A 68.20% net U.S. return is hard to trust when the short book pays no borrow fees. Luo, Yi, Zhang and Sun make a stronger case for ranking stocks. Their ablations show that Stock-JEPA's learned revision adds cross-sectional information beyond a plain Ridge forecast, while their own backtest disclosures leave the trading result exposed.

What the revision learns

The starting point is an expanding-window Ridge regression (alpha = 20), refit at block boundaries. It maps 40 point-in-time descriptors, 28 stock-level and 12 market-level, to eight targets: return, volatility, drawdown and market-relative return at 5 and 21 trading days. Its eight forecasts and eight block-level residual scales give each stock a 16-number daily prior.

JEPA predicts an encoded account of the future rather than its raw values. The target encoder turns the next 21 days of those same 40 features into 128 dimensions. A small network called the prior projector maps the 16-number prior into that space. This produces the anchor, an estimate of the future embedding from Ridge alone. A Transformer reads 126 days of history and predicts the revision: the portion of the gap between the realised future embedding and the anchor that history can forecast. Separate losses keep the revision from taking credit for information already in the prior.

The JEPA stage receives no return label. Ridge does: it fits realised 5- and 21-day returns, and the latter is also the readout's prediction target. With everything upstream frozen, a final MLP takes the history mean, prior, anchor and revision (312 inputs) and produces a 21-day return score. The intended division of labour gives slow, persistent structure to the prior and faster, history-predictable signal to the revision. For a fixed target encoder and unrestricted predictors, the authors prove that the optimal revision reduces the anchor's squared error by exactly E[‖Δ‖²]. They also acknowledge that "lower representation-prediction risk also does not by itself guarantee downstream economic gains." The tests have to establish the economic case.

The sample covers 10,152 U.S. securities from Norgate and 5,669 Chinese securities from Wind, including delisted names. Tests span January 2024 to June 2026, with three seeds per model.

Does it outrank the simpler inputs?

On U.S. RankIC, Stock-JEPA reaches 12.96%, ahead of FactorVAE's 11.18%, the best of 13 baselines. FactorVAE leads on Pearson IC, 12.85% against 11.67%. Stock-JEPA's advantage is in the ordering of names.

The matched readout comparison is more persuasive than the headline table. Using one frozen checkpoint and a fixed MLP, the authors change only the inputs. Add the revision to history, prior and anchor, and U.S. RankIC rises by 3.67 points; net Sharpe rises by 0.53. Replace the revision with the raw 128-dimensional context embedding from which it is computed, keeping input width identical, and RankIC falls by 2.27 points. A separate controlled ablation gives every readout the prior. There, plain JEPA with prior conditioning zeroed trails the separated model by 3.40 U.S. RankIC points (9.05% vs 12.45%). Those comparisons give the prior-relative revision a clearer claim on the ranking gain.

The loss separation itself has a smaller ranking effect. U.S. RankIC is 12.45% with separate losses and 12.37% with a joint loss; China favours the joint version, 11.49% to 11.11%. Portfolio results favour separation: net Sharpe is 1.51 versus 1.25 in China and 1.27 versus 1.09 in the U.S. The evidence also rests on three seeds and roughly 2.5 years of daily forecasts for overlapping 21-day returns. We did not find IC standard errors adjusted for that overlap.

China supplies a useful check on the scale of the gain. A Fama-MacBeth-style regression reaches 8.76% RankIC there, above every general deep model; LSTM tops out at 8.24%.

The daily book

Each morning, the U.S. portfolio selects the top 30 and bottom 30 names, weights them equally and enters at the next open. A batch stays in place for 21 sessions, without rebalancing, before closing at that day's close. Each new batch receives 1/21 of NAV per leg. With 21 batches live, entry exposure is roughly 100% of NAV long and 100% short, or about 2x gross.

The paper charges 5 bps per unit of executed notional at entry and exit, without netting trades across batches. By our arithmetic from those rules, daily trading is about 4/21 of NAV, or around 48 times NAV annually. At 5 bps, annual commissions come to roughly 2.4% of NAV. Against 68.20% annual return, that charge is small. It is also the paper's entire cost model.

Who pays for the shorts?

The authors state that the backtest "does not model market impact, execution restrictions, stock-borrow availability, borrowing fees, or funding interest." Their U.S. universe includes small-cap and highly volatile stocks, and they flag the exposure of 30-name short legs to upward jumps. The bottom 30 of that eligible cross-section is likely to contain scarce borrows. With the short book at about 100% of NAV, every point of average borrow fee reduces annual return directly. An unavailable locate prevents the trade altogether.

Impact matters too. Thirty equal-weighted picks from a universe that includes small caps could be small stocks. We did not find a liquidity screen or a capacity estimate. The paper defines daily turnover, yet we found no turnover figure in its results tables and no factor-adjusted alpha. Without that alpha, the contribution of size or short-term reversal to the 1.24 Sharpe remains unclear. Results also vary widely across seeds: net return is 68.20% ± 15.92 and drawdown is -40.12% ± 13.12. The authors accurately call that the least severe drawdown among U.S. methods. T-JEPA is closest at -42.86%, followed by MASTER at -43.31%. A 40% loss still warrants attention in a book that is dollar-neutral at entry.

China's portfolio is long-only, so its 45.66% net return includes market beta. We found no benchmark-relative figure. Its -31.09% drawdown exceeds PRISM-VQ's -13.64% in severity.

The ranking result is worth building on.

A U.S. rerun limited to liquid, borrowable names and charged realistic borrow fees would change my view of the portfolio. Under those same restrictions, Stock-JEPA would also need to beat a Fama-MacBeth-style regression rerun; that regression scores 1.11 under the paper's current rules. A Sharpe advantage there would show the revision paying for itself in dollars as well as rank correlation.