acceptodds
Under review as a conference paper at ICLR 2027

Open-world tokenized equities: a dataset linking blockchains and equity markets

Abstract

Tokenized equities are an emerging class of real-world assets that connect public blockchains with traditional equity markets. Yet this rapidly growing market remains underexplored. Existing AI-for-finance datasets model prices, text, and inter-stock relations, while blockchain datasets capture wallets and transactions; neither jointly observes public on-chain activity and the corresponding listed-equity information. We introduce, to our knowledge, the first large-scale ticker-aligned dataset bridging these domains, covering 890 tokenized equities, 249.5M transfers, and 31.3M swaps, together with exchange prices, news, and fundamentals. This alignment uncovers new empirical findings. Tokenized equities remain active when U.S. exchanges are closed, and fresh, liquid on-chain quotes can predict subsequent exchange price movements. A small set of liquid tokenized assets also carries information about the broader equity market, while account-level co-holding is more strongly associated with shared news attention than with return correlation. These findings motivate two complementary learning tasks enabled by the linked dataset: dark-window equity forecasting, which augments off-chain market data with evolving on-chain signals to predict the direction of the next opening price, and open-world next-asset prediction, which predicts participation in newly tokenized equities from account histories and asset information. Together, the dataset, findings, and tasks introduce a new learning setting spanning on-chain–equity forecasting and temporal asset recommendation.

Primary area:
datasets and benchmarks
Keywords:
financial datasets, tokenized stocks, dynamic transaction graphs

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_v95kOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a large-scale, ticker-aligned dataset connecting traditional equity markets with public blockchain activity. The dataset covers 890 tokenized equities across four blockchains, including approximately 249.5 million transfers and 31.3 million decentralized-exchange swaps, together with corresponding minute-level exchange prices, news, and company fundamentals. The paper uses this cross-market alignment to investigate on-chain activity during exchange closures, the information contained in on-chain prices, and account-level behavior across tokenized assets. The authors report four main empirical findings. First, tokenized equities remain active while U.S. equity markets are closed, and the divergence between on-chain prices and the latest exchange prices increases as the latter become stale. Second, when on-chain trading is sufficiently active and quotes are sufficiently recent, token returns during exchange closures are informative about the direction of the next exchange opening. Third, a small number of liquid tokenized index or sector assets may provide market- or sector-level signals for a broader set of equities. Fourth, account-level co-holding patterns are more strongly associated with shared news attention than with return correlations. Based on these findings, the paper formulates two learning tasks. The first is dark-window equity forecasting, which uses on-chain signals observed during exchange closures to augment existing equity forecasting models and predict the next opening gap. The second is open-world next-asset prediction, which uses an account’s historical holdings and interactions to predict the next tokenized equity it will acquire. The experiments suggest that incorporating on-chain information can improve cross-sectional RankIC for several equity forecasting models, while the next-asset prediction task highlights the cold-start and distribution-shift challenges caused by newly listed tokenized assets.

Critical strengths
  1. A novel and valuable dataset The paper aligns on-chain activity for tokenized equities with traditional equity prices, news, and fundamentals, connecting two research areas that have largely been studied separately. The resulting dataset provides a new basis for studying price discovery during exchange closures and account-level investment behavior.
  2. Strong connection between empirical findings and learning tasks The paper identifies dark-window signals, cross-asset proxy relationships, and co-holding patterns, and uses these findings to formulate dark-window equity forecasting and open-world next-asset prediction. These tasks are practically relevant and introduce useful research settings for financial forecasting and temporal recommendation.
Critical weaknesses
  1. The central forecasting claim requires stronger validation. The reported directional accuracy above 80% is obtained only under restrictive conditions involving high trading activity and fresh quotes. However, the corresponding sample size, asset coverage, and threshold-selection process are not clearly reported, making the generalizability of this result uncertain.
  2. The event definition in the account-level task requires further validation. Blockchain addresses and token transfers provide only an approximate representation of investor asset choices, since some events may result from airdrops, custody movements, or market-making activity. The paper should describe its filtering procedure more clearly and verify whether the main results remain valid when only confirmed purchase events are retained.
Questions for authors

1.For the reported directional accuracy above 80% under the conditions of at least 300 trades and a quote age of no more than one hour, how many windows, dates, and equities are included? Were these thresholds selected only on the discovery split and then fixed for evaluation on the confirmation split? 2.Do on-chain prices retain significant incremental predictive value after accounting for other publicly available off-hours signals, such as equity-index futures, overseas markets, premarket trading, or relevant news? 3.In the next-asset prediction task, how do the authors distinguish active purchases from ordinary transfers, airdrops, minting, custody movements, or market-making activity? Do the results remain valid when only verifiable purchase events are retained? 4.Does the association between news co-mentions and account-level co-holding remain significant after controlling for asset popularity, news coverage, trading activity, and token listing age?

