Team members: Tian Qi, Jessica W, Divya Bhargavi, Mahsa Ashabi

The goal of this class project is to use an time series modeling to predict Canadian monthly bankruptcy rates. With a collection of monthly bankruptcy rates data from January 1987 to December 2014, we aim to predict the bankruptcy rates for the following 36-month period. There are also corresponding data on unemployment rates, population rates, and Housing Price Index that could be used to predict forecast bankruptcy rates. We used both univariate (SARIMA, SMOOTHING) and multivariate time series model (SARIMAX, VAR).

We first did a basic EDA to visualize the series explore the underlying relationships between them. The validation data is the subset data from 2011 to 2014, we then used the validation data to tune parameters for each model. Then we refit the model on whole data we have, made the prediction for the next 3 years (2015-2017) and compared it with the true value. We used RMSE as the loss function during this procedure.

The whole training time of four models was less than 3 mins. Our final model is the emsemble of four models, which is the average of their predictions, this gives RMSE 0.19 on test data. We finally ranked top 3 among 22 teams in the class competition.

Check the Data folder for train and test data

Check the Rcode folder for SARIMA and VAR implementation code