acceptodds
Withdrawn from ICLR 2027

RICE-Alpha: Reliability-Informed Correction with Event Graphs for LLM-Agent Stock Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Equity-relevant news evolves through temporally dependent corporate events rather than isolated observations, making historical information useful only when event continuity, information availability, and transition reliability are modeled explicitly. Existing LLM-based financial agents increasingly incorporate historical evidence, yet they provide limited support for preserving issuer-specific chronology under point-in-time constraints and for identifying when historical transitions contribute information beyond the current forecast. We present RICE-Alpha (Reliability-Informed Correction with Event Graphs), a point-in-time stock-scoring framework that separates a history-aware multi-view Base Alpha from a reliability-calibrated residual correction derived from historical event continuation. A Multi-Tier Memory Layer grounds news interpretation in temporally eligible issuer-specific history, while a Typed Event Agent constructs event states whose successor relations are formed within issuers and pooled across firms only after valid local pairing. Matured transitions are calibrated by their empirical reliability, and the resulting graph signal is residualized against the Base Alpha and technical view to obtain the RICE Delta. On daily Nasdaq-100 and Hang Seng Index panels from 2024 to 2026, RICE-Alpha achieves the strongest results among the evaluated LLM-based agents and momentum across four predictive and four portfolio-level metrics in both markets. Its ICIR more than doubles that of the strongest baseline, while net Sharpe ratios reach 1.656 and 1.725 in the U.S. and Hong Kong, respectively. U.S. ablations further show significant reductions in IC and RankIC after Holm adjustment when major components are removed. These results indicate that historical event continuation adds incremental information when it is temporally grounded, reliability-calibrated, and introduced as a residual correction to a multi-view forecast.

Primary area:
learning on time series and dynamical systems
Keywords:
stock forecasting, large language model agents, structured memory, temporal event graphs, point-in-time prediction, reliability-calibrated correction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_jRvgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RICE-Alpha combines historical memory, multi-agent analysis, and an event graph correction for cross-sectional stock ranking. Sentiment, technical, fundamental, and macroeconomic views first produce a Base Alpha. The system constructs successor relations within individual issuers and then pools comparable transitions across firms. It weights graph evidence using historical support, relative transition frequency, lag structure, and matured successor returns. Residualizing this evidence against Base Alpha and the technical view produces a separate correction for the final score. The authors evaluate the framework on Nasdaq-100 and Hang Seng Index constituents from 2024 to 2026. The complete system achieves the best reported predictive and portfolio metrics among the evaluated methods. U.S. ablations, an alternative LLM backbone, constituent exclusions, and transaction-cost stress tests provide additional evidence for the contribution and robustness of the components.

Critical strengths
  1. Careful constraints on the use of event history. The distinction between NEW, UPDATED, and CARRIED records prevents continued coverage without new facts from creating duplicate occurrences. Successor relations are formed within issuers before pooling across firms, and observations enter graph estimates only after complete successor windows and relevant returns have matured. These rules address concrete problems in financial news backtesting. Keeping graph evidence as a separate correction also makes its incremental contribution relative to the base forecast easier to examine.

  2. Broad evaluation with component comparisons. Both markets use the same frozen hyperparameters, portfolio rules, and cost assumptions. In the U.S., the complete system achieves IC 0.0350, compared with 0.0191 for the memory-only configuration and 0.0260 for direct graph correction. Separate reliability-only and residualization-only variants help distinguish the two operations. Alternative-backbone, constituent-exclusion, and cost analyses further support the reported results.

Critical weaknesses
  1. The conclusion and Appendix F explicitly acknowledge that both pretrained backbones may contain information that postdates individual forecasts. The paper does not experimentally assess the effect of this information on the historical backtest. Restricting prompt inputs controls the evidence explicitly supplied to a model, but does not exclude prior knowledge of companies, events, or subsequent developments. Replacing the backbone with another model that may also contain later information does not resolve this concern. Source-data availability also needs clarification. Appendix D.1 admits U.S. financial reports using a fixed 90-day delay and describes Hong Kong financial snapshots within a historical window ending before the prediction date. It does not fully establish that the original publication timestamps and contemporaneous data versions were used. A fiscal period preceding a prediction date does not by itself establish that the corresponding information was publicly available at that time. Because temporal validity is central to the paper, these uncertainties affect the interpretation of the reported predictive gains. A backbone with a training cutoff preceding the test period, a prospective evaluation after model freezing, or targeted contamination controls would strengthen the evidence. The manuscript does not demonstrate that leakage occurred, but it also does not adequately exclude its influence.

