Why this study exists: the original preset criteria were selected by testing on a single six-month window, which risks overfitting: criteria chosen and judged on the same data tend to look better than they are. This study re-tested the criteria out of sample.
Design: 49 candidate criteria sets (the live criteria plus variants) were frozen in advance and run walk-forward over October 2023 to October 2026: each candidate was selected on a rolling 12-month window and scored only on the following 3 months it had never seen, across eight test windows. The universe was rebuilt to correct a construction flaw in the first study: 2,733 token-bearing protocol groups, with revenue collected per protocol across every chain it operates on, and an adapter-history gate so protocols only become eligible once their data existed. Outcomes are 30-day returns versus Bitcoin, net of a uniform 1% round-trip cost assumption, using DeFiLlama chart prices (sampled, not exchange closes).
Findings on performance: none of the 49 candidates beat Bitcoin on average. The best candidate's median 30-day result was -2.6 percentage points versus Bitcoin; the live Confirmed Momentum criteria measured -6.3 points across 269 flags; the median eligible DeFi token measured -10.4 points, so the screens rank DeFi tokens against each other better than chance without beating Bitcoin.
Findings on discovery: a separate audit asked whether the screens surface protocols that later become narrative-relevant (an objective cohort of 523 tokens defined by peak size, peak multiple, category leadership, and sustained relevance). The live screens captured 23.3% of that cohort versus 8.6% for a naive rank-by-revenue-growth baseline, about 2.7 times the rate, and captured 8.2% early (at least 30 days before the defining milestone) versus 3.3% for the baseline. The typical capture happened at the milestone rather than before it (median lead -2 days). Named cases: Derive flagged early (May 2025), Pons flagged early (July 2026), Hyperliquid and Lighter flagged late, Venice not capturable because its narrative formed before it was screen-eligible.
Limitations: fundamentals are DeFiLlama's current (restated) histories, not as-originally-published snapshots, so results are upper bounds; prices are sampled chart prices, which can miss intra-period peaks; historical market-cap and FDV figures are proxies (price times total supply) that failed validation against external FDV data and are used for rough banding only; some delisted protocols remain unrecoverable in any current registry. Conclusion: the criteria are frozen as discovery screens. They are not changed on the basis of this study, and no threshold in this product should be read as a tested predictor of returns.