these levels.
We hypothesize that in each moment of time, each value has the values of many technical indicators, which characterize the position of a given point and its previous writing.
Knowing the history of price movements, we calculate for each point the indicators that require standard input data, namely the values of the prices of open, close, highs, lows.
Using these data, we calculated indicator values from the library TaLib. During testing and development is used more than 30 indicators, in the future, after passing the tests and to improve the results of prediction are used are all available, and more than 150 indicators.
Normalizing on the price value, the indicator values, all of this is fed to our neural network. Further, we propose to make a prediction network based on random weights assigned to the neurons of the network during its generation (at this point, the network did not know how or understand).
Assessing the response of the network, we compare it with the real price value on the history and ask her again to try to give an answer to obtain an acceptable response with an error not more than specified.
At this point the weight in a neural network are distributed, and grid trained for a specific pair and timeframe given the kind of price movement.
And so we move to obtain a trained neural network for all available history and have adapted the neural network for the selected pair that is able to predict with a maximum specified error.
This solution is not a panacea and not a solution to all problems in algorithmic trading, but only one of the filters, which will be built into the strategy.
During testing I will be writing about the latest developments to keep you up to date.
Regards team Elinesoft.
The Project Easytrading.pw