5 comments

We thank the reviewer for recognizing the novelty of the dataset and the connection between our findings and the learning tasks, and for the helpful questions. We respond to each point below and have revised the paper accordingly, with changes to the text marked in blue.

W1: The central forecasting claim requires stronger validation...

Q1: For the reported directional accuracy...

A1: We thank the reviewer for raising this concern. We have revised Figure 4(b) to annotate each cell's sample size and to hatch cells too thin to support an estimate, and we have added the sample size and coverage of this region to Section 5.2.

1. Sample size. The above-80% region (more than 300 trades, quote no older than one hour) contains 1,917 windows over 270 dates and 32 equities. Its directional accuracy is 85.5% (95% CI [83.5%, 87.3%]), against 58.7% for always predicting an up move.

2. Coverage. The result does not depend on a few liquid names: accuracy is 84.6% without the top five tickers and 84.9% without the top ten. Only 34 of the 890 tickers ever exceed 300 trades in a window, but they account for 99.8% of swap volume.

3. Threshold selection. The cut-offs are the bin edges of a descriptive stratification and were set on the pooled sample, not on the discovery split alone. With the cut-offs held fixed, accuracy is above 80% on both splits: 82.3% on discovery and 87.9% on confirmation. The result is therefore not produced by selecting thresholds on the evaluation period.

W2: The event definition in the account-level task requires further validation...

Q3: In the next-asset prediction task..

A2: Thanks. We have described the filtering procedure more clearly in Section 6.2 and Appendix B, and we conducted an additional experiment that retains only verified purchases. Specifically,

(1) Current filtering. Filtering is applied to accounts, in two steps. We first exclude contract addresses, exchange wallets, issuers, and market makers, and then remove addresses flagged by six behavioural rules, such as mint/redeem addresses, same-block arbitrage, and one-to-many distributors (Appendix B). These rules remove 0.7% to 16.5% of wallets per chain. An acquisition is the first time a remaining account receives a token it does not hold. This removes the addresses that carry out minting, custody, and market-making.

(2) Additional experiment: verified purchases only. We further classify every acquisition event by the counterparty that delivers the token.

Method. For each event, we identify the counterparty that delivers the token in the same transaction and assign it to one of four channels. The labels combine three sources: address labels in the dataset and public exchange address lists; verified contract names for DEX pools, aggregators, and issuers; and the sender's behaviour over its full transfer history, where one-way batch senders are treated as airdrops or distributions and two-way high-degree hubs as custodial addresses. An event is left unlabelled when the evidence is insufficient. All thresholds were fixed before any model result was inspected.

Ondo and Binance are the two issuers on BSC. Event shares and counts split the BSC events by the issuer of the acquired token; MRR follows the folds of Table 5.

Table R8a. Share of acquisition events by sender.

ChainDEX pool / aggregator tradeCEX or custodial-hub withdrawalIssuer mintAirdrop / dividend / distributionUnlabelled
Solana42.4%9.6%3.8%23.5%20.8%
Binance29.7%18.7%0.0%41.1%10.5%
Ondo38.4%3.9%14.9%29.7%13.2%
Ethereum63.6%7.2%13.0%0.7%15.5%
RH-chain16.4%21.8%0.0%35.9%25.9%

As shown in Table R8a, airdrops and distributions account for 24% to 41% of acquisition events on Solana, Binance, Ondo, and RH-chain, and under 1% on Ethereum, so a substantial share of events on these four chains are not active purchases. Where our swap data cover the trading venues, the classification can be checked: on Solana and RH-chain, 93% and 96% of the events delivered by a DEX pool are confirmed by an independent swap record.

Q3 (continued).

(3) Results on verified purchases only. We define a verified purchase in two ways. Strict keeps an event only if its own transaction is a recorded DEX swap in which the account buys the token. Layered also keeps events delivered by an aggregator trade, an issuer mint, or an exchange withdrawal, where the channel indicates a purchase but no swap is recorded. Both exclude airdrops, distributions, and unlabelled events. We evaluate the same models, with the same folds and candidate sets, on the verified subset of the test set.

Table R8b. Events retained under each definition, and MRR on all test events → on the strict subset.