  2. Table 2 supports numerical graph correction, reliability weighting, and residualization within this system. It does not directly test the importance of forming issuer-local successor pairs before pooling across companies. Controls that preserve event frequencies and sample sizes while shuffling within-issuer temporal order, or event statistics without successor ordering, would help distinguish gains from genuine event continuation from gains associated with event categories and historical return summaries. The authors appropriately limit claims of superiority over prior event graph methods. Nevertheless, the absence of comparisons under a common protocol limits the assessment of the methodological contribution. Relevant candidates include TRACE, which combines graph structure with textual evidence, and CausalStock, which incorporates lag-dependent relations and news information. Statistical comparisons do not cover the strongest baseline. The U.S. IC advantage over momentum is only 0.0034, from 0.0316 to 0.0350. Table 9 reports paired baseline tests against AI Hedge Fund, but not momentum, MEME, or R&D-Agent-Quant. A significant positive IC for the complete system and a nonsignificant IC for a baseline do not establish a significant difference between the methods. Since the validity rules permit different stock subsets across methods, the comparison should also be confirmed on common stocks and dates.

Questions for authors
  1. Model temporal validity. What are the exact backbone versions and known training cutoff dates? Are there results from a test period after the models' knowledge cutoffs, or controls that assess the influence of historical event memorization?
  2. Data availability and configuration selection. Were financial reports and news reconstructed using actual publication timestamps, time zones, and contemporaneous versions? How is public availability guaranteed for Hong Kong financial snapshots? What historical replay and parameter-selection procedure is meant by freezing the configuration at the end of the 2023 warm-up?
  3. Contribution of event ordering. How does performance change when within-issuer temporal order is shuffled while preserving event category frequencies, support, and return sample sizes? This would help establish whether succession structure provides additional information.
  4. Alignment with the prediction horizon. Equation (9) estimates returns after successor event B occurs, whereas the prediction target starts at the next open after the current cutoff. How does the lag weighting in Equation (11) account for this difference? For CARRIED events that have already persisted for several days, is event age used to adjust the remaining successor probability?
6 comments
AuthorsOpenReview ↗

Q1 & W1 (Backbone versions, training cutoffs, and model knowledge):

We agree that backbone look-ahead requires direct validation. Neither submitted backbone has a documented cutoff ruling out look-ahead (both were released during/after the evaluation period). We therefore replayed the full pipeline with an earlier-cutoff Llama backbone and added entity masking (Llama, Qwen3-Max), a post-Llama-release window, and a post-submission extension. As shown in Table J1, the RICE-Delta increment remains positive and comparable across all Llama and DeepSeek settings, with significant full-window gains in both markets (wider intervals in shorter windows) and negligible masking effects. Although Llama has lower absolute metrics than DeepSeek, its US IC and ICIR remain within the range of the two submitted backbones.

Exact versions & cutoffs:

  • DeepSeek (main): API deepseek-v4-flash, build DeepSeek-V4-Flash-0731 (rel. 2026-07-31); no model-specific cutoff (public summary dates some data to ~May 2025 and web crawl to ~2024–Dec 2025, without a post-training cutoff).
  • Qwen3-Max (Table 2): API qwen3-max, snapshot qwen3-max-2026-01-23 (family rel. Sep 2025); no documented cutoff.
  • Llama (replay): meta-llama/Llama-3.1-70B-Instruct (rev 1605565b47bb9346c5515c34102e054115b4f98b); pretraining cutoff Dec 2023; rel. 2024-07-23.
  • Memory encoder (submitted only, unused in replay): BAAI/bge-m3 (rev 5617a9f61b028005a4858fdac845db406aefb181; rel. 2024-01-30); no documented cutoff.

Table J1. Backbone, window, and entity-masking controls (metrics as Table 1). Increment = RICE-Alpha IC − Base Alpha IC (RICE-Delta). Masking replaces tickers, names, and aliases with placeholders across all LLM roles. Qwen3-Max (2nd backbone) covers US predictive metrics only (masked run is new).

