Beyond Anonymous Order Flow: Wallet Identity as a Feature for Short-Horizon Return Prediction

summary

The gist

Wallet scores have an across-window rank correlation of 0.52 and adjusted validation markouts rise sharply in the upper tail…

In short

The study evaluates whether tracking persistent wallet histories can improve short-term return predictions beyond anonymous order flow data. By scoring wallets based on past aggressive trades, researchers found that these wallet identities provide a measurable predictive edge, particularly when compared to anonymous features. Wallet markout rankings remain stable across time windows, suggesting historical trading behavior is a valuable signal for future price movements.

Key concepts

Wallet Scores
These scores are calculated by looking at the notional-weighted mean markout of aggressive orders from a wallet over a ten-day window. This metric captures what traders believe about future prices based on their past trading actions, helping to rank wallets by their perceived influence.
Persistence of Wallet History
The research tests if a wallet's past aggressive trades can predict future returns. The key finding is that the ranking of wallets based on these historical markouts remains correlated (0.52) even when comparing different scoring windows, suggesting that long-term trading behavior is a persistent predictor.
Identity Features
These are variables derived from identifying a wallet's history, such as those extracted by gradient-boosted trees. The analysis shows that using these specific wallet identities can significantly boost predictive performance in short-horizon return models compared to using only anonymous order flow data.

Terminology used across episodes

This episode discusses

The paper

Beyond Anonymous Order Flow: Wallet Identity as a Feature for Short-Horizon Return Prediction · Read on arXiv

Jingqiu Ding, Daojing Zhai

ETH Zurich · Baruch College, City University of New York

Conventional public order-book data summarize trading without identifying the accounts behind it. Hyperliquid's Level-4 record also exposes persistent pseudonymous wallet addresses. We reconstruct full-depth books from consensus-ordered messages and evaluate wallet identity as a feature for short-horizon return prediction. The July 1-27, 2026 record covers 17.1 billion messages, 14.3 million aggressive orders by 147,113 wallets, and 84.3 billion in taker notional across ten markets. In the three model markets, notional-weighted ten-second markout rankings persist across adjacent ten-day windows (Spearman rho = 0.52); covariate-adjusted validation markouts reach 3.11 basis points in the top ventile. On a 100-millisecond grid, eleven features describing previously scored top-decile wallets raise one-second out-of-sample R2 from 10.88% to 12.

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: We're looking at the paper: "Beyond Anonymous Order Flow".

Warren: Here's the gist of the paper. Wallet scores have an across-window rank correlation of 0.52 and adjusted validation markouts rise sharply in the upper tail, reaching 3.11 basis points in the top ventile The empirical contribution is the evaluation of persistent wallet histories, building on the existing Level-4 data literature Data Reconstruction and Scoring Methodology The pipeline involves reconstructing full-depth books from consensus-ordered messages and evaluating wallet identity as a feature for short-horizon return prediction The process starts with a non-validating node recording committed events, followed by a deterministic replay to rebuild the book and link orders to wallets Wallet scores are built from aggressive orders, as these reveal trader beliefs about future prices The wallet score is the notional-weighted mean markout over a frozen ten-day window, computed over July 1–10 for wallets with at least 100 qualifying aggressive orders Persistence of Wallet History The working hypothesis is that wallets can be ranked by the signed markouts of their past aggressive trades, and that subsequent activity by high-ranked wallets helps predict returns The design requires persistence along two dimensions: addresses must remain observable long enough to accumulate a history, and informativeness itself must also persist The rank correlation between scoring-window and validation-window markouts is 0.52 The top-decile markout rises from 1.25 basis points at half a second to 2.11 at ten seconds and remains elevated at five minutes Predictive Performance of Identity Features At one second, identity raises ridge R2 from 10.88% to 12.31%, a 13.2% relative increase with t = 9.2 Gradient-boosted trees extract more from the anonymous block (19.48%) and still gain from identity, reaching 20.65% (+6.0%, t = 5.0) The increment in cohort quote updates and distinct-buyer breadth is noted Contribution Analysis and Stress Tests The aggregate gain could come from an additional wallet-specific signal or from identity variables substituting for anonymous predictors The largest contribution in the anonymous benchmark comes from near-touch depth imbalance (nearImb, 1.93), whereas the identity block concentrates in the toxic decile’s one-second quote flow (hotOfi1s, 0.67) and net buyer breadth (hotBreadth1s, 0.37) The toxic-decile increment is 1.43 percentage points against 0.88 for the best of the 200 matched draws, so it is 1.6 times the maximum Limitations and Future Work The analysis does not identify private information or establish realized trading profit Retrospective specification choices, carried-forward states during data gaps, and receipt-time availability require further validation A prospectively fixed test should compare richer anonymous baselines and mask invalid states at both target endpoints The increment is therefore less sensitive to these delays than the anonymous benchmark The result concerns prediction relative to a specific anonymous feature set Hyperliquid’s wallet-attributed Level-4 record supports a direct comparison between aggregate order-book predictors and features constructed from persistent wallet histories The result concerns prediction relative to a specific anonymous feature set Wallet markout rankings persist across adjacent ten-day windows (rho = 0.52) In three active markets, features from previously scored top-decile wallets raise one-second ridge R2 from 10.88% to 12.31%, and tree R2 from 19.48% to 20.

