Interpretable Machine Learning for Tokenomic Collapse Risk: Evidence from 155 Real-World Tokens and a Small Prospective Holdout

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Can a token’s collapse be predicted from design parameters fixed at launch, before it has market history? We compile on-chain data for 155 tokens (79 survived, 76 collapsed) and benchmark seven classifiers under leave-one-out cross-validation. Random Forest attains the highest area under the curve (AUC; 0.909; nested cross-validation 0.906), though no pairwise test separates the leading models. The result is not a category prior: out-of-fold AUC stays high within categories (decentralized finance 0.943 on 64 tokens; layer-1/layer-2 0.923 on two collapsed tokens, hence unstable), and excluding the two categories whose labels are near-definitional costs 0.014. Leave-one-category-out validation yields a mean AUC of 0.764 but fails for memecoins (0.571), which bounds the claim. Using only four unambiguously pre-launch features costs 0.038; restricting to long-exposure tokens does not degrade the model (0.921). A 13-token prospective test under distributional shift, with three collapsed cases, is merely indicative. Shapley-value attributions, permutation importance, and Sobol sensitivity agree that inflation rate dominates; team allocation is second by the first two and by standardized logistic coefficients, whereas Sobol ranks whale concentration second. The Safe Operating Envelope, counterfactuals, and fragility screen are in-sample and illustrative. Collapse risk is substantially predictable from design parameters within the categories represented here.