Data warehousing in lending helps decide whether to rebuild a scorecard

At the monthly risk committee, the outcome report is open in front of the head of credit risk. The score’s Gini on last quarter’s approvals has dropped, and someone proposes rebuilding the model. You have the policy, the rules and the score in front of you. The report was built by joining this month’s defaults to the current loan and customer tables, so the performance figure and the information alongside it come from different points in the lending process. Before agreeing to a rebuild, what are you actually measuring: a weaker model, or a report reading information the decision never saw?

Use the record available at decision time

A lower Gini can reflect weaker score performance, a change in the book’s mix or a mismatch in the information used for review. The Default of Credit Card Clients dataset covers 30,000 card holders in Taiwan in 2005. The score was fitted only on information known at the end of June and judged against October defaults, then the same unchanged score was fed the same customers’ records from the end of September.

June data, score Gini 0.387, flagged customers who defaulted 54.3%, defaulters caught 28.2%. September record, score Gini 0.484, flagged customers who defaulted 69.4%, defaulters caught 32.5%.
Score and rule results using decision time and later records

The later record made the score appear stronger and the rule appear to catch more defaulters. The customers, score and rule were unchanged. Only the information used in the review changed. These results show how timing can affect a measurement in one public credit-card dataset and one score. Portfolio expectations require assessment against each lender’s own data.

Joining defaults to today’s loan and customer records seems reasonable because those records are available now. But the report is then judging an earlier decision with information that arrived later. That makes the model look better than it was when the lender used it. The Gini coefficient must be read against the information the decision actually saw.

Separate score performance from portfolio mix

Even with the right decision-time record, a lower portfolio Gini can reflect weaker score performance or a change in the book’s mix. The score ranks some segments less well than others, so a change in the book’s mix can lower the aggregate Gini.

In the same test, the score ranked younger customers and customers with higher credit limits less well than other groups. When the share of higher-limit customers increased, the aggregate Gini fell even as the book’s default rate fell too.

Higher-limit shareGiniDefault rate
52%0.38722.1%
68%0.37620.1%
76%0.36619.1%
84%0.35318.1%

The score stayed the same. The book got safer, but the aggregate measure fell because more of the book came from a segment the score ranked less well. A portfolio Gini combines score performance with the mix of customers entering the book. Check both before proposing a rebuild.

Connect each decision to its later outcomes

A policy says what the lender intends to do. To see whether the rules worked, connect each outcome to the decision that caused it, using the information available at the time.

Keep each decision with its inputs, score, rule and strategy version. Preserve each source as delivered, then retain dated versions when customer or loan records change. Decision-time and event snapshots let the review use what was known at the decision or event, rather than a record corrected later.

Collections contacts, promises to pay and payment outcomes are part of that history too. They help connect the decision to what happened to the account, as described in this review of debt recovery tools and communication channels.

A data warehouse brings source records and outcomes together for analysis. That history lets the risk team test whether a change in results came from the decision, later information or a different mix of customers.

Put these checks into a decision flow

Create rules for credit, fraud or collections, then run them on live decisions.

Preserve decision history across warehouse layers

The layers keep the decision, every later version of the record and the outcome, each readable as it stood.

Sources, landing, vault, marts and outer serving area. Month ends are checked, approved, published and frozen. Each decision returns as a source with the inputs it read.
Lending warehouse layers from source records to live decisions

Sources first arrive in a landing area and remain as delivered. In the vault, hubs hold each business key from its first sighting, links connect keys and satellites keep every version, so each value received stays readable with the date it was loaded. The marts follow Kimball’s dimensional design: each fact table declares its grain and does not mix grains. Conformed dimensions let facts from separate tables be combined in one report. The vault keeps the history, and the dimensional marts built from it work together. A separate outer serving area can publish fixed values for live decisions while the analytical tables are being rebuilt.

Keep each fact at one grain. A decision row answers what happened at one decision. A month-end loan row answers what the account looked like at a close. A cohort row groups loans by origination period. Comparing a cohort at the same months on book is what makes a vintage comparison meaningful.

Close each month-end snapshot only after checking that balances roll forward, account records remain continuous across closes and currencies agree. Approve, publish and freeze the close before using it for official figures. A corrected file becomes a new load. If a published month-end record is wrong, restate it and retake later closes in order, approving each one. The history stays visible.

