1
$\begingroup$

I'm trying to model a future balance for any time t in a list of bank accounts.

I have historical balance & transactional data, and other factor data that affects the balance of an account (such as interest payments, fees, etc) that I'll try to fit in the model as well.

My question is, what model/kind of model should I use to estimate a future value of the Balance variable?

A GARCH/ARCH model?

Would a multivariate regression be sufficient to accurately predict future values?

I don't have much background experience in econometric analyses, so I would appreciate if you could just point me to the right direction and I'll research and dig deeper into it.

Improve this question
$\endgroup$

2 Answers 2

Reset to default
0
$\begingroup$

The common industry approach is to model pools of accounts, grouped by vintage and other characteristics. It's often assumed that the lagged rates drive the decay rates. You can try to do a fancy analysis like GARCH, but I guarantee that it won't work better than simple approaches. There's too much noise on account level analysis, you'll abandon it eventually

Improve this answer
$\endgroup$
5
  • $\begingroup$ A colleague suggested I should go for a mean-reversion model with a stochastic element. I can model my best estimate of the balance in the future using certain parameters (contributions, interest rates, inflation, investment returns, etc.). Using that model against my historical data, calculate the error terms, and that would be my stochastic element (the standard deviation of the errors). I should add I plan to do this on a client level analysis. $\endgroup$
    – MrPmosh
    Commented Feb 16, 2018 at 19:25
  • $\begingroup$ @NZambrano, Are working with non maturity deposits? what kind of a product you deal with? there's very few studies published on the subj. you can lookup deventer's papers $\endgroup$
    – Aksakal
    Commented Feb 16, 2018 at 19:38
  • $\begingroup$ I'm working with savings and checking accounts only. $\endgroup$
    – MrPmosh
    Commented Feb 19, 2018 at 17:15
  • $\begingroup$ Thank you so much for pointing me to the Deventer's Papers. I found a paper exactly on what I'm trying to accomplish. Non-Maturity Deposit Modeling in the Framework of Asset Liability Management $\endgroup$
    – MrPmosh
    Commented Feb 21, 2018 at 18:17
  • $\begingroup$ Unfortunately, the publisher of the journal from which you cited the paper is in Beall's list :( it doesn't automatically invalidate the paper's content, but it's just something to be aware of $\endgroup$
    – Aksakal
    Commented Feb 21, 2018 at 20:47
0
$\begingroup$

I would consider a daily analysis using the historical daily balance data. Cash infusion and withdrawal activity often are systematic behavior/habits/ by day-of-the-week , day-of-the-month , holiday effects , week-of-the-month effects, month-of-the-year effects et al. The approach would be similar to Is is possible to fit discrete data to a continuous distribution, and use this to simulate discrete outcomes? predicting # of people expected for daily lunch and in terms of daily demand for cash here http://autobox.com/cms/index.php/afs-university/intro-to-forecasting/doc_download/53-capabilities-presentation slide 50 plus.

I would just try to form a daily predictive model for each separate account to build up useful examples. If the data set is long enough interest rates might be meaningful .

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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