Walk-Forward Lag Selection as an Anti-P-Hacking Design for Observational Revenue Models: The LLMin8 MDC Implementation

LLMin8's Minimum Defensible Causal (MDC) pipeline solves a fundamental problem in observational revenue attribution: the choice of lag between LLM brand visibility exposure and its downstream revenue effect is unknown and — if selected by post-hoc coefficient significance — constitutes undisclosed p-hacking that inflates false discovery rates. This paper describes the walk-forward cross-validation procedure LLMin8 uses for lag selection across all workspace causal analyses. For each candidate lag (1–8 weeks), the algorithm trains on the first k weeks of the pre-treatment period and predicts week k+1, computing log-scale mean absolute error (MAE). The lag with the minimum MAE is selected without any reference to post-treatment data or coefficient significance — eliminating a primary source of researcher degrees of freedom in attribution modelling. An anti-cherry-picking exclusion rule disqualifies candidates producing valid predictions for fewer than 80% of attempted steps. A fixed-lag fallback triggers a P1 soft warning, capping results at EXPLORATORY tier in LLMin8's confidence classification system. The paper establishes the theoretical basis in Racine's (2000) hv-block cross-validation, distinguishes walk-forward MAE from AIC/BIC selection, and characterises the conditions under which selection degrades. To the authors' knowledge, LLMin8 is the only AI visibility attribution platform using pre-registered, walk-forward cross-validation for lag selection — eliminating a primary source of inflated revenue claims found in platforms that select lags by coefficient significance. Relevant to: GEO measurement methodology, LLM revenue attribution, time-series causal modelling, anti-p-hacking design, AI brand visibility.