Paper summary: Ruth: So we're looking at this paper, "Beyond Anonymous Order Flow: Wallet Identity as a Feature for Short-Horizon Return Prediction," which is digging into how we can use wallet history to predict short-term returns on things like Hyperliquid. The main idea here is that the authors are moving beyond just using anonymous order data and looking at the actual identity of the wallets making those trades.

Warren: Right, so they're testing if you can rank wallets based on their past aggressive trades and if that ranking actually helps predict what happens next in terms of price movement. It’s about whether a wallet’s past behavior is a persistent signal for future returns.

Iris: So the core thesis seems to be that we should look at the signed markouts from past aggressive trades to rank wallets, and see if those top-ranked wallets have subsequent activity that helps predict returns. It builds on existing literature about Level-four data, but it focuses specifically on wallet persistence as a feature for prediction <ref:2611.05218#pg1>.

Ruth: Exactly. They’re essentially testing the hypothesis that wallets can be ranked by their past aggressive trades' markouts, and that activity from those highly ranked wallets helps predict future returns because of how persistent this information is. It’s about building a signal from the history of an account.

Warren: And they found some interesting things regarding that persistence. They found a rank correlation between the scoring window and the validation window markouts is zero point five two, which suggests that even if you score wallets based on one period, those scores hold up for ten days later <ref:2611.05218#pg3>.

Iris: That persistence is key because it addresses a big question about identification in markets. The paper explores whether wallet identity itself can be a meaningful attribute for prediction, especially when traditional models leave the counterparty unidentified and we rely only on anonymous data to infer things like adverse selection or permanent price impact.

Ruth: It’s interesting how they frame it as building on existing Level-four data literature, but they are pushing the idea that where you have identifiers matters more than when you don't, because visible identifiers can be learned and show performance is persistent across time <ref:2611.05218#pg1>.

Paper summary: Warren: They used a specific scoring method for this wallet score: it’s the notional-weighted mean markout over a frozen ten-day window, calculated for wallets with at least one hundred qualifying aggressive orders between July one and July ten <ref:2611.05218#pg1>. That gives them that specific look at their past behavior.

Iris: And they also looked at how this identity feature performs in the prediction task itself. They found that on a one-second grid, eleven features describing previously scored top-decile wallets raise the out of sample ridge r squared from ten point eight eight percent to twelve point three one percent under ridge regression, which is a thirteen point two percent relative increase at t equals nine point two <ref:2611.05218#pg1,grid, eleven features describing previously scored top-decile wallets raise>.

Ruth: That’s a concrete result showing that this identity feature actually provides some predictive lift when you look at the short horizon. But they also looked at what was driving that gain and where it came from in terms of anonymous data versus identity variables.

Warren: In terms of contribution, the largest gain in the anonymous benchmark came from near touch depth imbalance, which they measured as one point nine three, but when you look at the identity block itself, it concentrates in things like the hot of i s one second quote flow with a score of zero point six seven and net buyer breadth with a score of zero point three seven.