MktWindow or controlLLMICICIRRankICRankICIRARR (%)SharpeMDD (%)Base ICIncrement [LB]
USMom. (12–1), fullnone0.03160.12740.03220.132418.940.88921.46
USRICE-Alpha, full (Tabs 1,2,9)DeepSeek0.03500.29280.03600.292730.861.65612.090.0191+0.0159 [+0.0091], p 0.0002
USRICE-Alpha − Mom., full: Δ [LB]DeepSeek+0.0034 [−0.0171]+0.1654 [+0.0810]
USRICE-Alpha, full (Tab 2)Qwen3-Max0.02810.19580.02910.2012
USRICE-Alpha, full, maskedQwen3-Max0.02740.19020.02850.1961
USMasked − unmasked (95% CI)Qwen3-Max−0.0007 (−0.0039, +0.0025)
USRICE-Alpha, fullLlama0.02980.22650.03090.230425.361.40214.370.0164+0.0134 [+0.0058], p 0.0018
USRICE-Alpha − Mom., full: Δ [LB]Llama−0.0018 [−0.0222]+0.0991 [+0.0124]
USRICE-Alpha, from 2024-07-24 (post-rel.)Llama0.02930.22310.03040.227024.581.37114.370.0161+0.0132 [+0.0044]
USRICE-Alpha, full, maskedLlama0.02910.22100.03020.224824.711.36814.520.0158+0.0133 [+0.0057]
USMasked − unmasked (95% CI)Llama−0.0007 (−0.0040, +0.0026)
USMom. (12–1), 2026-03-31–2026-10-23 (post-sub.)none0.02010.08320.02140.088912.840.70213.92
USRICE-Alpha, 2026-03-31–2026-10-23 (post-sub.)Llama0.02850.21660.02960.220323.121.2979.840.0158+0.0127 [−0.0023]
HKMom. (12–1), fullnone0.01600.06450.03270.131223.101.09813.23
HKRICE-Alpha, full (Tabs 1,2,9)DeepSeek0.02830.19860.03950.251436.641.72511.830.0195+0.0088 [+0.0003], p 0.0446
HKRICE-Alpha − Mom., full: Δ [LB]DeepSeek+0.0123 [−0.0129]+0.1341 [+0.0219]
HKRICE-Alpha, fullLlama0.02460.17020.03460.210530.121.48812.960.0168+0.0078 [+0.0007], p 0.0360
HKRICE-Alpha − Mom., full: Δ [LB]Llama+0.0086 [−0.0167]+0.1057 [−0.0071]
HKRICE-Alpha, from 2024-07-24 (post-rel.)Llama0.02410.16680.03390.206329.051.45212.960.0166+0.0075 [−0.0007]
HKRICE-Alpha, full, maskedLlama0.02400.16620.03380.205629.441.45513.100.0163+0.0077 [+0.0006]
HKMasked − unmasked (95% CI)Llama−0.0006 (−0.0050, +0.0038)
HKMom. (12–1), 2026-03-31–2026-10-23 (post-sub.)none0.01120.04580.02360.096115.310.81210.44
HKRICE-Alpha, 2026-03-31–2026-10-23 (post-sub.)Llama0.02210.15290.03100.189127.351.3829.120.0150+0.0071 [−0.0070]
AuthorsOpenReview ↗

Q2 & W1 (Data availability and configuration selection):

We agree that verifying point-in-time data availability and clarifying the configuration-selection procedure are essential. We address both concerns below, supported by the source-level audit in Table J2.

  • Timestamps and versions. News is admitted based on provider publication timestamps (UTC converted to exchange time), with date-only records deferred to the next trading session. We audited every financial statement and news item used in the submitted run against publication records, including the exact version available at each forecast cutoff. The independent Llama replay reconstructs a document-versioned source panel from archived records.
  • Financial statements. US financial statements are admitted 90 calendar days after fiscal-period end, while HK snapshots use the latest fiscal period ending at least three months before the prediction date. The publication audit confirms that all admitted issuer-period results were publicly available before use. We will clarify the HK rule in Appendix D.1.
  • Freezing. The 2023 warm-up replays the same daily pipeline on historical data to initialize system state, without contributing to reported metrics. Parameters in Tables 5 and 7 were manually specified based on rule design and 2023 replay diagnostics (score scales, coverage, and cross-sectional stability). No grid search was performed, and no 2024–2026 forecasts were evaluated before the configuration was frozen.

As shown in Table J2, all financial statements and news items used in the submitted run were public before admission, no forecast used a version unavailable at its cutoff, and no forecasts required exclusion. Excluding all flagged stock-dates leaves IC nearly unchanged in both markets. The original Table 1 results remain unchanged.

