# Should I use Monte Carlo or a classifier for this Decision Making problem?

I want to build a model to support decision making for loan insurance proposal.

There are three actors in the problem: a bank, a loaner applicant (someone who ask for a loan) and a counselor. The counselor studies the loaner application and if it has a good profile it will propose to him loan from banks that fits his profile. Then the application is sent to the bank but the bank could refuse the applicant (based on criteria we don't know).

The counselor has also to decide whether or not he will propose to the loaner applicant a loan insurance.

The risk is that some banks reject loan applicant who accepts a loan insurance and other banks accept more applicants with a loan insurance. But there aren't rules regarding banks since some banks accept or reject applicants with loan insurance according of the type of acquisition applicants want with their loan for example.

Thus, the profile of the applicant can matter in their rejection from banks but all criteria influencing the decision are quite uncertain.

I've researched online and found several scholarly articles on using Monte Carlo for decision making. Should I use Monte Carlo or a simple classifier for this Decision Making problem ?

I saw that Monte Carlo (possibly Monte Carlo Tree Search) can be used in Decision Making and it is good when there is uncertainty. But it seems that it would forecast by producing some strategy (after running a lot of simulations) but what I want is an outcome based on both the profile of the loaner applicant and the bank knowing that criteria from banks (to accept loaner applicant from could change every six months. And I would have too model banks which seems quite difficult.

A classifier seems to me to not really fit the problem. I am not really sure. Actually, I don't see how a classifier like a decision tree, for example, would work here. Because I have to predict decision of the counselor to propose or not based on the decision of banks (and I don't know their criteria) to refuse or accept applicants who were proposed loan insurance and accepted it.

The data I have is former applicants profile who were sent to banks and if they were accepted or not by the bank, if they wanted a loan insurance or not and the type of acquisition they wanted to make with their loan.

I am new to Decision Making. Thank you!

A classifier seems to me to not really fit the problem. I am not really sure. Actually, I don't see how a classifier like a decision tree, for example, would work here. Because I have to predict decision of the counselor to propose or not based on the decision of banks (and I don't know their criteria) to refuse or accept applicants who were proposed loan insurance and accepted it.

The data I have is former applicants profile who were sent to banks and if they were accepted or not by the bank, if they wanted a loan insurance or not and the type of acquisition they wanted to make with their loan.

Why does this seem to you like something where a classifier wouldn't fit? Unless I'm missing something, it sounds like a prototypical example of a classification problem to me.

You have:

• Input features (applicants' profile)
• A clear (binary?) prediction target: propose or don't propose (equivalent to predicting whether or not the bank would accept, because you'll always want to propose if the bank would accept, and never propose if the bank wouldn't accept).
• Old training data containing both the input features and the matching prediction targets.

Approaches like Monte-Carlo Tree Search can only be used if you have a forward model or simulator. In your setting, you could view the features (applicants' profile) as a "game state", and model the problem as a game with two actions (propose or not propose). However, you don't have a forward model (a function that, given a current state and action, generates a possible reward and subsequent state).

In applications where MCTS is often used (such as games), you do have such a forward model: for a game like Go or chess, you can easily program the game's rules, program how you transition from one state into another when you select an action, etc. This does not appear to be the case for you.