Concetti Chiave
DoubleAdapt is an end-to-end framework that adapts both the data and the model to cope with distribution shifts in the online environment for stock trend forecasting.
Sintesi
The paper proposes DoubleAdapt, a meta-learning approach to incremental learning (IL) for stock trend forecasting. The key insights are:
Data Adaptation: DoubleAdapt introduces a data adapter that transforms the features and labels of the incremental data and the test data to mitigate the effects of distribution shifts. The data adapter contains a multi-head feature adaptation layer and a multi-head label adaptation layer.
Model Adaptation: DoubleAdapt introduces a model adapter that provides a good initialization of model parameters for each IL task and then adapts the initial parameters to fit the incremental data.
DoubleAdapt casts each IL task as a meta-learning task and optimizes the data adapter and the model adapter to minimize the forecast error on the adapted test data. Experiments on real-world stock datasets demonstrate that DoubleAdapt achieves state-of-the-art predictive performance and shows considerable efficiency compared to rolling retraining methods.
Statistiche
Stock price change rate of the next day is used as the stock price trend label.
The feature vector of each stock includes 360 technical indicators looking back 60 days.