Empirical determination of the weights of startup investment attractiveness factors: from expert distortions to machine learning and SHAP analysis

The paper develops and empirically verifies a methodology for the objective determination of weighting factors of investment attractiveness of high-tech startups, completely excluding subjective expert judgments. Based on a critical analysis of the epistemological limitations of classical methods (Saati, Delphi, direct rankings), the fundamental inconsistency of expert weights is proved due to cognitive distortions (accessibility, anchoring, confirmatory bias), the inability to represent nonlinear and threshold effects, as well as the lack of reproducibility. A descriptive epistemology is proposed as a radical alternative: weights are extracted from large arrays of retrospective data (thousands of startup cases) through training an ensemble of decision trees (gradient boosting) and subsequent SHAP analysis– an axiomatically based tool for distributing predictive contributions based on cooperative game theory and Shapley values. The resulting weights are normalized to the sum of 100%, have asymptotic consistency (variance tends to zero as the sample grows) and natural stage differentiation: retraining the model on a subsample of successful projects shifts the focus from institutional signals at the early stages to technological drivers at the scaling stage. The methodology allows us to build reproducible, objective and adaptive weighing systems for any high-tech sectors to the stage of project development. Empirical validation on deferred test samples and real market cases confirmed the superiority of machine scales over expert scales in terms of AUC-ROC metrics, lift factor, reproducibility, and sensitivity to industry and stage specifics. The proposed approach sets a new standard for quantifying success criteria in venture analysis.