Table J2. Source-availability audit of the submitted run and independent document-versioned replay. Flagged stock-dates involve a pending periodic report, financial line items differing from their first release, or subsequently revised news, regardless of whether the version used was already public at the forecast cutoff.

USHK
Submitted run (DeepSeek): every statement and news item checked against the publication record
Issuer–periods whose statements entered a forecast948437
Results published before admission948 (100.00%)437 (100.00%)
Days from publication to admission, median / shortest58 / 1124 / 1
Periodic report also on file with the SEC before admission945n/a
Line items equal to the first-published value99.6%99.7%
News items that entered a forecast148,112100,935
of which later revised by the provider1,173 (0.79%)868 (0.86%)
Forecast uses of a version not yet public at the cutoff00
Forecasts requiring exclusion00
IC without every flagged stock–date (share of the panel flagged)0.0352 (3.63%)0.0281 (3.33%)
Submitted IC (Table 1)0.03500.0283
Independent replay (Llama): source panel rebuilt from archives with full revision histories
News records in the archive152,384104,517
admitted at their own cutoff147,602 (96.86%)99,841 (95.53%)
admitted from the next session (date only)3,1073,318
not admitted (duplicates or no usable timestamp)1,6751,358
of the admitted records, served in the latest version public at the cutoff because a later revision exists1,212905
IC of the replay0.02980.0246
AuthorsOpenReview ↗

Q3 & W2 (Contribution of event ordering):

We conduct both requested controls and four additional variants, modifying only the numerical event-graph pathway while keeping Base Alpha and memory text fixed. Within-issuer type permutation preserves event category frequencies, pair counts, and return sample sizes; cross-issuer successors preserve per-edge support; and succession-free statistics retain each event type's matured return history without successor information. As shown in Table J3, all six controls lower IC in both markets with both backbones. The gain over succession-free statistics remains significant after Holm correction, demonstrating that successor structure contributes predictive information beyond event categories and their return histories. The succession-free variant itself retains positive IC, while incorporating successor structure provides a further improvement.

Table J3. Contribution of event ordering and controlled graph-construction alternatives (IC; extending Table 2). ΔIC denotes the improvement of local-first construction over each alternative, averaged equally across the two markets on shared valid dates (as in Table 9). The final two rows report two confirmatory contrasts designated before the controls were run, with Holm-adjusted p-values ((p_H)); individual-rule p-values are unadjusted. All variants recompute only the numerical correction, keeping Base Alpha and memory text fixed from the original run for both backbones.

Construction ruleDeepSeek US ICDeepSeek HK ICDeepSeek pooled ΔIC [LB]pLlama US ICLlama HK ICLlama pooled ΔIC [LB]p
Local-first (RICE-Alpha)0.03500.02830.02980.0246
Event types permuted within issuer (category frequencies, pair counts and return sample sizes preserved)0.02930.0240+0.0050 [+0.0006]0.02960.02500.0209+0.0043 [+0.0003]0.0366
Cross-issuer successors at identical per-edge support0.03030.0247+0.0041 [+0.0001]0.04740.02580.0215+0.0036 [−0.0001]0.0564
Predecessor-reversed: same pairs and returns, reversed query direction0.02990.0242+0.0046 [+0.0002]0.04240.02560.0211+0.0038 [−0.0001]0.0536
Pooled-first: events of all issuers pooled before pairing, so a successor may come from another issuer (same edge thresholds)0.02910.0238+0.0052 [+0.0005]0.03320.02480.0207+0.0045 [+0.0002]0.0408
Lifecycle-free: repeated coverage also creates occurrences0.03170.0251+0.0032 [−0.0005]0.07460.02710.0219+0.0027 [−0.0006]0.0918
Succession-free: each event type keeps its own return history, no successor information0.02850.0228+0.0060 [+0.0011]0.0210 (p_H 0.0260)0.02430.0199+0.0051 [+0.0007]0.0270 (p_H 0.0360)
Mean of the five rules other than succession-free0.03010.0244+0.0044 [+0.0012]0.0130 (p_H 0.0260)0.02570.0212+0.0038 [+0.0008]0.0180 (p_H 0.0360)
Reviewer_wGZZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies point-in-time stock ranking from current financial information and issuer-specific event history. RICE-Alpha combines a history-aware multi-agent Base Alpha with a reliability-weighted residual signal from an issuer-local temporal event graph. On U.S. and Hong Kong equity panels, it reports favorable prediction and portfolio results, supported by component ablations and robustness analyses. I find the problem relevant and the overall framework coherent, but the evidence for several specific design choices and for broader investment utility remains incomplete.

