As I understand the tournament, it is in everybody’s interest for the Training and Validation data to be representative of the patterns that our models will then need to predict for the Live data.
I am uneasy however at the idea of training our models on monthly eras and then apply them to weekly eras, where I believe it’s possible that regularities detected at a month level may vanish with a time horizon 4 times shorter (or the other way around). Has this issue been discussed before?
Were these possibilities considered (or even tested):
- Training/Validating on weekly eras (thus increasing the dataset size)
- Predicting on monthly eras (thus being able to trade only monthly)
I apologize if this has been discussed before or I’m missing a basic point!