> For the complete documentation index, see [llms.txt](https://matterhorn-doc.mometic.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://matterhorn-doc.mometic.com/understand-the-scores/ml-enhanced.md).

# ML enhanced Beta

Turn on Matterhorn ML enhanced Beta, understand its 70% component and 30% learned-pattern blend, and compare it with standard scores.

**ML enhanced** offers a second way to prioritize research. The LightGBM ranking model learns historical combinations of growth, acceleration, demand, quality, valuation, and market behavior associated with stronger subsequent relative returns. Transparent financial components keep the resulting view interpretable.

## Turn it on

Select the **ML enhanced** checkbox beside the **Beta** chip in the workspace navigation. Hover the circled **?** for a brief explanation and, when enabled, the data date. The choice is remembered in this browser. Uncheck it to return to standard rankings.

![Matterhorn Overview with ML enhanced Beta enabled](https://981865776-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FNOJXLd7UJBniObM136Nu%2Fuploads%2Fgit-blob-dcc9a46617dd55b9f06f06c0d531dccdd5d4315e%2Foverview-ml.png?alt=media)

## What changes

| Surface                                                                       | Behavior with ML enabled                                                                                                     |
| ----------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------- |
| Top 10, ranked watchlist cards, Companies rank and opportunity score          | Use matching predictions from the published beta bundle                                                                      |
| Watchlist history beneath the ML score                                        | Shows the separately labelled seven-day Matterhorn score comparison; it is standard-score history, not ML prediction history |
| Theme company preview lists                                                   | Sort matching theme members by ML score; companies without a matching prediction are omitted                                 |
| Company score breakdown                                                       | Shows the ML score, financial-component score, learned-pattern score, rank, evidence quality, and data date                  |
| Company and comparison radars                                                 | Show the eight-axis opportunity shape                                                                                        |
| Analysis, Compounding, Price, History, Evidence, Filings                      | Continue to show the existing company research; they are not replaced by the beta score                                      |
| Detailed checklist, standard investor lenses, and standard percentile panels  | Retain their original basis, even below an ML chart                                                                          |
| Moved this week, theme allocation/stress calculations, and standard analytics | Retain their standard logic                                                                                                  |

In the Companies table, **Emerg**, **Rate**, completeness, and rankability retain their standard meanings. The ML badge does not make every number on the page an ML output.

## What 70/30 actually means

**ML enhanced score = 70% × financial-component score + 30% × learned-pattern percentile.**

The financial-component side combines growth, forward demand, quality, market confirmation, valuation, evidence, and the separate durability estimate. It is **not the original five-component Matterhorn score**. That distinction explains why multiplying the standard score by 0.70 will not reproduce the beta result.

Using rounded displayed values for illustration, a component score of 61 and learned-pattern score of 93 give approximately **70.6**, displayed as **71**. Exact calculations use unrounded values. The split is a fixed release policy, not a proven optimal allocation and not an automatic adjustment by sector.

## What the model is trying to learn

The emergence model learns to rank historical company-month observations by subsequent **24-month sector-and-size-adjusted total-return percentile**. In plain English: which companies later performed unusually well compared with observed contemporaneous peers of similar industry and size? The target is relative performance, not a promised CAGR.

A separate durability model examines whether favorable operating economics persist across the following four quarters. Neither output comes from an LLM assigning a score. See [methodology](/reference/methodology.md) for the training target and validation boundaries.

## Frozen data is part of the contract

The published bundle fixes the model version, feature version, company identities, and observation date. New filings can appear in the ordinary company research before they are incorporated into a newly validated ML bundle. **Refresh** reloads published predictions; it does not retrain them.

Uncovered companies remain **ML unavailable**. If the bundle fails integrity checks, the app falls back to standard mode. A stale but valid bundle can still be displayed with its age; check the date before using it.

> **Pro Tip — Switch modes to expose a question, not to pick the answer you prefer.** When ML lifts a company, identify which combination it favors and inspect the weakest supporting area. When it lowers one, test whether the model has missed a business characteristic the source evidence supports.

This beta changes the viewing and research-prioritization surfaces described above. It has not graduated to automatically control production promotions or execute trades.

Next: [read the eight axes](/understand-the-scores/opportunity-shape.md), [data freshness](/keep-research-current/refresh-and-freshness.md).