search

lancopku / CGM

Code for IJCAI 2021 main conference paper "Long-term, Short-term and Sudden Event: Trading Volume Movement Prediction with Graph-based Multi-view Modeling"
22 stars 11 forks source link

question regarding missing data in prediction phase #1

Open victorxie996 opened 2 years ago

victorxie996 commented 2 years ago

Hi!

Appreciate for sharing such greate work!

My concern regarding to the pratical usability of this whole pipeline is: if you select a stock universe i.e., stocks in the CSI300 index, or just simply select n stocks (doesn't really matter), to train the model, then how is the model gonna functioning well (do the right prediction) if some stocks quit the market, or stopped trading for a short period? From my understanding the vertices and edges cannot be missing, or nan?

zhao1iang commented 2 years ago

The stopped trading stocks may negatively affect the performance of its neighbor stocks, because those neighbors cannot gain useful information from this stock. If a stock its all the neighbor stocks stopped trading in prediction phase, the CGM will degenerate into a lstm-like model which only take time series data into consideration for prediction.