Weak Factor Alpha Strategy 1(中文)
這個策略的想法源自於分類器中的弱學習器，首先生成技術指標並找到最佳參數，再將這些技術指標視為弱因子（特徵），並使用CatBoost分類器預測未來趨勢的相對強弱，通過買入相對最強的組合，賣出相對最弱的組合，形成Alpha策略。本篇主要在探討如何生成更好的特徵，下篇則會說明如何運用分類器建構Alpha策略。
The idea of this strategy comes from the weak learner in the classifier, which first generates technical indicators and then finds the best parameters through the algorithm and sees these technical indicators as weak factors. Finally uses these weak factors (features) as input for training CatBoost classifier with relative strength. By long the relatively strongest prediction combination, short the relatively weakest prediction combination, finally implement the alpha strategy.
In this article, we focusing on how we improve the method from past literature and use more reasonable approaches to construct the strategy. In the feature selection part, We use distance correlation as our target function. Compared to the traditional method of Pearson correlation, distance correlation can describe the non-linear relationship which supplies more information. Also, we use the k-means clustering model to avoid selecting the outlier as the fitted parameter.
Our data is SPY 500 symbols from 2015 to 2022/11 daily Open, High, Low, Close, Volume. We apply python package TuneTa to generate and select technical indicators. By this package, it encompass not only traditional technical indicators, but also unique indicators from TradingView.
Figure 1 shows the difference between Pearson’s correlation and distance correlation. If Pearson’s correlation is 0, there are not necessary independent(might be non-linear). However, if distance correlation is 0, we can say that they are independent.
Figure 2: shows the final outcome of the indicators, we can see that every technical indicator have an unique and fitted parameter.
In summary, we spend more process on generate diverse features rather than using unselected features. Also, by apply distance correlation as our target function and K-Means clustering to find suitable parameters for each indicator, making the feature selection more rational. In next report, we will describe how we use these features as an input to training ML classification model and construct an alpha strategy.
Contact us at info@gammaparadigm.com for the full report.
David Wang
Quantitative Research Intern, Quantitative Finance, Gamma Paradigm
