54%Real bank data from an open research dataset
Four statement packs, one history
higher average risk forecast when the statements were joined as they came, against the full account record
- What was observed
- Current accounts with every transaction and running balance, and their loans, reviewed each quarter: 1,733 reviews. At each, two years of statements arrived as four packs that overlapped by one or two months and left one month uncovered.
- What a generic approach says
- Join the packs as they come: overlapping months counted twice, the missing month read as a quiet one. The average forecast of a missed repayment in the next six months rose from 1.71% to 2.64%, when 1.68% happened.
- What the engine read
- Reconciled the packs into one history: 39,306 duplicated transactions counted once, every missing month reported and never filled, and every covered month equal to the bank's full record at all 1,733 reviews. Forecasts from it averaged 1.73% and ranked customers as well as forecasts from the full record (AUC 0.714 against 0.716).
- What it meant for the decision
- Older statements can be used at renewal without overstating risk. Here the two year history also ranked customers better than the latest six months (AUC 0.714 against 0.648), but on only 12 loans that went on to miss, so that gain is measured on your own renewals before anything changes.