Craft reports a 2.334 Sharpe on recent SP500, but the direction calls behind it barely register on MCC. The model gets 53.9% right (± 0.006 across five seeds), with Matthews correlation of 0.009 and a seed standard deviation of 0.014. This test runs to December 2024, after a validation window ending November 2023.
What Craft sees
Soun, Lee and Kang predict each stock's next-day up/down direction using daily adjusted closes alone. For each stock and day, they take a window of w prices; the search covers 5, 10 and 20. Dividing each price by the first produces w-1 relative returns. A k-means fit on the training split pools those vectors across stocks and days, then assigns each vector to its nearest centroid. The paper calls that centroid a token and uses the centroids as fixed embeddings. Validation accuracy selects k: 200 in two test regimes and 150 in the third. A token is meant to retain the recent path's shape while discarding stock-specific wiggle.
Peers supply the other input. Each day, a 60-day correlation matrix of returns across the universe yields its top 10 eigenvectors, which serve as stock coordinates. The model forms each stock's "global context" by averaging other stocks' current tokens, giving less weight to stocks farther away in that space. In the paper's illustration, Tesla's nearest neighbour changes from AAPL in June 2021 to NVDA at the end of 2024. MSFT enters the top five; AVGO and AAPL remain in the top three.
Masked self-attention processes the two token sequences and learns to predict the next token. Attention across the cross-section then takes every stock's last-step vector and outputs p(up) for day t+1. The trading premise is that short price shapes recur more usefully when the model also knows its peers' shapes. Across six datasets in four markets and three test windows (Q1 2020, H1 2024, and a longer recent window), Craft leads on accuracy in 12 of 14 settings. Recent CSI300 is its best cell: 0.556 versus 0.515 for the runner-up, Kalman+LSTM, with MCC of 0.104.
How much of 53.9% clears always-up?
The paper treats accuracy above 0.50 as signal because up-day base rates are near 0.50. Recent SP500 has an up-day base rate of 0.52, leaving Craft 1.9 points ahead of an always-up rule. On CIKM24, the other US dataset, its 0.531 accuracy exceeds the 0.52 base rate by 1.1 points. MCC there is 0.019, tied with plain ALSTM, although Craft's accuracy beats ALSTM's 0.499. Against the strongest baseline, the Diebold-Mariano test on SP500 gives p = 0.001. That competitor gap survives a Bonferroni cut across all 14 cells. The always-up gap remains 1.9 points.
The methods also receive different treatment in the paper's setup. Craft gets a grid search over seven hyperparameters plus k; baselines use repository defaults. MASTER and MATCC receive Alpha158 plus 63 market features. Yet on the recent US tests, MASTER's MCC is negative: −0.017 on SP500 and −0.010 on CIKM24. MATCC records 0.003 on SP500.
The authors acknowledge the small margins. Accuracy in their tables runs from 0.50 to 0.56, and MCC values "remain close to zero". They argue that gains over 0.50 "translate into materially different investment outcomes". The paper does not report the accuracy that would show that translation on the days the strategy trades. Its 2.334 annualized Sharpe on recent SP500 comes from a gated book: Craft goes long when p(up) exceeds 0.60, shorts when p(down) exceeds 0.60, and equal-weights each leg. We did not find accuracy for that confident subset. The published accuracy tables cover every prediction, including those that never enter the book.
Same-close fills
The signal uses the day-t close, and the simulation enters at that close before exiting at the next one. The authors describe "an idealized fill that ignores intraday execution latency". Applying the protocol to every competing method supports their ranking comparison. It leaves the level of the reported 2.334 Sharpe exposed to the fill assumption.
The simulation charges 6 bps one-way on sells only, without spread or impact. With a two-sided 6 bps charge, Craft still has the highest Sharpe on 4 of 6 recent datasets. We did not find turnover. On a daily close-to-close book, that missing figure bears directly on the cost assumption.
Which stocks survived to March 2025?
Each of the four new universes takes the union of the top 20% of stocks by market cap or volume within each sector. Eligibility also requires continuous listing from January 2020 through the March 2025 download. The authors say "this continuous-listing filter introduces survivorship bias" and argue that it "applies uniformly to all evaluated methods", leaving rankings unaffected. They assert that ranking claim without testing it.
The restriction reaches back into the COVID test: the March 2020 short leg cannot hold a stock that delisted later. The reported S&P500 cumulative return of +33%, versus +26% for the market, is therefore a survivor-conditioned level regardless of the ranking. We did not find when the top-20% cut is measured, or a stock count. Taking per-asset latency as rebalance latency divided by universe size suggests roughly 37 to 55 names per dataset (31.2 s over 568 ms, 21.4 s over 578 ms).
The US run we can make
We cannot trade the paper's exact universe. Our test is US-only, with an annually selected US market-cap universe in place of exact historical S&P 500 constituents. Tokens and correlation context require only prices, so the mechanism carries over. The paper's cross-market results and exact S&P 500 figures do not transfer to our universe.
Its portfolio figures remain reference points for our run, rather than our results. On recent SP500, the paper reports a 2.334 Sharpe and 6.1% drawdown. Its drawdown claim reaches "up to 7.0 points" on recent CSI300, where Craft's 4.0% compares with MATCC's 11.0%. Craft does not have the lowest drawdown on recent SP500: MATCC posts 5.8% against 6.1%. NI225 likewise has MATCC at 9.6% against Craft's 11.2%. The COVID-window S&P500 Sharpe of 1.954 spans about three months. The authors note that Sharpe differences in short windows rarely reach significance; Jobson-Korkie is significant in 1 of 8 such cells.
Without a confident call, the gate leaves the book flat.
One result would change our view. If Craft entered at the next close on a point-in-time universe, paid the two-sided costs the paper already tests, and still exceeded the paper's 1.089 Sharpe for an equal-weight SP500 book, it would be carrying more than its 0.009 MCC suggests.