Hi birinhos, I think you have misunderstood a bit the scope of this work. It’s related not to re-training a machine learning model that predicts something in the future but to calibration of financial models for option pricing to the current market state. There is a huge difference between these two tasks.

Concerning multitask learning paper, at the start of section 3, I’ve mentioned, that we use multivariate time series for forecasting (high, open, low, close prices and volume).