Critical strengths
  • Relevant problem and intuitive motivation. Linking current news to temporally eligible historical event continuations is meaningful for financial forecasting.
  • Coherent overall framework. The separation between multi-view Base Alpha and event-driven residual correction is understandable and reasonably motivated.
  • Useful empirical coverage. The study evaluates two markets and considers point-in-time eligibility, transaction costs, component ablations, and falling-market behavior.
Critical weaknesses
  1. The portfolio evaluation relies on a fixed investment rule. Stock ranking is a meaningful task, but actual investment decisions also involve trading directions, positions, and allocations. Applying the same long-only, equal-weighted top-10% rule to all methods ensures a fair comparison of their ranking signals, yet it also imposes predetermined trading decisions on every baseline. Accordingly, Panel B demonstrates performance only under this constrained setting. Additional strategy-level comparisons would better establish the method's practical investment value.
  2. The reliability of LLM-generated Base Alpha scores is not directly established. The paper does include normalization, confidence filtering, matured-feedback adjustment, and adaptive reconciliation. Nevertheless, it remains unclear whether the LLM's raw scores are comparable across stocks, whether self-reported confidence reflects realized forecast quality, and how much the feedback-based calibration contributes beyond raw scores. Direct calibration diagnostics would strengthen the foundation of Base Alpha.
  3. The novelty and specific benefit of issuer-local pairing require clearer justification. Prior work such as CausalStock and TRACE has already explored temporal relations and graph-enhanced stock prediction. The authors should therefore better explain what is genuinely new about their local-first pairing design and why it provides more informative historical relationships. From an empirical perspective, the C2 ablation removes numerical graph evidence altogether; it does not isolate local-first pairing against alternative graph-construction rules under an otherwise matched pipeline. Consequently, the contribution of this particular design is not yet established.
  4. The numerical correction pipeline relies heavily on hand-designed choices. Event filtering, reliability weighting, graph-signal clipping, and final fusion involve multiple manually specified thresholds, functions, and coefficients. Although such heuristics may be reasonable and their values are disclosed and frozen before evaluation, their individual roles and necessity are not sufficiently justified. The existing component ablations and transaction-cost tests do not show whether the reported gains depend on these particular settings. Targeted sensitivity analyses or simplified alternatives would help assess the robustness of the design.
  5. Baseline competitiveness and adaptation need stronger support. First, the simple 12–1 momentum signal outperforms all three evaluated LLM-agent baselines on every predictive metric in Panel A, raising concerns about whether these baselines were adequately and fairly adapted and tuned. Second, mapping agentic trading systems to stock scores and forcing them to follow the same portfolio rule may suppress their original decision-making capabilities. The evaluation should therefore also include representative classical or competitive end-to-end investment strategies, rather than relying exclusively on ranking-based portfolio construction.
Questions for authors
  1. What is the specific novelty of issuer-local pairing relative to existing temporal event-graph approaches? Can the authors isolate its benefit by comparing local-first pairing with pooled-first or other graph-construction rules while keeping the event representation and correction pipeline fixed?
  2. Are raw Base Alpha scores comparable across stocks, and does the LLM's self-reported confidence correlate with realized forecast quality? What incremental improvement comes from confidence filtering and feedback-based calibration?
  3. How were MEME, R&D-Agent-Quant, and AI Hedge Fund adapted and tuned for the common scoring protocol, especially given their weaker Panel A results than momentum? Can the authors also report comparisons with established or strong end-to-end investment strategies that retain their native portfolio decisions, rather than only the prescribed top-10% rule?
  4. How sensitive are the results to key hand-specified settings, such as reliability-saturation thresholds, graph-signal clipping bounds, and the final correction strength? Would a simpler scoring or fusion rule yield comparable performance?
6 comments
AuthorsOpenReview ↗

W1 (Fixed portfolio rule):

We agree that the original top-10% rule limits strategy-level evaluation, but retain it to isolate stock-ranking quality. To assess broader investment value, we evaluate four additional policies (top-20%/30%, rank-weighted with volatility targeting, and dollar-neutral long–short), R&D-Agent-Quant with its native TopKDropout rule, and two classical score-free benchmarks. As shown in Table W1, RICE-Alpha outperforms Momentum in return, Sharpe ratio, and drawdown across all matched policies, including long–short. Its original top-10% portfolio also outperforms the additional strategy-level benchmarks in both markets, demonstrating that the gains are not limited to the original investment rule.

