Defense Beats Offense: An Honest, Free-Data Study of What You Can and Cannot Predict in Equities and Crypto

We ask a single question with unusual discipline: using only free, point-in-time public data, can you predict which stocks or crypto tokens will rise sharply — and if not, what can you reliably do? Across two asset classes and four pre-registered ex perimental lines, with pre-committed gates, sealed consume-once hold-outs, deflated performance metrics, survivorship-aware label construction, and permutation tests, we find a consistent asymmetry. Return / pump prediction has no tradeable edge. An apparent cross-sectional ranking signal collapses under realistic exit rules, transaction costs, and survivorship correction; on a 29-name mega-cap equity backtest the deflated Sharpe is −1.377 (held-out, GO bar +0.50) and the strategy underperforms buy-and-hold by 7.3 per centage points of CAGR. On the broader S&P-600 small-cap gate, only one of six classic factors carries the correct sign at |t| > 1.5; the equal-weight composite ranks stocks backwards (IC t = −4.73). In crypto, a pump classifier achieves AUC 0.708 and top-decile lift 2.59× — but every long-only basket loses money at every realistic cost (top-decile −1.6 / −6.6 / −11.6% at 5 / 10 / 15% one-way cost on a 30-day horizon), and labelled “risk flags” (buy_tax, n_rug_flags) carry the wrong sign — they raise measured pump rate, an artifact of survivorship and degenerate-trader behaviour. Forensic defense is real and provable. A crypto rug/scam detector validated at AUC 0.83 on a 3,677-token balanced evaluation sample drawn from a 61,339- label pool (ETH 0.921 via Honeypot.is, SOL 0.754 via RugCheck) catches 89 of 100 ETH and 97 of 100 SOL scams at a fixed 0.5 threshold. A separate price/volume only coin death detector achieves AUC 0.743 on 1,267 cached CoinGecko coins (36 died, 1,231 survived). On the equity side, an abstention-first 3-way forensic screen reduces false-alarm-on-SOUND from 5.34% to 1.03% on its diagnosis set and 1.65% on a held-out date the tool never saw, with a permutation-real recall of 22.7%–25.0% on blow-ups (perm p = 0.0005). The defensive screens are honestly one-sided: they reliably refuse to insult healthy names, but they are weak on clean accounting fraud (full-set fraud recall ~3.5%) and on growth/biotech blow-ups that lack a balance sheet tell (held-out false-clear 2 of 8, 25%, PRLB/TNDM). We explain the asymmetry mechanistically: ex-ante, big winners and look-alike non-movers are statistical twins on every directional factor (rocket-vs-flop AUC ≈ 0.70, bullish-factor max |Cohen’s d| = 0.187 ≈ 0), separable only by a volatility / issuance / illiquidity capacity signature (realized 3-month volatility d = +0.56). Di rection is set by future catalysts absent from any public snapshot; structural failure leaves verifiable forensic traces in present-day fundamentals. We release all code, figures, and negative results in full