SolanaBinanceOndoEthereumRH-chain
All acquisition events98,99184,30348,91712,91222,323
Strict coverage39.0%1.9%0.7%6.7%12.0%
Layered coverage55.7%48.4%57.1%83.8%38.2%
EASE0.311 → 0.3150.068 → 0.0650.269 → 0.3090.184 → 0.2610.229 → 0.255
SASRec0.311 → 0.3200.066 → 0.0420.307 → 0.2160.234 → 0.3460.288 → 0.318
FDSA0.317 → 0.3140.073 → 0.0510.319 → 0.2590.237 → 0.3540.295 → 0.339
UniSRec0.314 → 0.3150.091 → 0.1380.339 → 0.2920.218 → 0.3390.297 → 0.357
DyGFormer0.233 → 0.2370.379 → 0.2910.286 → 0.2890.071 → 0.0820.139 → 0.175
CAWN0.293 → 0.3010.234 → 0.2740.273 → 0.2880.191 → 0.2980.178 → 0.158

Table R8c. MRR on all test events → on the layered subset.

SolanaBinanceOndoEthereumRH-chain
EASE0.311 → 0.3080.068 → 0.0510.269 → 0.1750.184 → 0.1830.229 → 0.254
SASRec0.311 → 0.3030.066 → 0.0490.307 → 0.1810.234 → 0.2270.288 → 0.292
FDSA0.317 → 0.3040.073 → 0.0570.319 → 0.1810.237 → 0.2320.295 → 0.305
UniSRec0.314 → 0.3090.091 → 0.0610.339 → 0.1830.218 → 0.2100.297 → 0.312
DyGFormer0.233 → 0.2350.379 → 0.4340.286 → 0.1610.071 → 0.0670.139 → 0.164
CAWN0.293 → 0.2960.234 → 0.2420.273 → 0.1770.191 → 0.1890.178 → 0.146

We summarize the main findings as follows. (1) On Solana, Ethereum, and RH-chain, the results hold. On Solana, every method changes by less than 3% on the strict subset. On Ethereum and RH-chain, strictly verified purchases are easier to predict for almost every method, and the layered results stay within about 0.03 of the original. (2) On Ondo and Binance, the layered definition is the informative one. Trades there go mostly through aggregators, the issuer, or the exchange and leave no swap record, so the strict subset keeps under 2% of events. On Ondo, all methods fall from 0.27–0.34 to 0.16–0.18 under the layered definition: the removed airdrops and distributions concentrate on a few assets and were the easiest events to predict. On Binance, the sequential models stay below 0.1 under the layered definition and DyGFormer remains the strongest. (3) Training on verified purchases gives the same result. On Solana, we retrain on strict events only and keep the test set unchanged: SASRec 0.311 → 0.320, UniSRec 0.314 → 0.325. Overall, the conclusions hold when only verified purchases are retained. On Ondo, airdrops and issuer distributions raise the absolute accuracy. We have added this analysis to Section 6.2 and the full results to Appendix B.

Q2: Do on-chain prices retain significant incremental...

A3: Thank you for this question. On-chain token prices and other off-hours signals react to the same overnight information, so some overlap is expected. On weekday nights (Tuesday–Friday), equity-index futures trade continuously and dominate overnight market-wide price discovery, so the market-wide part of token prices largely follows futures. Tokens can add information in two cases. During futures closures (weekends and holidays), they are the only instrument still trading. At the level of individual equities, futures and overseas indices price the market rather than single stocks, and news is text rather than a traded price. In the revision, we add this discussion to Section 5.2 and the full details to Appendix F.

Setup. We regress the opening gap y (open over previous close, minus one) on the on-chain signal s together with the off-hours controls X that are available before the 04:00 ET decision time: y = θ′X + γ·s + ε. The controls are the returns of ES and NQ index futures from the close to 04:00 ET (Yahoo Finance, hourly bars, each scaled by the equity's beta), the Nikkei 225 (daily close), overnight news sentiment and article counts, and the equity's own after-hours return. Premarket trading starts at 04:00 ET and is therefore not available. We run three tests:

  • Market level: s is the dark-window state. Weights are fitted on the discovery period and evaluated on the confirmation period (106 trading days, 52,300 instances).
  • Stock level: s is the return of the equity's own token, with a day fixed effect and standard errors clustered by day (959 instances over 200 nights and 28 equities since on-chain data became available).
  • Futures closures: y is the gap at which ES futures reopen and s is the return of the SPY token over the same closure (48 weekend closures since July 2025).

