Use of time-varying prediction histories

Determine how the time-varying sequence of rolling LSTM predictions for an individual cryptocurrency asset can be used to improve further cryptocurrency-death predictions.

Background

The authors apply the LSTM classifier to every available sliding window for an asset and observe that the resulting prediction curve can change over time. In particular, the changing prediction trajectory for the asset hbar is described as a surprising result. The paper leaves unresolved how this temporal prediction history could be incorporated into subsequent prediction or early-warning procedures.