Development and Evaluation of a Finish-Time–Conditioned Split Model for the Athens Marathon: A Retrospective Analysis of the 2024 and 2025 Races

Background: Race-specific split calculators can translate a proposed marathon finish time into intermediate passage times. Their descriptive accuracy, however, must be distinguished from evidence that following those times improves performance. This study developed and evaluated a finish-time-conditioned model for the Athens Marathon. Methods: A deterministic, stratified sample comprised 4,000 race participations from 2024 and 2025, with net finish times of 2:30:00–5:00:00. Of 3,870 technically complete records, 3,714 passed development-referenced screening for unusually uneven pace profiles. Possible repeat participants were grouped across development (n = 2,138), calibration (n = 525) and original holdout (n = 1,051) sets. Four prespecified model families allocated total time across nine course intervals. Model selection used grouped cross-validation, numerical checks and a parsimony rule. The selected eight-parameter model, H1, was frozen before a further 500-record challenge; 484 records were complete and 466 passed the frozen screen. Comparators were even pace and a frozen grade-adjusted pace implementation. The primary metric was mean absolute error (MAE) across eight intermediate checkpoints, equally weighted across observed year-by-finish-time cells. Results: Original holdout MAE was 138.1 s for H1, 236.5 s for grade-adjusted pace and 182.5 s for even pace. Challenge MAE was 142.8, 253.5 and 193.5 s, respectively. H1 improved challenge MAE by 110.7 s (95% bootstrap interval 99.5–121.8) and 50.7 s (41.7–60.0). Improvement persisted when all complete challenge records were included. H1 error increased from 83.3 s at 2:45–3:00 to 225.0 s at 4:45–5:00. Nominal 80% empirical ranges achieved 78.0% pooled pointwise coverage but 54.1% simultaneous coverage across all eight checkpoints. No fresh challenge records covered 2:30–2:45. Conclusions: H1 described intermediate splits more accurately overall than the tested baselines among sampled finishers from these two races. Individual errors, incomplete fast-end coverage and outcome-based screening constrain interpretation. The model provides a descriptive reference conditional on finish time; it does not establish an optimal pacing strategy or predict whether a runner can achieve the supplied target. Publication status: Preprint, version 1.0. This work has not been peer reviewed. Author contributions: Georgios Svarnas conceived the study, defined the research questions and sampling goals, directed the modeling and evaluation workflow, and interpreted the findings from a coaching perspective. All AI-assisted work was carried out under his explicit instructions, strict guidance and supervision. He personally read and reviewed all results, edited their presentation, and revised and approved the manuscript and all final research deliverables. He takes responsibility for the work and its final content. Use of artificial intelligence: Selected OpenAI models accessed through Codex were used selectively under explicit, task-specific instructions and the strict guidance and supervision of Georgios Svarnas. Their roles included data-collection orchestration, code generation, data processing, statistical implementation, computational checking, figure production, and assistance with the initial drafting and structure of the manuscript and supporting documents. The author directed the analytical choices and personally reviewed all results. All manuscript text, figures and final deliverables were personally reviewed, edited where appropriate, and approved by him. AI tools provided technical and editorial assistance; the author retained control of the research decisions, interpretation and final content, and assumes responsibility for the deposited work. Funding and competing interests: This research received no external funding. Georgios Svarnas owns Svarnas Coaching, the endurance-coaching business whose online resources include the Athens Marathon split calculator evaluated here. The author declares no additional competing interests related to this study. Deposit contents: The main research manuscript, supplementary methods, and supporting materials containing the frozen H1 model, reference inference implementation, course features, aggregate tables, scientific figures, and a computational audit. Source-linked participant records and original acquisition artifacts are excluded. The supporting materials reproduce model inference; they are not a stand-alone reproduction of participant-level acquisition, fitting, or evaluation.