# 11. CVPAT

## Purpose

Does the loyalty model predict better than its comparison benchmark across held-out folds?

## Practice study

- File: [loyalty-prediction.csv](../samples/loyalty-prediction.csv)
- Prepared project: [loyalty-prediction.qpls](../projects/loyalty-prediction.qpls)
- Observations: 420
- Data type: deterministic synthetic instructional data
- Complexity: a realistic multi-construct model with multi-item measurement blocks

## Exact procedure

1. Start QuickPLS and choose **Open Project**. Select [loyalty-prediction.qpls](../projects/loyalty-prediction.qpls). This prepared teaching project contains the exact dataset, analysis-ready base model, bindings, and a saved layout. Complete any method-specific term, group, or higher-order instruction stated below before calculation.

2. Open **Data** and confirm **loyalty-prediction.csv** contains 420 observations. Return to **Model**. The linked CSV remains available separately for inspection and re-import practice.

   ![CVPAT sample data](../screenshots/cvpat/01-data-ready.png)

3. Confirm that the prepared model contains these constructs and measurement blocks: Ease (ease_1–ease_3), Quality (quality_1–quality_4), Trust (trust_1–trust_4), Intention (intention_1–intention_3), and Loyalty (loyalty_1–loyalty_3), all reflective.

   Confirm these saved structural paths: Ease → Trust; Quality → Trust; Ease → Intention; Quality → Intention; Trust → Intention; Trust → Loyalty; Intention → Loyalty.

4. The prepared project already contains a saved, neatly arranged model layout. Do not choose **Arrange** or otherwise change the canvas before this tutorial calculation. If you later edit the model, arrange it once and save the edited project before calculating.

   ![CVPAT prepared model or variable roles](../screenshots/cvpat/02-arranged-model.png)

5. Choose **Validate** and confirm there are no blockers. Do not save the unchanged prepared project; save only after you intentionally edit it.

6. Use **Calculate → PLSpredict / CVPAT**. QuickPLS calculates the indicator-level prediction results and single-model CVPAT benchmark assessments together; there is no separate mode selector in this release.

   ![CVPAT calculation setup](../screenshots/cvpat/03-calculation-setup.png)

7. Complete the method-specific settings:

- Confirm the fixed 10-fold × 10-repeat cross-validation design displayed by QuickPLS.
- Set the seed only when you need a reproducible alternative assignment.

8. Read **Readiness**, leave **Open Results when finished** selected, and choose **Start calculation** once. Do not close the project while the estimate is running.

> **QuickPLS 2.62.8 qualification note:** the installed release completed this example's cross-validation estimate, but publication of the saved Results document then failed because the current prediction result omitted its required typed score-execution contract. The screenshots and Results steps below remain pending rather than showing simulated evidence. Use this page for preparation guidance until a later QuickPLS release repairs that application pipeline.

9. In **Results**, confirm **CVPAT** is selected and inspect these outputs in order:

1. CVPAT benchmark comparison
2. Average loss difference
3. Prediction-error chart
4. Fold accounting

   ![CVPAT primary result](../screenshots/cvpat/04-results-primary.png)

   ![CVPAT secondary result](../screenshots/cvpat/05-results-secondary.png)

10. Choose **Save Report** to preserve the exact result. Use **Export** for the needed table/report format. Close and reopen the project once and confirm the saved result remains selectable.

## Reading guidance

- The data are deterministic synthetic teaching data. They are suitable for reproducing the workflow, not for substantive publication claims.