Machine learning method to predict event date

Question

Let's say I have a big dataset consisting of variables including but not limited to the start/end date of loans, their notional amount, a loan prepayment indicator etc.

My goal is to create a model that will be trained on past data in order to predict the prepayment date of current loans and I was wondering which ML method would be most suitable for this case. My first thought was to handle this as a classification problem, using interval dates to predict a prepayment date interval, but I believe that there should be a more robust & sophisticated approach to this.

Answer 1

You can both utilize regression and classification models in here. Also I suggest log transform your financial features (eg: debt, last credit risk, average loan amount etc) when using algorithms except decision tree based models (since they are not affected by monotonic transformations).

Classification

You can construct a model such that predicts if a customer would pay "x" days after payment day or not. In this model you need to find best "x" by comparing results of different x day models. Also you need to find optimal confidence range in each model in comparison.

Regression

In this setting, you will predict how much day a customer will pay after payment date. Again you should obtain confidence intervals such that you calculate 3 days for a customer which means she will pay 3 days after (it can be earlier for negative predictions). However you know that your model is powerful on +2 -2 day interval. Then you should say this particular customer will pay in 1 or 5 days after payment date.

Algorithms

I suggest xgboost, lightgbm and RNNs. But its not so clear maybe you have a linear space and it converges fast with SVMs, line fits etc. Yet they are pretty good, kaggle winner learners.

I hope it helps. I thinks it will be a good starting point.