AI for algorithmic trading: rethinking bars, labeling, and stationarity
In a series of articles I was applying a very straightforward approach to forecast financial time series: take the whole dataset, using a sliding window approach generate X and Y, split it into historical and out-of-sample data, train some machine learning models to map X to Y and backtest simple long-short strategy. But as I showed in the last blog post I started to realize that pipeline for “normal” static data like images, text, audio, tabular data or even less chaotic time series can’t be used for financial time series analysis.