Findings. On weekday nights, futures dominate market-wide price discovery, and the market-wide part of token prices follows them. Adding the dark-window state to futures does not raise RankIC (0.219 to 0.226, t = 0.49). On-chain prices add information in two cases, where the other signals are silent:

  1. When futures are closed. Over weekends, tokens are the only instrument still trading. The return of the SPY token predicts the gap at which futures reopen (r = 0.42, permutation p = 0.003).
  2. At the level of individual equities. Futures and overseas indices price the market, not single stocks, and news is text, not a traded price. After removing everything common to all equities on a given day and adding all controls, the return of an equity's own token remains significant (γ = 0.46, t = 6.95). The result holds separately in the discovery and confirmation periods (t = 3.89 and t = 5.48). Coverage is currently about five equities per night.

Q4: Does the association between news co-mentions...

A4: Yes, the association remains positive and significant after controlling for all four factors.

We run a pair-level regression of co-holding on news co-mention, with asset popularity, news coverage, on-chain trading activity, token listing age, and return correlation as controls. Because pairs that share a ticker are not independent, we use standard errors clustered on both tickers and a QAP permutation test (2,000 permutations).

Table R9. Coefficient of news co-mention.

Co-holding definitionWithout controlsWith controls95% CI (dyadic-robust)QAP p
Ever co-held (count)0.4960.235[0.184, 0.286]< 0.001
Concurrently co-held (count)0.4850.227[0.172, 0.282]< 0.001
Ever co-held (ratio)0.03550.0222[0.017, 0.027]< 0.001
Concurrently co-held (ratio)0.02670.0173[0.013, 0.021]< 0.001

As can be seen, the association remains significant under all four definitions of co-holding. Every confidence interval excludes zero, and none of the 2,000 permutations reaches the observed coefficient. The controls do absorb part of it, about half on the count outcomes and a third on the ratio outcomes, with popularity the strongest among them on the count outcomes. The remaining effect is therefore not a proxy for popularity or coverage. We have added this analysis to Section 5.3 and Appendix E.

Reviewer_Frs1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a ticker-aligned dataset connecting tokenized equities on four blockchains with traditional equity prices, news, and fundamentals. It reports empirical findings on off-hours trading, the informativeness of on-chain prices, and account co-holding patterns. It further introduces two benchmark tasks: predicting overnight equity returns using dark-window on-chain signals, and predicting an account’s next asset acquisition in an evolving asset universe. The experiments suggest that on-chain signals improve cross-sectional equity ranking, while newly introduced assets pose substantial challenges for ID-based recommendation methods.

Critical strengths

S1: A useful and timely data setting. Connecting public account-level blockchain activity with conventional equity information creates opportunities for research that neither source supports independently. The dataset is potentially valuable for both financial forecasting and temporal recommendation.

S2: Well-motivated empirical analysis. The distinction between liquid, fresh quotes and sparse, stale observations is sensible. The co-holding analysis also offers an interesting behavioral perspective beyond price prediction.

S3: Complementary benchmark tasks. The two tasks cover both market-level applications and account-level behavior. Chronological evaluation and the inclusion of several model families are positive aspects of the experimental design.

Critical weaknesses

W1: Inconsistencies in the dataset and evaluation description. The introduction reports 140.1 million minute bars, whereas Section 4 reports 156.7 million. More importantly, the discovery period begins in September 2024, while Table 2 places the earliest on-chain observations in June 2025, and Section 4 states that aligned coverage begins at each asset’s on-chain launch. The manuscript does not explain how the earlier period is used or how missing on-chain signals are handled. These issues make it difficult to reconstruct the evaluation timeline.

W2: The forecasting experiments do not sufficiently isolate the source of the gains. The dark-window state combines an asset’s own token return, a market proxy, and a theme adjustment, but their individual contributions are not established in the provided manuscript. Simple on-chain-only and linear residual baselines, together with component ablations, would help determine how much of the improvement comes from the additional information versus the correction mechanism. The results support incremental ranking value, but market-level improvements are mixed.

W3: Missing details of coverage and uncertainty. The reported directional accuracy above 80% applies to windows selected for high activity and quote freshness. The manuscript does not clearly establish the sample size, asset concentration, and uncertainty for this specific subset. Since swap coverage is substantially narrower than transfer coverage, it is also unclear how broadly this result applies across the advertised asset universe.

