SVM's SVC classifier was the initial machine learning algorithm that was evaluated. It was created using the following key parameters to determine both the algorithm and that training data:
short_window = 4
long_window = 100
DateOffset = 3 monthsThe overall accuracy score of this SVC model with these parameters was 0.55.
Based on the comparison of cumulative returns between the actual returns of the ETF and the returns using the SVC model predictions, predictions using the SVC model returned a slight out-performance (SVC 1.52 vs actual 1.386)
For the first alteration, the short and long window periods were kept at their original values of 4 and 100, respectively, and the training period adusted to 24 months.
short_window = 4
long_window = 100
DateOffset = 24 monthsThe cumulative returns that resulted from chaning the periods associated with the training and testing data can be seen in the following image.
By changing the training window, both the model accuracy score, and the cumlative performace of the strategy did improve. The accuracy score was reported at 0.56 and the cumulative preformace at 1.634%.
Overall, increasing the training window imporved the performance of the model.
For this version up updating the SVM machine learning model, the short window was adjusted to 50 periods, and the long window to 200 periods. The DateOffset parameter stayed the same at 3 months.
short_window = 50
long_window = 200
DateOffset = 3 monthsAn image of the cumulative retrun data for this version of the model is seen below:
The accuracy score of the SVC model using the above parameters was roughly in-line with the previous models 0.54.
In terms of cumulative performance, the SVM strategy actually underperformed using these parameters, returning only 1.437 versus the actual return value of 1.565.
Three months was clearly not enough time to train the SVC model with a short window of 50 periods and a long window of 200.
It is interesting to note that if the DateOffset period was adjusted to 24 month in order to accomodate the longer short and long window periods, the performance does improve drastically.
short_window = 50
long_window = 200
DateOffset = 24 monthsAs seen in the following image, the model actually returns 1.83% versus actual returns of 1.398%.
Not surprisingly, the longer windows allows for better pattern recognition with regard to the behavior of the ETF over time, and the longer training period gives the model additional opportunity to learn based on the data.
The original parameters, as seen below, were evaluated using the AdaBoost Classifier model.
short_window = 4
long_window = 100
DateOffset = 3 monthsThe overall accuracy of the AdaBoost classifier model was in line with the SVC's at 0.55
The AdaBoost Classifier returned an on overall performance of 1.571 as seen in the following image.
Overall, the AdaBoost classifier model outperformed the baseline SVC model, returning 1.57 versus 1.52.
It appears that the AdaBoost classifier model is better at predicting the signals for the short-long algorithm with the original parameters than was the SVC model.
Overall, it appears that the AdaBoost classifier is a better fit for the long-short algorithm trading strategy presented in this Challenge (1.57 vs 1.52).
However, given the performance data that was returned by fine-tuning the parameters for the SVC model, it would be beneficial to evaluate the AdaBoost classifier under those same conditions to confirm.