ref：Andrey Shtrauss. 'Building an Asset Trading Strategy', 2020. https://www.kaggle.com/shtrausslearning/building-an-asset-trading-strategy
This kaggle passage develop a mechine learning stragtegy for Bitcoin trading. I'm going to implement the same strategy in Shanghai Securities Composite Index.
Use trading signals as target variables to extract feature values from price information and technical indicators for machine learning. Specifically, taking the dual moving average trading strategy as an example, a buy signal is formed when the short-term moving average breaks above the long-term moving average (set to 1), and a sell signal is issued when the short-term moving average falls below the long-term moving average (set to 0), so that the signal variable constitutes a (0-1) categorical variable, which is convenient for us to further use the machine learning model for evaluation. Here, the parameters of the short-term moving average (SMA1) and long-term moving average (SMA2) are set to 10 and 60 respectively. The settings of the two have a certain degree of arbitrariness. The choice of parameters will affect the subsequent results, so ideally it is required Perform parameter optimization to find the optimal value.
The linear correlation values between the current features open, high, low, close, volumes and the target variable are very small. Maybe they are not ideal predictive feature variables, so feature construction and selection are needed below.
To facilitate analysis, the following introduces a feature matrix with technical indicators as features, which specifically include the following technical indicators:
Moving average: Moving average indicates the trend of price movement by reducing noise.
Stochastic Oscillator %K and %D: Stochastic Oscillator is a momentum indicator that compares the closing price of a particular security with the price range within a certain period of time. %K and %D are slow and fast indicators respectively.
Relative Strength Index (RSI): A momentum indicator that measures the magnitude of recent price changes to assess the overbought or oversold conditions of stocks or other assets.
Rate of Change (ROC): Momentum oscillator, which measures the percentage change between the current price and the past price in n periods. The higher the ROC value, the more likely to be overbought, and the lower the ROC value, the more likely it is to be oversold.
Momentum (MOM): The speed at which the price or volume of a security accelerates; the speed at which prices change.
In the figure above, you can see that a set of features that are clearly linearly related are created as a result of feature engineering. If the basic data set feature is used in the feature matrix, it is likely to have little or no impact on the change of the target variable. On the other hand, the newly created feature has a fairly wide range of correlation values, which is quite important; the correlation with the target variable (trading signal) is not particularly high.
The following uses commonly used machine learning algorithms to fit and evaluate the data respectively.
For comparison, firstly we use baseline features only to fit the model. We can see that the test scores are not very pretty, which indicates that only using baseline features related to one asset is not suitable for predicting accurate asset directionality.
We can see that compared with the baseline model, the accuracy score has been significantly improved.
The performance of linear discriminant analysis is surprising, not only on the training set, but also in cross-validation. It is also one of the fastest methods, making it one of the most effective methods for large data sets.
Among the LDA models with higher scores, there is no doubt that the advanced models have higher score, like GBM, XGB, RF.
Compared with the supervised learning model, kNN and GaussianNB unsupervised models perform slightly worse.