I face a data which records the default rate of loans by cohort.e.g. My company currently hold a portfolio comprising personal loans whose starting date was in a range from 2014 Jul till now. The loans were divided into each monthly cohort like loans drawn in 2014 Jul, in 2014 Aug, ... in 2015 Feb,... in 2017 May etc.
Therefore each cohort had a fixed loan population but the defaulted amount was increasing as time goes. So we record the defaulted rate (defaulted amount / total loan amount) of each cohort on a temporal axis (daily basis.).
I create a sample data which is isomorphic to my real dataset. The first column is the cohort and the other columns are for dates on the time axis.
My task is to create a model to predict the defaulted rate for each cohort in a time window like 360-day, i.e. how about the defaulted rates of each cohort after 360 days given its input into the model?
I have tried some ARIMA models but the AR order is too large, e.g. AR(100) and the computation take long time to run.
I wanna try a machine learning method which allow me to create a dataset not only with the time labels but also some other features which can contribute to the prediction of the defaulted rate.
Please advice how can I do that or please recommend some papers/textbooks for reference. I know some concepts about the sequential learning but not to the executable level.