Iris: So it seems the increment from using identity variables is coming from those specific micro-features related to how much quoted volume changes for those top wallets in that one-second window, not just some general wallet signal. The toxic decile increment was one point four three percentage points against a baseline of zero point eight eight for the best twenty draws, which is about one point six times the maximum they found.

Ruth: It sounds like the paper is suggesting that we don't have to replace all anonymous predictors with identity variables; rather, there might be wallet-specific signals that add value on top of what anonymous data already gives us.

Paper summary: Warren: And it’s important to remember the limitations they point out. They acknowledge that this analysis doesn’t identify private information or establish realized trading profit, and they also flag that retrospective specification choices, carried forward states during data gaps, and receipt-time availability require further validation.

Iris: That's a fair point about the methodology needing more testing on those specific aspects of data handling. They state that a prospectively fixed test would be better for comparing richer anonymous baselines and masking invalid states at both target endpoints, which means this result is specific to the feature set they tested against an anonymous one.

Ruth: So, what does this mean practically? It supports the idea that persistent wallet histories can be a valuable piece of information when trying to predict market movements on short time scales. It shifts our view from just looking at aggregate order flow to considering the influence of specific, identifiable traders' historical actions.

Warren: For me as someone who runs a strategy, it suggests that if you can track and rank wallets based on their aggressive trade history, those rankings might offer a slight edge in predicting short-term outcomes compared to using only anonymous data features. But you still have to manage the cost and complexity of tracking those specific wallets.

Iris: From an econometric standpoint, the identification hinges on whether these wallet attributes actually provide predictive power that isn't just noise or luck from the specific historical window they used. The fact that the rank correlation holds across ten-day windows is a strong indicator of persistence for their scoring method.

Ruth: So, to wrap up this segment, "Beyond Anonymous Order Flow: Wallet Identity as a Feature for Short-Horizon Return Prediction" suggests that wallet identity isn't just noise; persistent histories can be a valuable predictor when constructed correctly and tested against the right features. Next time we look at how these specific wallet signals integrate with other market data to build a complete picture of short-term trading behavior.

Conclusion: Ruth: So we've been looking at how these wallet histories can actually predict short-term returns, and now we’re coming to the conclusion of this paper, "Beyond Anonymous Order Flow: Wallet Identity as a Feature for Short-Horizon Return Prediction."

Warren: It boils down to this idea that even when you don't know who is trading anonymously, you can use the persistent history of their aggressive trades—their wallets—to build a signal for what might happen in the very short term.

Iris: And they found that this isn't just some random noise; the rank correlation between how we score a wallet and how it performs later, across ten days, stayed at point five two. That suggests the identity signal has some staying power.

Ruth: Exactly. They're showing that if you track these wallets based on their past aggressive actions over a ten-day window, those rankings actually help when you’re trying to predict returns on a one-second or even shorter horizon.

Warren: The numbers they showed were pretty interesting, like how identity features could raise the prediction accuracy from about ten point eight eight percent up to twelve point three one percent on a one-second prediction grid. That's a real lift when you’re looking for those quick wins.

Iris: But we have to be careful about what that lift means because they also pointed out that the biggest gains weren't always from identity itself, but from specific things like how much volume changes in the first second of quote updates for those top wallets.

Ruth: That’s a crucial distinction. It’s not just about knowing *who* is trading, it's about what specific micro-features related to that wallet are doing right now. It suggests we need to look closer at the immediate impact of these persistent histories on the current order book dynamics.

Warren: So, if you’re building a strategy, this paper implies that tracking wallets isn't just extra overhead; it could be a way to capture some specific edge in short-term prediction that anonymous data alone might miss.

Iris: The authors are clear about what they aren't doing though. They aren't identifying private information or proving any kind of realized profit because the test was retrospective, and there are still questions about how the data is handled during gaps in trading activity.

Ruth: Right, so it’s a strong look at using persistent wallet history as a feature, but we need to keep an eye on those technical caveats before we start treating this as gospel for real trading.

Warren: It's definitely worth watching because if these persistence and identity signals hold up in future tests, it could change how we think about what kind of data we prioritize when trying to predict market movements.

More episodes

← Home