I'm working on a forecasting problem, and I'm not sure if the data requires any transformation because the unit of time is dynamic.
I'm working with an education data set where I have data on students attempting questions. We however allow re-attempting of questions, so therefore, the questions' data is captured for each attempt for each question for each student. So plotting the time series of say time taken to solve questions per attempt yields a time-series, but the unit of time there would be attempt and not time-day,hour,month,etc.
Is this ok- or does it require any special treatment? I'm trying to forecast the accuracy of the next attempt for a student, can I retain attempt based data or should I take the timestamp values- I'm worried if I take the timestamp values, there will be a lot of empty regions/no data because attempts can be taken quickly or after a long time- entirely upto the student. Will this affect the accuracy of the forecast? Is this a special case that needs to be handled in a different way? If I change the time-units from 'attempt' to say 'hour', I'll have to figure out a way to transform, say average/max/etc where there were >1 attempts by a student within an hour- any way to avoid this?
I'm unsure how to handle missing values in such cases- would appreciate some guidance on how missing values impact time-series model predictions and say if I simply place a 0 at missing places, wouldn't that mis-interpret the data- implying a poor accuracy from the student, instead of saying-no attempt.
Should I have 1/0- attempt, no_attempt sent as a time-series also into the model- will it be worth doing this?
I'm planning to use aws-forecast, would appreciate if anyone has insight on the algorithms/recipes (currently thinking MDN) that would be best for me.