A Pre-Registered Evaluation of JEPA Pretraining for Cross-Sectional Return Prediction

Do self-supervised representations carry incremental alpha? A pre-registered, sealed-test evaluation of JEPA pretraining for cross-sectional return prediction on a point-in-time cryptocurrency universe. Self-supervised learning has reshaped vision and language, and the joint-embedding predictive architecture (JEPA) offers a collapse-resistant recipe for latent-space prediction. We ask, under pre-registered hypotheses and honest evaluation, whether JEPA representations of raw price and volume history carrycross-sectional predictive information beyond classical characteristics. On a point-in-time crypto universe with a sealed test set, an append-only trials ledger, purged combinatorial cross-validation, and a deflated Sharpe ratio computed against the full trial count, the answer is negative. A linear probe onfrozen JEPA embeddings has out-of-sample predictive power (H1 supported, test rank IC 0.047, Newey-West t = 8.7), but it is weaker than two trivial controlsbuilt from the same windows: a 128-component PCA projection and, most tellingly, a randomly initialized, untrained encoder of the same architecture (test rank IC0.077). Adding JEPA embeddings to a gradient-boosted model on classical characteristics does not help and on the sealed test set significantly detracts(primary endpoint: paired daily rank IC difference -0.006, Newey-West t = -5.5). We identify and verify a mechanism: the isotropic-Gaussian regularizer that prevents representational collapse also flattens the embedding variance spectrum, removing the anisotropic directions that carry the volatility and momentumsignal, so the trained encoder is more isotropic and less predictive than the untrained one. Contributions: A pre-registered, sealed-test evaluation with a clean negative answer at the primary endpoint. A random-encoder control that beats the trained encoder, quantifying how much of an apparent self-supervised gain is really an architecture prior. A verified mechanism tying the failure to the collapse-prevention regularizer. A reusable, fully reproducible evaluation harness: sealed test set, append-only trials ledger, purged combinatorial cross-validation, and a deflated Sharpe ratio against the full trial count.