we should:

1. train with a cutoff date at ~2006

2. simulate information flow (financial data, news, earnings, ...) day by day

3. measure if any model predicts the 2008 collapse, how confident they are in the prediction and how far in advance