What cross validation is: a complete guide to reconciling values across documents

Cross validation takes the same field from the several documents that make up one case, compares the values against each other, decides whether they agree, and passes only the fields that disagree to a person.

Why one document never finishes the work

A review ends in reconciliation.

Watch a financial review operation and the task the operator spends longest on is not reading. It is laying the merchant application, the business registration certificate, the passbook copy, and the corporate registry certificate side by side and confirming that the representative's name is the same, that the account number matches, that the business registration number agrees.

Measurements put three to eight minutes per case on document reconciliation and validation alone. Where an average case arrives with six document types, that work takes a large share of the total time. However accurately a single document is read, automation stops halfway while this step remains.

What cross validation actually is: how it differs from field validation

Field validation and cross validation across three axes

First, the target differs. Field validation works on values inside one document, checking digit counts and date formats. Cross validation works on the same field scattered across several documents.

Second, the method of judgment differs. Field validation compares against a defined format. Cross validation compares documents against each other, so there is no separate correct answer. Only the fact of disagreement is established; which one is right is for a person to judge.

Third, what it exposes differs. Field validation catches recognition errors and format errors. Cross validation exposes discrepancies in the documents themselves, including possible forgery. In one real case, comparing 46 fields across several documents automatically separated 45 agreements from a single mismatch.

The two are not alternatives. Values that both pass format checks can still disagree between documents, so both are needed.

Five conditions that make cross validation work

Five conditions that hold in real operations

First, documents have to be grouped by case. Handling them as loose pages leaves nothing to compare against. A step that splits the received file and determines which case each piece belongs to has to come first.

Second, field correspondences have to be defined. Someone has to establish in advance that the representative's name on the application and the representative director on the registry are the same field.

Third, notation differences have to be absorbed. One document commonly uses the full legal name where another uses an abbreviation. There are real cases where standardising bank names was necessary.

Fourth, temporal ordering has to be handled. For fields where sequence carries meaning, such as transaction history and filing timestamps, it is not only whether the values match but whether the order holds.

Fifth, the character of a mismatch has to be distinguished. A mismatch caused by recognition error and a genuine discrepancy between documents call for different responses. Presenting the source position alongside lets a reviewer separate the two quickly.

How cross validation is applied in practice

Draw the reconciliation matrix first

The first deliverable of the design is a reconciliation matrix. Which fields are compared between which documents, which document is treated as the reference, and what happens on a mismatch, laid out as a table.

With that table in place, the review screen composes itself: each field showing the value from every document side by side, with the mismatched cell marked.

Set the mismatch handling criteria with the business

Not every mismatch is grounds for rejection. A difference in how a floor is written in an address can pass; a difference in an account number has to be checked.

Agreeing with the business owner how much variance is tolerable field by field sharply reduces the number of cases routed to exceptions.

Choose the fields to reconcile on business criteria

There is no need to reconcile everything across every document. Selecting only the fields that actually inform the decision reduces processing load and makes each mismatch meaningful.

Financial reviews commonly select the representative's name, business registration number, account number, amount, and date. In one bank's relief automation, name, account number, bank name, contact number, claimed amount, and incident timestamp were compared across three document types, and among 46 fields the system automatically surfaced a single mismatch on the bank name. The full legal name and an abbreviation had arrived mixed together.

What that case shows is that how notation differences will be treated has to be settled in advance. Differences to be read as the same value should be standardised, and only genuine differences raised as mismatches, or unnecessary review piles up in front of the operator.

Cross validation in the Korean environment

Korean financial operations receive around six document types per case, a substantial share of them faxes or photographs. The same value is recognised at different quality across documents, which makes absorbing notation differences especially important.

In the public sector, frequent form revisions mean the same field can be named differently from document to document. Managing field correspondences as data and versioning them avoids rebuilding everything at each revision.

Frequently asked questions