Effect of the "Balance of all Accounts" on Defaults

[Jogi]

My question concernes the "Balance of all Accounts", which is given by the field tot_cur_bal in Lending Clubs data set.

I downloaded the whole available dataset of all financed loans ans plotted the default rate within the quantiles of the total balance.
To be more precise, I define a quantiles' default rate as the number of defaults within a certain quantile over the total number of observations

And here comes my intuition:
The more balance is reported on the credit report, the more debt the borrower has.
So a high balance on the credit card accounts or on mortgage accounts lets a large part of the borrowes income be used to finance his debt.
Therefore less money is left for additional new debt and the borrower has fewer financial flexibility.

Consequently, a high "Balance of all Accounts" should have a negative impact on the borrowers PD.

But the plot say something different:
The more Balance a borrower has (which means that he is in a larger quantile) the lower his PD?
https://forum.lendacademy.com/proxy.php?request=http%3A%2F%2FBalance.png&hash=a4488b924306a057f6de098fe25eb746" alt="" class="bbc_img" />
For me this seems counterintuitive...

Do you have any guess?

[rawraw]

From what I recall, borrowers with no or low  debt underperformed those with debt in back tests. The common explanation is that they have more experience in budgeting and meeting payments. I personally like to screen for refinance requests that are roughly the size of the outstanding balances. I find that performed the best

Sent from my SAMSUNG-SM-G935A using Tapatalk

[TravelingPennies]

@rawraw: Very nice explanation! Do you have any links or sources to such studies?

[Fred93]

I think you need to look at the absolute amounts, rather than drop them into some arbitrary tiles, and also perhaps look at more than one variable at a time.

I suspect that what is going on here is that many of the folks with very low balances are folks who don't use credit much, either they have little EXPERIENCE with credit, or they have not been allowed to use credit much because of some variables we can't observe with this limited view (bad FICO, or no job, or whatever).

Therefore, while looking at nothing else, credit balance is acting as a proxy for more traditional credit variables that would be more intuitive to you.

If on the other hand, you took a set of people with other credit background common, ie all have jobs, own homes, credit score in a limited range perhaps, etc, and THEN looked at default rate vs credit balance, you might get a different result.  I suspect, based on the analysis I've done in the past, that you will still find the very low balance people have higher default rates.  I don't claim to understand that really.  I just attribute it to lack of credit experience.   Similar things happen when you look at the #accounts variables.

[TravelingPennies]

https://forum.lendacademy.com/index.php?topic=4484.msg41244#msg88888888Quote"> from: Jogi on July 03, 2017, 02:08:08 PM

[AnilG]

You might want to consider following suggestions based on my insights gained through analyzing Lending Club and Prosper data for past few years

Instead of analyzing all the available data, pick a quarterly/yearly vintage and use that for your analysis. This will alleviate impact of any changes that occurred over time. Preferably pick a vintage that is few years old so that you have larger fraction of loans either paid off or defaulted. Also ignore loans that haven't reached final status of fully paid or charged off in your analysis. The transient status loans can bias your conclusions.

While your intuition is correct that defaults should go up with tot_cur_bal, there are confounding factors that may influence the impact of tot_cur_bal on defaults. You need to identify these confounding attributes. For example, the capacity to pay (income and/or DTI) may have more influence on how much tot_cur_bal a borrower can handle. You may also find that tot_cur_bal might be connected to open credit lines or total credit lines, etc.

Instead of jumping directly into reviewing defaults with tot_cur_bal, do an exploratory analysis of all available attributes and identify related attributes. The loan and borrower attributes haven't stayed the same across all the periods and you need to identify those time period and whether and how they impact your analysis. You will be surprised how many people claimed that large loan amounts (>25,000) have lower defaults few years ago and same is being suggested with >35000 loan amount now which logically makes no sense.