W4: Evaluation is restricted to accounts with exactly one previously acquired item that subsequently acquire a different item. This is a valid task, but it does not establish performance on richer account histories or general portfolio evolution. The cold-start interpretation would also be stronger with explicit results separating assets seen during training from newly introduced assets, alongside simple popularity and recency baselines.

Questions for authors

Please address the points that directly correspond to the concerns raised in the Weaknesses section. No new experiments are required. I would consider a higher rating if the authors could clarify these.

5 comments

We thank the reviewer for recognizing the value of the data setting, the empirical analysis, and the two complementary tasks, and for the detailed and constructive comments. We respond to each weakness below and have revised the paper accordingly, with changes to the text marked in blue.

W1: Inconsistencies in the dataset and evaluation description...

A1: We thank the reviewer for pointing out these inconsistencies. We have corrected the bar count and now state the timeline and the handling of missing on-chain signals explicitly in Sections 4 and 6.1 of the revised manuscript. Specifically:

1. Minute-bar count. 140.1 M is the correct figure, while 156.7 M mistakenly corresponded to a later data snapshot. 140.1 M counts the full price corpus after de-duplication: 890 tickers, all available history through August 2026, of which 82.6% falls in regular trading hours. The revision reports this one scope consistently in the Introduction and Section 4.

2. The two start dates describe different data layers. The on-chain layer starts in June 2025, and each asset is covered from its own first on-chain activity through August 2026. The off-chain layer (minute prices, news, and fundamentals) starts as early as September 2024 to give the forecasting backbones a longer training history.

3. How the earlier period is used. Discovery runs from September 2024 to March 2026 and confirmation from March to August 2026 (n = 106 trading days). The earlier period is used only to train and validate the backbones. All reported numbers are computed on the confirmation split, which lies entirely inside the on-chain period. Where on-chain data does not yet exist, the dark-window state is masked to zero and never imputed.

W2: The forecasting experiments do not sufficiently isolate the source of the gains...

A2: Thanks. We have added the on-chain-only and linear residual baselines and a component ablation of the dark-window state. All results use the same confirmation split and target as Table 4. Specifically,

1. On-chain-only and linear residual baselines. On-chain only uses the state s directly as the prediction, with no off-chain model. Original method (Table 4) retrains the backbone on y − s and adds s back with a fixed coefficient of one. Linear residual fits y = a + c·ŷ_off + b·s on the validation period and applies it unchanged. Adaptive combination is a new method that weights the ranks of the state and of the model, with a weight chosen each day from past days only, so it can fall back on the state alone.

Table R4. RankIC on the confirmation split.

ModelOff-chain onlyOriginal method (Table 4)Linear residualAdaptive combination
MSGCA0.0820.1580.1410.196
GS-FUSE0.0450.0700.1820.194
StockMixer0.0570.1260.1900.190
MASTER0.0150.0470.1760.189
CI-STHPAN0.0150.0380.1710.189
DoubleAdapt0.0430.1100.1680.188
MAN-SF0.0260.0480.1540.184
On-chain only0.189

As shown in Table R4, the improvement comes from the additional information rather than the correction mechanism. The on-chain-only baseline (0.189) exceeds every backbone, with or without the state under the original method. The linear residual and the adaptive combination recover most of this signal, but neither exceeds the on-chain-only baseline by a meaningful margin.

2. Component ablation. If an equity has a reliable token quote, s is the token's own return; otherwise s is a market proxy, the equity's beta times the on-chain market move of that night. The theme adjustment applies only to three sector ETFs outside the S&P 500, so it is never active in this benchmark; we have removed it from Section 6.1.

Table R5. Source of the state on the confirmation split.

Source of the stateCoverageRankIC
Market proxy97%0.186
Own token return1%0.331
No state2%NA

As shown in Table R5, the market proxy accounts for almost all of the benchmark-wide gain, because it covers 97% of instances: using the market proxy alone gives a RankIC of 0.184, against 0.189 for the full state. The own token return is the stronger signal where it exists. On the same equities, the market proxy reaches only 0.029. This comparison uses about four equities per day, so it is indicative rather than precise.

3. Market-level results. We agree that the market-level results in the submitted Table 4 were mixed: market correlation rose for six backbones and fell for MSGCA. Since the task is motivated by the direction of the next opening, we now report directional and balanced accuracy instead; balanced accuracy improves for all seven backbones.

In the revision, we add the on-chain-only baseline to Table 4, present it together with the adaptive combination in Section 6.1, and report the remaining results in Appendix H.

W3: Missing details of coverage and uncertainty...

