摘要： How Machine Learning Helped Improve Our Trading Strategy.
Usually, trading strategies are affected by changes in market regimes. These periods of changing market behavior strongly affect the profitability of trading strategies. Over the past few months, we have been working on creating an adoptive system that would allow us to be less sensitive to changes in the rapidly changing cryptocurrency market.
We took our trading strategy from the price return to the mean and made some changes.
This challenge consisted of the following steps:
1. Search for statistical patterns that have predictive power.
2. Automatic generation of trading rules when the market mode is changed.
3. Development of a machine learning algorithm to filter the best situations in the market.
The biggest mistake in developing trading strategies is not paying enough attention to finding features that have predictive power.
As practice shows, each trading instrument has its own behavior, its own statistical patterns that work for some time.
Here we can split the patterns into two parts:
1. Those that are intuitive. Such as abnormally large trading volumes or open instruments
2. Those that are created by a generative neural network. With the help of a neural network, we can generate features through an autoencoder that have predictive power
Finding signs that have predictive power is probably the hardest part. Many regularities in the cryptocurrency market tend to exist for some time, then they disappear. But the interesting point was that in combination with calendar effects and trading volumes, these patterns have a longer period of existence.
Changing market conditions
If you have been trading for a long enough time, then you probably faced the problem of how your trading strategy becomes unprofitable. This effect is associated with a change in market conditions. Just as a neural network rebalances its weights to optimize its cost function, you must also generate a new rule packet from statistically significant patterns.
By adding machine learning ensembles to this, we were able to achieve even more consistent results. With the help of machine learning, we have filtered out the most insignificant situations on the market, and also were able to find the best combinations of money management for trading instruments.
Result before improvement:
Result after improvement
若喜歡本文，請關注我們的臉書 Please Like our Facebook Page： Big Data In Finance