Many a times, for use cases such as forecasting a stock’s trend, it’s more beneficial to detect if the stock is going up or down by at least some percent value higher, and then it becomes a classical machine learning problem, especially you want to add in information such as ‘news sentiments’ or ‘expert opinions’ as features.
There are three ways to engineer time series based features:
- Build different time series forecasting models and simply use forecasted values as features
- Delve more into signal processing, such as computing peaks and troughs, and design features such as average slope since last trough, average trough to peak distance.
- Understand the domain, and most important technical indicators trades use, and convert them as features.