# How ThumbPrint Works

[ThumbPrint](https://www.moneythumb.com/thumbprint/) is MoneyThumb's patented fraud detection engine for PDF documents. It analyzes a bank statement and returns a score indicating how likely it has been altered or fabricated, along with the specific factors behind the score.

## What does ThumbPrint Actually Detect?

[ThumbPrint](https://www.moneythumb.com/thumbprint/) is built to catch statement-level tampering—signs that a real document was edited or that a document was fabricated to imitate a real one. Concretely, it surfaces patterns, such as:

* Altered dollar amounts or modified account balances.
* Copied or duplicated transaction lines.
* Fonts that are inconsistent with a bank's known statement template.
* Layout or formatting anomalies relative to that bank's established structure.


It does not attempt to verify the borrower's underlying financial situation. See [Known Limitations](#known-limitations) below.

## Methodology

[ThumbPrint](https://www.moneythumb.com/thumbprint/) uses AI and machine learning to detect inconsistencies both *within* a single statement and across the population of statements MoneyThumb has processed from that same bank, effectively building a "ThumbPrint" of what genuine statements from that institution look like.

Both analyses feed into a single weighted score.

The underlying technique is patented (US Publication No. 2023/0359815, *"Detection of Altered Documents"*). We don't publish the specific factor set or weighting logic since that's the proprietary part, but the patent filing describes the general architecture.

## Interpreting the Score (-1 to 1,000)

| **Score**  | **Meaning** | **Suggested Action** |
|  --- | --- | --- |
| -1 | Not scorable. Typically, an image-based PDF rather than a native document. Insufficient data to analyze. | Route for manual review or request a native file. |
| 0 | No abnormalities found. | Proceed normally. |
| 1–299 | Low risk | Proceed normally. Spot-check at your discretion. |
| 300–700 | Moderate risk | Flag for a closer look before final decisioning. |
| 701–999 | High risk | Escalate to manual review before proceeding. |
| 1000 | Direct evidence of editing. | Treat as confirmed alteration. Escalate accordingly. |


These bands are a starting point. Some customers tune their internal thresholds (e.g., where "flag for review" begins) based on their own risk tolerance and portfolio performance.

## Using the Contributing Factors

Any score above 0 comes with a list of the specific factors that contributed to it. Rather than routing to the same queue, teams can use the factors to route intelligently. Where available in our [Insights UI](https://insights.moneythumb.com/), the **View in PDF** link highlights the exact location in the document tied to a factor, which shortens manual review time considerably.

## How Thumbprint Compares to Other Approaches

| **Approach** | **What it catches** | **What it misses** |
|  --- | --- | --- |
| Manual Review | Whatever a trained reviewer happens to notice. | Inconsistent, slow, doesn't scale, misses subtle structural tampering. |
| Rules-Based/Metadata Checks | Simple edit markers, "last modified" flags. | Anything where metadata has been scrubbed or the file re-saved. |
| Statement Issuer Verification (e.g., direct login to the bank) | Ground-truth confirmation the account exists and matches. | Costly, slow, raises privacy/liability concerns and applicants often decline to grant access. |
| ThumbPrint | Structural tampering and fabrication signals that survive metadata scrubbing at automated scale. | See [Known Limitations](#known-limitations). |


## Known Limitations 

[ThumbPrint](https://www.moneythumb.com/thumbprint/) is a statement-integrity check, not a full financial or identity investigation. Be aware of what it isn't designed to do:

* **It can't tell you the applicant is lying about their financial picture using real documents.** A borrower who submits genuine, unaltered statements — but only the flattering ones, omitting other accounts — won't trigger a high score, because nothing in those documents was tampered with.
* **It can't catch a fabricated institution it has no history with.** The inter-document model relies on having seen prior genuine statements from that bank. A convincing fake from an obscure or newly-fabricated "bank" may score lower than expected until MoneyThumb builds a baseline for it.
* **It operates at the statement level, not the account or identity level.** It won't detect account takeover, identity fraud, or synthetic identities — only whether the document itself appears authentic.
* **Scores can start high on legitimate new formats.** When a bank changes its template, or a new bank is introduced, early scores may run higher than warranted until ThumbPrint learns the new baseline — expect some early false positives on genuinely new formats.


## Next Steps

Thumbprint Quickstart
Go here for a walkthrough of our API.