Trend Following with Logistic Regression

In this post, we'll cover a pragmatic logistic regression classifier to mimic a trend following strategy for the S&P 500 ETF, SPY. The pipeline takes in daily prices for SPY along with several SPDR sector ETFs and macro ETFs for gold, Yen, Swiss Franc etc. Once all Open, High, Low, Close, and Volume data has been received from yfinance, a feature space (the set of columns if thinking in a spreadsheets world) is built using select indicators included in TA-lib. The features are then reduced to 4 n-components with Principal Component Analysis; the model is trained on these n principal components, using ground truth labels generated by a brute force optimized dual moving average crossover. Initially, I opted to use the default boundary of .5 for the binary classification. On visual inspection, there is a gap in this logic -- as the classifier appears exceedingly optimistic (subjective).