A wallet address helps forecast the next second on Hyperliquid. Ding and Zhai measure a 1.43-point gain in one-second out-of-sample ridge R2 over a literature-based ten-variable anonymous model. Their figures support the claim. How much a second team could recover remains open, and several choices that the paper candidly documents bear on the answer.
Following the wallets
Every submission, cancel, rejection and fill on Hyperliquid carries a persistent pseudonymous wallet address. To reconstruct full-depth books, the authors ran a non-validating node and replayed 17.1 billion messages across the ten most active perpetuals from July 1 to 27, 2026. They connected 14.3 million aggressive orders to 147,113 wallets across those ten markets. The economics are familiar adverse selection: some takers move prices after trading. A persistent address lets their current activity inform a forecast of the next second.
Each wallet gets a score: its notional-weighted ten-second signed midpoint markout over July 1 to 10. Eligibility requires at least 100 aggressive orders; TWAP and liquidation orders are excluded. This leaves 2,314 wallets. The paper calls the top decile, 231 wallets, the toxic cohort. For the 965 wallets scored in both windows, rankings persist into July 11 to 20 with Spearman 0.52. Validation markouts remain at or below 0.60 bp through D8, before rising to 1.09 bp in D9 and 2.20 bp in D10. There is no mean reversion: the top decile's markout grows from 1.25 bp at half a second to 2.11 bp at ten seconds.
BTC, ETH and SOL enter the forecasting test on a 100 ms grid. The comparison starts with a ten-variable anonymous block covering imbalance, order-flow imbalance, signed flow, recent return, volatility and spread. The authors then add eleven identity variables. These include cohort versions of quote and flow measures and four address-based extras: net buyer breadth over one and thirty seconds, the cohort's share of best-quote depth, and its contribution to best-quote imbalance. At one second, ridge R2 rises from 10.88% to 12.31% (t = 9.2). LightGBM rises from 19.48% to 20.65% (t = 5.0).
What does the score capture?
The authors use "toxicity" as a descriptive label for a high markout. Measurement starts at the pre-block midpoint, so the score includes impact from the trade's own block. They use the ten-second oracle to test the obvious latency-arbitrage explanation. Top-decile wallets trade in the direction of the oracle-versus-mid gap only 41.3% of the time; the bottom decile does so 57.9% of the time. Adding gap controls moves the D10 minus D1 difference from 1.72 to 1.75 bp. The authors explicitly acknowledge the limit: a ten-second oracle cannot detect sub-second cross-venue moves.
Full instructions, four sensitive choices
The fitting recipe is specified in full. Ridge penalties are 10^(-7+0.5j) for j = 0 to 16, selected over nine expanding daily folds. LightGBM uses learning rate 0.05, 127 leaves, a 5,000-observation leaf floor and seed 20251201. The paper even spells out the depth-share clamp. Four specified choices deserve particular attention from a replicator.
Near-touch volatility scaling matters most. The coin's previous-day one-second volatility multiplies both near-touch imbalance measures, anonymous and cohort. With that scaling, the December 2025 identity increment is +1.80 points (12.36% to 14.16%, t = 3.9). Without it, the increment is -1.17 points. The authors describe the December result as specification-dependent and say the scaled December figures have not yet been audited end to end. They also disclose that they selected this scaling after seeing results.
One transform reverses the December check's sign.
Second, the blocks use different constructions for order-flow imbalance. Cohort OFI adds changes in bid-minus-ask USD depth; anonymous OFI follows the price-indicator transitions of Cont et al. The paper discloses the difference. We could not determine from the text whether a common construction would raise or lower the increment.
Third comes stale state. Emitted states carry forward on the grid during data gaps and oracle-gate closures. The gate was closed for 1.83% of evaluation-window coin-time, and July 27 alone contains a 28,326-block gap. Excluding July 25 and 27 without refitting leaves ridge R2 at 11.72% for the anonymous model and 13.18% with identity. For trees, the corresponding figures are 19.75% and 20.90%. Those two days therefore do not generate the finding. Stale states on the other five days remain in place, and the authors say a fully gap-masked rerun is still outstanding.
Fourth, consensus time indexes every feature. There is no audited local receipt-time clock. The 200 and 300 ms delay tests therefore cannot establish that a trader received the state in time.
The benchmark adds another choice. The authors describe it as non-maximal: it omits multi-level depth, cancellations, rejections, intensity and cross-venue predictors. Those additions, they say, "may reduce the measured increment," and they cite earlier work finding that rejected orders predict returns on this venue.
Trees retain most of the gain
Measured in points, most of the identity gain survives the move to trees. The anonymous R2 nearly doubles, reaching 19.48% against 10.88%, while the identity increment shrinks only from 1.43 to 1.17 points. Drop-one refits help locate the source. Removing near-touch imbalance costs the anonymous model 1.93 points. With identity included, removing it costs 0.24. Cohort one-second OFI (0.67) and breadth (0.37) contribute most among the identity variables. On its own, the identity block reaches 10.06%.
The authors suggest that identity may offer a cleaner view of the same aggregate flow. To examine whether any similarly active subset would suffice, they draw 200 activity-matched cohorts. The best produces 0.88 points, or 61% of the 1.43-point increment. The toxic cohort beats every draw through ten seconds. Donor limits leave the comparison cohorts smaller than the toxic cohort, however, and the authors describe the ranking as descriptive.
Delay looks more forgiving in relative terms than in points. At a 300 ms lag, ridge retains a 9.1% relative gain, while anonymous R2 falls to 6.40%. Our arithmetic puts the remaining increment at about 0.58 points, down from 1.43. For trees, anonymous R2 declines from 19.48% to 9.64%; the relative gain declines from 6.0% to 4.0%. This leaves about 0.39 points, compared with 1.17 undelayed. The authors interpret these same figures as evidence that delay affects the increment less than the anonymous benchmark.
Seven days behind every t-statistic
Seven paired daily MSE differences produce t = 9.2. The authors call their t-statistics descriptive and unadjusted. They also disclose that the decile cut, scaling and July window were chosen after results were available. Ding and Zhai call the design "a chronological out-of-sample comparison rather than a prospectively locked test." That is an accurate description. The top-quintile result (13.1%) suggests the cut is not fragile. The December sign reversal says the scaling is.
Our data could not support the wallet test
We have neither crypto perpetuals nor order records tied to persistent wallet IDs. Our finest bars span one minute, too coarse for a one-second target constructed from 100 ms features. A minute-bar proxy would test a different signal.
From forecasts to trading returns
Mean pre-trade spreads were 0.20 bp for BTC, 0.62 for ETH and 0.26 for SOL. The top-ventile adjusted markout was 3.11 bp. On paper, that gap makes the idea look tradable. But measurement begins before the toxic wallet's own block, so some of the markout is impact that a follower reaches too late to collect. The paper does not model fees, queue position or fill probability. Its latency treatment consists solely of the 200 and 300 ms feature delays, indexed to consensus time.
The paper plainly states that its results do not "establish realized trading profit." Our earlier note on Shenzhen exact-match orders encountered the same limit: the mid-quote forecast is measured before the spread that the signal's own orders must cross.
A prospectively fixed rerun with gap masks, still delivering around a point of one-second R2, would move this from a promising field to a signal worth building.