Allow multiple rows of lagged data to produce forecasts. Currently, columns are shifted in situ, and this is confusing when identifying which features have been lagged and which have not. We have defined a convention for variable names, e.g., close[1] is the previous close, but we want to make this explicit in the data frame itself, so that you can clearly see when multiple variables have been lagged, e.g., close[1], close[2], etc. This also applies for future forecasts, e.g., target[+2] or target[+10]. This capability will allow easy transformation for Keras sequence-to-sequence learning in the future. Expect a new version around the end of August, along with other fixes.
mrconway
commented
6 years ago
Allow multiple rows of lagged data to produce forecasts. Currently, columns are shifted in situ, and this is confusing when identifying which features have been lagged and which have not. We have defined a convention for variable names, e.g., close[1] is the previous close, but we want to make this explicit in the data frame itself, so that you can clearly see when multiple variables have been lagged, e.g., close[1], close[2], etc. This also applies for future forecasts, e.g., target[+2] or target[+10]. This capability will allow easy transformation for Keras sequence-to-sequence learning in the future. Expect a new version around the end of August, along with other fixes.