Looking into the future is always exciting, especially if it’s the future of markets…
So now we are ready to test our engine. As a simple example we will describe 3000+ experiments with following settings:
Currency pair: ETH USDT
Candle time interval: 1 hour
Exchange:
Y labels: If in the next period the price has grown for 3% or more we put “buy” label, if the price went down for 1% we put “sell”, else “wait” label
Neural Network Architecture: Feed Forward
Seeds: 0 -100
Epochs: 1–34
Class weights: for class “buy”: 3, “sell”: 8
Test set start date: 11.14.2018
Test set end date: 2.19.2019
indicators: MACD(moving average convergence/divergence), RSI (relative strength index), Ichimoku Cloud, OBV (On-balance volume), SR(support and resistance), BB(Bollinger Bands), DMI( Directional Movement Index)
Results:
So after running experiments in these settings, the best model we got was able to make almost 35% profit in 3 months. Out of 3414 experiments, only 307 or 0.9% yielded profitable models, detailed results can be found . History of trades of the most profitable model is .
The experiments we ran is like 0.0000000000001% of the total amount of all experiments that possibly can be conducted. This is not much, but we still were able to get good results.
What do you think about the results? How can we improve them? We appreciate your feedback. See you in a week.
Published at Fri, 19 Apr 2019 09:15:00 +0000