A3: We thank the reviewer for this comment. We have revised Figure 4(b) to annotate each cell's sample size and to hatch cells too thin to support an estimate. We have also added more details to Section 5.2, summarized as follows.

1. Sample size and uncertainty. The above-80% region (more than 300 trades, quote no more than one hour) contains 1,917 windows over 270 dates and 32 equities. Its directional accuracy is 85.5%, with a date-clustered 95% confidence interval of [83.5%, 87.3%]. Three cells of Figure 4(b) contain 16 windows or fewer; they are now hatched and not used in the text. The estimate is therefore well supported: even the lower bound of the interval is above 80%, and far above the 58.7% obtained by always predicting an up move.

2. Asset concentration. The five most frequent tickers account for 55.5% of these windows, but the result does not depend on them. Accuracy is 84.6% without the top five and 84.9% without the top ten. The result is thus not driven by a few heavily traded names.

3. Coverage across the asset universe. Of the 890 tickers, 216 have swap data and only 34 ever exceed 300 trades in a window. These 34 account for 99.8% of swap volume. The subset is therefore narrow in ticker count but covers almost all trading, and the revision presents the result as a property of liquid, freshly quoted windows, not of the whole tokenized-asset universe.

W4: Evaluation is restricted to accounts with...

A4: Thanks. To examine richer account histories and to make the cold-start interpretation explicit, we conducted two additional experiments, both with simple popularity and recency baselines. They use the same 15 temporal folds and candidate sets as Table 5 and are reported per chain. Most-Popular ranks assets by first acquisitions before the training cutoff, and Recent-Popular by first acquisitions in the last 7 days. We have added main findings to Section 6.2 and the full results, including both baselines, to Appendix G.

1. Performance on richer account histories. We extend the test set to every acquisition by an account that previously acquired h distinct assets, grouped as h = 1, 2, 3–4, and ≥5. Histories use strictly earlier acquisitions only.

Table R6. MRR of the best learned model / the best popularity baseline in each cell. Per-model results are in Appendix G.2.

HistoryOndoBinanceSolanaEthereumRH-chain
h = 1.323 / .303.122 / .144.313 / .296.242 / .235.289 / .252
h = 2.203 / .187.172 / .201.343 / .309.237 / .233.330 / .351
h = 3–4.176 / .148.207 / .194.290 / .284.211 / .208.298 / .303
h ≥ 5.129 / .105.173 / .170.270 / .256.148 / .135.332 / .326

As shown in Table R6, we have two main findings. (1) Accuracy falls with history length on Ondo, Solana, and Ethereum, because the targets become harder. The decline appears for the popularity baselines as well as the models, since accounts with long histories acquire less popular assets. (2) The gain over popularity depends on the chain. Learned models exceed the best baseline in 16 of 20 cells: at every history length on Ondo, but only for accounts with three or more assets on Binance.

W4 (continued).

2. Seen versus newly introduced assets, with popularity and recency baselines. We split the 220,837 test events by whether the target asset appears in the fold's training data: 187,996 are seen and 32,841 are unseen, 85% of the latter on Binance. Both baselines are computed once at the cutoff, like the other static models.

Table R7. Seen and unseen targets, MRR.

MethodOndoBinanceSolanaEthereumRH-chain
SeenMost-Popular.239.118.303.273.252
Recent-Popular.307.199.310.190.252
EASE.266.095.324.214.229
SASRec.304 ± .009.097 ± .018.324 ± .004.274 ± .009.306 ± .023
FDSA.320 ± .006.101 ± .002.330 ± .002.274 ± .008.300 ± .012
UniSRec.341 ± .021.130 ± .005.330 ± .002.268 ± .012.289 ± .003
LLM-ESR.282 ± .022.167 ± .017.326 ± .002.242 ± .006.255 ± .017
LLMRank†.191 ± .003.119 ± .000.213 ± .000.197 ± .003.207 ± .001
DyGFormer†.284 ± .052.472 ± .017.240 ± .020.078 ± .014.139 ± .034
CAWN†.274 ± .091.285 ± .019.305 ± .002.219 ± .005.178 ± .081
UnseenStatic methods≈ 0≈ 0≈ 0≈ 0n/a
LLMRank†.009 ± .000.016 ± .000.023 ± .001.018 ± .002n/a
DyGFormer†.009 ± .001.281 ± .011.051 ± .011.024 ± .003n/a
CAWN†.009 ± .001.200 ± .121.026 ± .002.019 ± .003n/a

†Updated before each test week; all other methods use only data up to the training cutoff. RH-chain has only 2 unseen events.

