Predicting 12-Month Recovery Rates for Non-Performing Loan Portfolios

A debt collection agency needed a more consistent method for valuing non-performing loan portfolios before auctions. SoftwareHut developed an ARIMA-based model that estimated the percentage of a portfolio’s face value likely to be recovered within 12 months and presented the results in Excel reports.

The client was a debt collection agency specialising in debts related to communication and transport services.

The company purchased non-performing loan portfolios through auctions and needed to estimate how much of each portfolio’s face value could realistically be recovered before placing a bid.

Portfolio sellers provided large volumes of data before auctions, but much of it had limited forecasting value.

The bidding team relied mainly on the judgement of experienced employees. This increased the risk of overpaying for low-potential portfolios or rejecting profitable opportunities.

The client needed a repeatable method for estimating expected recoveries based on historical data.

SoftwareHut interviewed employees responsible for auction bids and analysed data from previous portfolio sales.

The project focused on variables that could be used consistently across portfolios, including:

• the period in which obligations arose,
• the age of the debt,
• the historical performance of portfolios sold by a particular partner.

After evaluating several approaches, the team selected an Autoregressive Integrated Moving Average (ARIMA) model.

The model predicted the percentage of a portfolio’s face value expected to be recovered within 12 months.

It was implemented in SoftwareHut’s AI Farm system. Users could enter portfolio data, generate a valuation and download the report as an Excel file

The model provided an analytical reference point for decisions that had previously relied mainly on expert judgement.

It helped the client identify portfolios with stronger recovery potential and avoid auctions involving less promising assets.

The reported model accuracy exceeded 17%. The client considered the model more accurate than expert estimates, although the calculation method and benchmark were not specified.