rbaa-validation: Back-Test of Risk-Based Attention Allocation (RBAA) on Real Open Data — Eurostat AI Adoption and 284,807 Real Transactions
Abstract
Abstract:This software package is an empirical companion to the article "Cognitive Optimisation of SME Decision-Making through Risk-Based Allocation of Managerial Attention" (Administrative Sciences, MDPI). Its self-contained Google Colab notebook validates the risk-based attention allocation (RBAA) rule with real, openly licensed data in two parts. Part A documents the context from Eurostat (isoc_eb_ai). In 2025, 55.0% of large EU enterprises used AI, compared with 20.0% of all enterprises. Part B is a counterfactual back-test on 284,807 real card transactions with confirmed fraud outcomes (ULB/Worldline, ODbL). An AI model trained on earlier data supplies calibrated error probabilities. RBAA is then compared, in euros and under the same managerial attention budget, with amount-based approval limits, random spot checks and an AI-score alert queue. At the default assumptions (€5 per review), RBAA reviewed 0.10% of 129,878 test operations and achieved: net benefit of €10,352, i.e. €481 per review hour (ROI 16×); 56% of fraud value stopped; +€14,295 over approval limits and +€11,046 over random checks (both p = 0.001, hour-block bootstrap); +€1,474 over the AI-score queue, which is not statistically significant (p = 0.08). The load-aware stopping rule outperformed the isolated rule, consistent with Hypothesis 4. The package also reports the price of the hard floor, sensitivity across 48 scenarios, robustness across six temporal-split × model combinations, an HTML report, tables, figures and a list of supported and unsupported findings. Reviewer behaviour is assumed rather than observed. No open data record the use of RBAA in enterprises, so this is a back-test of the allocation rule, not a measured effect in any firm. Raw data are downloaded at run time and verified by checksum; they are not redistributed. Keywords: managerial attention; risk-based review; human–AI decision-making; back-test; fraud detection; AI adoption; large enterprises; Eurostat; managerial economics; Google Colab Ці назви й анотації я вже вписав у .zenodo.json та CITATION.cff обох архівів і перепакував їх. Під час завантаження архіву Zenodo підставляє ці поля автоматично. Notebook і результати не змінювалися. Що зробити далі: Пов'яжіть записи між собою. Спочатку опублікуйте rbaa-sim. Потім у записі rbaa-validation додайте в Related works його DOI зі зв'язком is supplement to / references, а також DOI статті, коли він з'явиться. Оновіть Data Availability Statement у статті. Зараз там сказано: «No supplementary datasets or simulation results accompany this version». Після публікації на Zenodo це речення суперечитиме двом записам, тому його варто замінити посиланням на обидва DOI. Файли: /tmp/claude-0/-home-user-bnt/06b9826d-69e1-5db5-8edf-9cbdffb2c8c7/scratchpad/ZENODO_titles_abstracts.md /tmp/claude-0/-home-user-bnt/06b9826d-69e1-5db5-8edf-9cbdffb2c8c7/scratchpad/build/rbaa-sim-v1.0.0.zip /tmp/claude-0/-home-user-bnt/06b9826d-69e1-5db5-8edf-9cbdffb2c8c7/scratchpad/val/rbaa-validation-v1.0.0.zip