Use consistent records to decide whether to rebuild

Use the records that match the measure. Review first payment default and days past due against decision-time inputs and the relevant outcome observations. Read provisions from month-end snapshots. Measure cure rates over a defined period, and compare vintages at the same months on book. For calculating provisions, IFRS 9 starts the allowance at expected losses over 12 months and moves it to lifetime losses when credit risk has increased significantly since the loan was booked, so the lender has to keep the credit risk it saw at origination. Large banks are also held to the Basel Committee’s principles for risk data aggregation, including accuracy and timeliness, but those principles do not bind every lender.

Then cut Gini and rule hit rates by segment and check whether the book’s mix changed. Scorecard monitoring is useful only if it keeps those differences visible, rather than treating one portfolio figure as a diagnosis. If the point-in-time results still indicate weaker performance within comparable segments, replay the current and proposed strategy over past decisions. A champion challenger test lets the team inspect which decisions would change and compare those decisions with their observed outcomes.

Rebuild when the point-in-time, segment-level evidence still shows weaker score performance and the replay supports changing the strategy. If the movement reflects a changed mix, use it to guide review of the affected segment’s policy or cut-off, rather than rebuilding the score. The committee needs a reason for the change, supported by more than a lower number.

  1. Store every decision with the inputs it read, its score, the rule that fired and the strategy version in force at that moment.
  2. Keep every source as delivered. When a customer or loan record changes, add a dated version and preserve the earlier one.
  3. At each month end, check that balances roll forward, account records remain continuous across closes and currencies agree. Approve, publish and freeze the snapshot. Record corrections as new loads, and restate a published period through approved later closes.
  4. Build separate marts with one row per decision, outcome observation, loan at month end and origination cohort. Use the same customer, product and segment definitions across them.
  5. Use inputs available at decision time and relevant frozen month-end records as the comparison point for outcomes, keeping the review tied to the historical record.
  6. Compare Gini and rule hit rates by segment, then check whether the book’s mix changed. Replay the current and proposed strategy over past decisions, inspect which decisions change and compare their observed outcomes.
  7. Read provisions from month-end snapshots, cure rates over defined measurement periods and vintages at the same months on book. Rebuild the score only when weaker performance remains in comparable segments and the replay supports the change. If the book’s mix explains the movement, review the affected segment’s policy or cut-off and test that change instead.

In this practitioner’s view, AI now does much of the analysis a data scientist used to do, and faster. That makes the warehouse more necessary, not less, because it is the layer AI reads. It has to be built properly for regulators and to show how the data relates to the outcomes.

Questions about data warehousing in lending

What is a data warehouse used for in lending?

It connects company data and recorded outcomes to the decisions that caused them, then organizes that history for analysis. Keep the decision and outcome records linked so the risk team can assess the strategy against what followed.

Why does a lender need snapshots?

Snapshots preserve what was true at a decision, event or month end after operational records change. Use the snapshot that matches the decision or outcome being reviewed.

What is the difference between a data vault and a Kimball data mart?

A data vault preserves business keys, their relationships and successive versions of descriptive data. A dimensional mart presents facts at defined grains for analysis. They serve different purposes and can be used together.

How do you know whether a credit rule is working?

Store the rule and strategy version with each decision, then compare its outcomes against the information available at that decision. Review hit rates by segment and use consistent outcome records.

Why did our scorecard’s Gini drop?

The score may have weakened, the book’s segment mix may have changed, or the review may be using information that arrived after the decision. Compare point-in-time results by segment and check the mix before deciding what to change.

Which lending metrics need month-end snapshots?

Provisions should be read from month-end snapshots. First payment default and days past due should be tied to the relevant outcome records, cure rates to a defined measurement period and vintages to the same months on book.

Does AI make the data warehouse less necessary?

In a practitioner’s view, AI makes the warehouse more necessary, not less. AI now does much of the analysis a data scientist used to do, and faster. The warehouse is the layer AI reads, and it needs to be built properly for regulators and to show how the data relates to outcomes.

Put these checks into a decision flow

Create rules for credit, fraud or collections, then run them on live decisions.