When a Valid Customer Metric Answers the Wrong Decision

Description

frozen protocols, executable notebooks, parameter manifests, reported summary outputs, README, computational chronology and software requirements.

Steps to reproduce

### Steps to Reproduce **1. Download and prepare the archive.** Download and extract `Article_7_Replication_Package.zip`, preserving its folder structure. Read the README and computational chronology. Install Python 3, JupyterLab or Jupyter Notebook, and the packages listed in `00_README_AND_DOCUMENTATION/requirements.txt`. Ensure that notebook working directories and output paths are correctly configured. **2. Study 1: Rank and selection portability.** Read the frozen confirmatory protocol in `01_STUDY_1`. Run the confirmatory notebook from beginning to end using its specified parameters and random seed. The notebook generates synthetic customer-level data and calculates rank concordance and top-k selection disagreement across customer-value measures. Next, read the frozen robustness protocol and execute the corresponding robustness notebook. **3. Study 2: Economic regret.** Read the frozen primary protocol in `02_STUDY_2`. Run the primary notebook using its specified parameters and random seed. Compare allocations based on alternative customer-value measures with the decision-relevant Profit Lift oracle under capacity constraints. Reproduce allocation values and normalized economic regret. Then read the frozen robustness protocol and execute the robustness notebook. **4. Study 3: Dynamic policy value.** Read the frozen primary protocol in `03_STUDY_3`. Run the primary notebook using its specified parameters and random seed. The analysis incorporates heterogeneous future customer-state effects, persistence, decision horizons and spillovers. Compare allocations based on current Profit Lift with those based on prospective Policy Value. Then execute the frozen robustness protocol and notebook. **5. Verify the results.** Compare generated summaries and diagnostics with the archived files in `04_REPORTED_OUTPUTS`. Examine rank concordance, top-k disagreement, allocation value and normalized regret, as applicable. Confirm that Study 3 recovers its structural-null conditions, in which current Profit Lift and prospective Policy Value coincide because no additional heterogeneous future economic information enters the decision. Allow for minor floating-point differences. **6. Preserve the study chronology.** The studies were developed sequentially, with frozen primary and robustness protocols. Study 1 informed subsequent theoretical refinement. The computational program was closed following Study 3 robustness testing. All analyzed data are synthetic. Full replication-level datasets are not included because they can be regenerated using the archived notebooks, parameter settings and random seeds. The findings establish computational relationships and boundary conditions, not the prevalence or economic magnitude of measure substitution in actual organizations.

Institutions

• Grand Canyon University

  Arizona, Phoenix

Categories

Licence