Im thinking of applying machine Learning algorithms on a data set of a Portfolio of credit contracts. The dataset is huge. I got around 22 Million contracts with around 100 Million data rows from a 2000 to 2015 (Panel Data) and up to 30 individual characteristic variables.

With this data set I want to estimate the so called "prepayment risk". Generally speaking, I have a regression problem where I want to estimate the parameters in order to forecast the probability of prepaying a loan within a group of clients

In the literature such models are basically estimated with a logistic Regression because the dependend variable is usually discretized. With some extended things were also modelled in a survival Analysis modell.

The goal of my research should be, how or if neural networks can improve the estimation compared to a logistic regression. Hence I want to get a step further and estimate my parameters with a deep learning ANN. Unfortunately I don't have any experience with machine learning, but I'm a graduate student of econometrics.

Now I want to read some opinions from some "experts". Do you think this is a good idea for a master thesis? I'm afraid that in the neural Network Approach I might get stucked at some point.

Could you give me some advise about some good papers and books? Maybe some online lectures which are worth to watch?

Which programm language should be preferred? I'm pretty good in MATLAB and SAS, but I think R or Python should be the way to go.

  • $\begingroup$ When considering or comparing techniques or languages, try a smaller subset of the data to determine which seems best for you - after that, then try the entire data set for a problem solution. $\endgroup$ – James Phillips Jun 1 '17 at 9:37

Some considerations in random order:

1) I'm not sure that would make a proper thesis, but it's mainly my opinion. A more interesting thesis from my point of view would be to evaluate more learning algorithms (random forests, svm, neural nets...) on the task and see which one performs better. Picking neural networks a priori doesn't seem fair to me.

2) if you are not a machine learning practitioner, you have a long journey ahead of you. A good start might be the course that professor Ng teaches on the coursera site (which is taught in Octave, an open source version of MATLAB).

3) matlab and SAS knowledge is fine for your needs, although python and R are mainstream at the moment.

4) you don't necessarily need a deep neural network for your task. They are essential in image recognition tasks to obtain good results, but with numerical data it is often the case that a shallow network (1-2 hidden layers) is enough.

  • $\begingroup$ Thanks for the fast Response! Your idea about comparing different techniques sounds interesting, but I guess it would go to far. I still want to compare mainly one technique to a logistic regression. Hmm I'm getting confused. In what sense it is not a regression problem? The Goal is to estimate the effect of f.e. in change in interest rates on the propability of prepayment? That is a classic regression problem from an econometric point of view? $\endgroup$ – Kosta S. Jun 1 '17 at 9:46
  • $\begingroup$ Oh alright then, I misread. Gonna edit and remove that point. $\endgroup$ – mp85 Jun 1 '17 at 9:50
  • $\begingroup$ As far as the thesis type, I come from a CS background, so my point of view is different than yours. Might be the case that for your purposes comparing only a logistic regression to neural networks is perfectly fine. I think your professor might answer in a more satisfying way than me to that question. $\endgroup$ – mp85 Jun 1 '17 at 9:54
  • $\begingroup$ agree deep NN is mainly for image processing. $\endgroup$ – seanv507 Jun 1 '17 at 13:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.