As shown in Table R7, we have three main findings. (1) How well popularity works depends on how fast demand shifts. On Binance, popular assets turn over quickly, so cumulative popularity is outdated by the test period (0.118) and recent popularity is much stronger (0.199). On Ethereum, where the asset set changes slowly, the opposite holds (0.273 against 0.190). (2) On seen assets, learned models add the most where popularity explains the least. FDSA and UniSRec exceed both static baselines on Ondo (UniSRec 0.341 against 0.307), the models are close to the baselines on Solana, Ethereum, and RH-chain, and on Binance static recency exceeds every static model. Adding popularity features to the models confirms this: it raises all four ID-based models on Binance (by 0.065 to 0.113) and lowers the three sequential models on Ondo (by 0.021 to 0.046), where the models already capture more than popularity. (3) Cold start is mainly a recency problem. Static methods cannot rank a newly introduced asset: their MRR on unseen targets is close to zero on every chain. The temporal graph models, which are updated every week, can. On Binance, where 85% of the unseen targets are, they reach 0.281 and 0.200.

Reviewer_AC4EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a large-scale, time-aligned dataset, which correlates the public chain activities of 890 tokenized stocks on four blockchains with traditional stock market data. By utilizing this alignment mechanism, the authors have made empirical discoveries on the activities during non-trading periods and the information content on the chain, and have proposed two tasks: (i) Dark Window Stock Prediction, which adds a frozen pre-market chain dark window state into the offline prediction factors; (ii) Next Asset Prediction in an Open World, which is based on the account-asset interaction behaviors over the time dimension. A large number of experiments have shown that introducing the chain state can continuously improve the stock ranking IC value, while the next asset ranking is faced with real cold start and market institutional transformation challenges.

Critical strengths
  1. The problem setting, dataset construction, and alignment strategies are all clearly explained, along with an intuitive interpretation that helps understand why the signals on the blockchain are important.

  2. This study focuses on a timely and increasingly important area, namely real assets and tokenization. Currently, high-quality and reproducible datasets are extremely scarce.

Critical weaknesses
  1. Section 4 states that the on-chain data began from June 10, 2025, and the data coverage period for each ticker starts from the first on-chain activity of that ticker. However, Section 6.1 reports that the Discovery phase began in September 2024 and lasted until March 2026. This is confusing. 
  2. The Introduction section reports 140.1 million one-minute stock price records, while Section 4 reports 156.7 million. The author needs to verify these two figures and their respective statistical scopes. 
  3. The paper is mainly motivated by predicting the direction of the next opening, but Table 4 mainly reports Market Correlation and RankIC. 
  4. For the dark window prediction, there is currently no clear benchmark method that only uses the dark window state to model the incremental values in complex off-chain models. For example, mapping si,ts_{i,t} to yi,ty_{i,t} with a simple linear or equidistant mapping.
Questions for authors

I would greatly appreciate it if the author could answer some of the questions I have encountered in Critical Weaknesses.

2 comments

We thank the reviewer for recognizing the timeliness of tokenized equities and the clarity of our dataset construction and alignment, and for the helpful suggestions. We respond to each point below and have revised the paper accordingly, with changes to the text marked in blue.

Q1: Section 4 states that the on-chain data began from June 10, 2025....

A1: We thank the reviewer for pointing this out. The two periods describe different data layers, and we now state this explicitly in Sections 4 and 6.1 of the revised manuscript. Specifically:

  1. June 2025 is the start of the on-chain layer. The first tokenized-equity event is in June 2025, and each asset is covered from its own first on-chain activity through August 2026.
  2. September 2024 is the start of the off-chain layer. We extend the off-chain layer to before the on-chain layer to give the forecasting backbones a longer training history, since the on-chain period alone is too short for training. The backbones use only off-chain features, so discovery can begin in September 2024.
  3. Where on-chain data does not yet exist, the dark-window state is masked to zero and never imputed. All reported numbers are computed on the confirmation split (March–August 2026), which lies entirely inside the on-chain period. All components of the state are frozen before listed premarket prices become available, so no future information enters the forecast.

Q2: The Introduction section reports 140.1 million ...

A2: We clarify that 140.1 M is the correct figure: it counts the full price corpus after de-duplication, covering 890 tickers and all available history through August 2026, of which 82.6% falls in regular trading hours. The 156.7 M in Section 4 mistakenly corresponded to a later data snapshot. The revision now reports this one scope consistently.

Q3: The paper is mainly motivated by predicting the direction of the next opening, but Table 4 mainly reports Market Correlation and RankIC.

