10-04-2019, 10:39 AM
Hi
I wanted to build a strategy that could learn how to adjust.
Parameters: ADA, 3 months with 60min candles
My first step was to test tried and tested indicators. I started with Tulip, ema.
Attachment: using indicators shows market -37% but strategy did 16%
Next step was to normalize inputs. NN doesn't like 0.00000567. So I normalized to 5.67. I also made sure that all inputs were of the same power(Math) so that no input would overwhelm the algorithm.
I used a feed forward neural network with backpropagation with several hidden layers.
For training, Backtest saved a json file with the results, then it takes the roundtrip data, sees profit % (- or +) and generates training data. This way I have automated the training cycle.
Same parameters with the NN is attachment using NN.
result is 26%
This was just an early test. Going to try add/remove inputs.
I couple of things I've seen. Overfitting the neural network is definitely an issue here. If I feed it too much data, results aren't as good, plus it will behave poorly on other assets. Current version behaves ok when switching assets, but I have seen that is better to train the neural network for each asset as they have distinct behaviours.
I wanted to build a strategy that could learn how to adjust.
Parameters: ADA, 3 months with 60min candles
My first step was to test tried and tested indicators. I started with Tulip, ema.
Attachment: using indicators shows market -37% but strategy did 16%
Next step was to normalize inputs. NN doesn't like 0.00000567. So I normalized to 5.67. I also made sure that all inputs were of the same power(Math) so that no input would overwhelm the algorithm.
I used a feed forward neural network with backpropagation with several hidden layers.
For training, Backtest saved a json file with the results, then it takes the roundtrip data, sees profit % (- or +) and generates training data. This way I have automated the training cycle.
Same parameters with the NN is attachment using NN.
result is 26%
This was just an early test. Going to try add/remove inputs.
I couple of things I've seen. Overfitting the neural network is definitely an issue here. If I feed it too much data, results aren't as good, plus it will behave poorly on other assets. Current version behaves ok when switching assets, but I have seen that is better to train the neural network for each asset as they have distinct behaviours.