A3: We reported RankIC and market correlation because they are widely used metrics in equity forecasting and give a comprehensive view of overall forecast quality: they use the magnitude and ordering of the predicted gaps across the whole cross-section. We agree that neither measures directly whether the predicted direction of the next opening is correct.

We have therefore additionally reported Directional Accuracy (DA) and Balanced Accuracy (BA) on the full benchmark. DA is the share of instances where the predicted and realized gaps have the same sign. BA is the mean of the recall on up gaps and on down gaps.

Table R1. Confirmation split, same rows and target as Table 4.

ModelDA (off-chain)DA (+ on-chain)BA (off-chain)BA (+ on-chain)
MASTER0.5061 ± 0.00270.5172 ± 0.00350.5047 ± 0.00640.5122 ± 0.0044
DoubleAdapt0.5120 ± 0.00500.5306 ± 0.00320.5162 ± 0.00430.5205 ± 0.0041
StockMixer0.5245 ± 0.00190.5418 ± 0.00390.5136 ± 0.00490.5321 ± 0.0054
CI-STHPAN0.5009 ± 0.00660.5072 ± 0.01020.5027 ± 0.00510.5089 ± 0.0087
MAN-SF0.5099 ± 0.01040.4850 ± 0.02130.5064 ± 0.00220.5123 ± 0.0092
MSGCA0.5346 ± 0.01370.5296 ± 0.00630.5213 ± 0.01820.5290 ± 0.0134
GS-FUSE0.4998 ± 0.01320.5127 ± 0.02420.5232 ± 0.00610.5288 ± 0.0099

As can be seen, BA improves under on-chain augmentation for all seven backbones, by 0.83% to 3.60% in relative terms, and DA improves for five of the seven, by 1.26% to 3.63%. For MAN-SF and MSGCA, DA decreases while BA and RankIC both increase, because DA weights up and down gaps by their frequency while BA weights them equally. In the revision, Table 4 now reports RankIC, DA, and BA for each setting, and Section 6.1 defines the two new metrics.

Q4: For the dark window prediction, ...

A4: Thanks for this suggestion. We have added the requested benchmark, which uses the dark-window state only. Inspired by its strength, we further propose a new way to combine the state with the off-chain models.

1. Benchmark using only the dark-window state. We map si,ts_{i,t} to yi,ty_{i,t} with the simplest mapping: si,ts_{i,t} is used directly as the prediction, with no fitted parameter. Any increasing linear mapping gives the same ranking, so the RankIC is identical. We compare it with each model without the state (Off-chain) and with the state under our original method, in which the model is retrained on the residual target yi,t−si,ty_{i,t}-s_{i,t} and si,ts_{i,t} is added back.

Table R2. Confirmation split.

MethodRankICDABA
Off-chain models (seven)0.015 – 0.0820.500 – 0.5350.503 – 0.523
Off-chain models + state, original method0.038 – 0.1580.485 – 0.5420.509 – 0.532
Dark-window state only0.1890.5120.508

As shown in Table R2, the dark-window state alone is a strong benchmark. Its RankIC of 0.189 exceeds every off-chain model, with or without the state under our original method. On direction, the state alone is comparable to the models (DA 0.512, BA 0.508).

2. A new combination method. Inspired by this benchmark, we propose a combination method that can rely on the state alone. Each day, we rank equities by the state and by the model's predictions, and score each equity by w⋅rank(s)+(1−w)⋅rank(model)w\cdot\mathrm{rank}(s)+(1-w)\cdot\mathrm{rank}(\mathrm{model}). The weight ww is chosen each day as the value with the highest average RankIC over all earlier days, so it uses past data only and reduces to the state alone (w=1w=1) when the model does not help.

Table R3. RankIC on the confirmation split.

ModelOff-chain onlyOriginal method (Table 4)New method
MSGCA0.0820.1580.196
GS-FUSE0.0450.0700.194
StockMixer0.0570.1260.190
MASTER0.0150.0470.189
CI-STHPAN0.0150.0380.189
DoubleAdapt0.0430.1100.188
MAN-SF0.0260.0480.184
Dark-window state only0.189

As shown in Table R3, the new method raises RankIC for all seven models to between 0.184 and 0.196, the level of the benchmark. This shows that the dark-window state gives strong signals for the next opening gap, and that these signals benefit all seven models regardless of their backbone. In the revision, we add the state-only benchmark to Table 4, present the benchmark and the new method in Section 6.1, and report the details in Appendix H.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.