Table W1. Alternative portfolio policies and strategy-level benchmarks. All strategies trade at the next open under the costs in Section 4.1, with additional annual borrow fees of 3% (US) and 4% (HK) for short positions. Top-10% results are reproduced from Table 1.

PolicyMethodUS ARR (%)US SharpeUS MDD (%)HK ARR (%)HK SharpeHK MDD (%)
Top 10%, equal weight (Table 1)RICE-Alpha30.861.65612.0936.641.72511.83
Top 10%, equal weight (Table 1)Momentum (12–1)18.940.88921.4623.101.09813.23
Top 20%, equal weightRICE-Alpha27.041.53113.0233.411.64812.31
Top 20%, equal weightMomentum (12–1)17.620.86419.8422.371.07113.88
Top 30%, equal weightRICE-Alpha24.181.41714.2630.481.51713.56
Top 30%, equal weightMomentum (12–1)16.550.82118.7721.021.01914.65
Rank-weighted top 20%, 15% volatility targetRICE-Alpha24.311.56110.4226.871.6529.71
Rank-weighted top 20%, 15% volatility targetMomentum (12–1)13.980.92815.2116.741.07010.98
Dollar-neutral long–short, top vs bottom 10%RICE-Alpha19.621.7808.9412.611.22610.37
Dollar-neutral long–short, top vs bottom 10%Momentum (12–1)3.860.28426.412.120.20622.07
Native ruleR&D-Agent-Quant, own TopkDropout rule13.050.67223.4118.660.93416.85
Score-free1/N over the index constituents12.460.64222.8519.870.88119.62
Score-freeMinimum variance (Ledoit–Wolf)15.730.97815.6417.210.96814.27
AuthorsOpenReview ↗

W2 & Q2 (Reliability of the LLM-generated Base Alpha scores):

We conduct three diagnostics to assess raw score comparability, confidence reliability, and feedback calibration. As shown in Table W2, each raw view exhibits positive ranking IC, and Base Alpha retains predictive information beyond stock-specific effects. Self-reported confidence is positively associated with realized ranking skill (significant except for the technical view in Hong Kong). The 2×2 factorial analysis confirms that confidence gating and matured feedback both improve IC, bringing the fused score to the IC of its strongest raw view with higher ICIR. The technical and fundamental views mainly contribute stability rather than mean IC.

Table W2. Base Alpha diagnostics. Raw-view scores are evaluated before calibration. ρ denotes the partial Spearman correlation between confidence and realized ranking skill, controlling for date and score extremity, with Holm-adjusted p-values (p_H) across the three views. Confidence gating and matured feedback are evaluated through their main effects on Base Alpha IC in a 2×2 factorial design, with memory context held fixed.

DiagnosticUSHK
(a) Raw sentiment view: IC0.01860.0188
(a) Raw technical view: IC0.00500.0061
(a) Raw fundamental view: IC0.00400.0038
(a) Base Alpha: IC (Table 2 for the US)0.01910.0195
(a) Base Alpha: IC after removing each stock's mean0.01740.0170
(a) Stock fixed effects: share of Base Alpha's variance (%)18.421.1
(b) Confidence–skill ρ (p; p_H), sentiment view0.0610 (0.0018; 0.0036)0.0550 (0.0062; 0.0124)
(b) Confidence–skill ρ (p; p_H), technical view0.0430 (0.0118; 0.0118)0.0220 (0.1796; 0.1796)
(b) Confidence–skill ρ (p; p_H), fundamental view0.0790 (0.0002; 0.0006)0.0710 (0.0002; 0.0006)
(c) Three views fused, no gate, no feedback: IC0.01060.0119
(c) Confidence gate: effect on IC [LB]+0.0031 [+0.0009]+0.0023 [+0.0002]
(c) Matured feedback: effect on IC [LB]+0.0054 [+0.0029]+0.0053 [+0.0028]
(c) ICIR: raw sentiment view0.11830.1062
(c) ICIR: Base Alpha (Table 2 for the US)0.14690.1345
AuthorsOpenReview ↗

W3 & Q1 (Novelty and isolated benefit of issuer-local pairing):

CausalStock learns lag-dependent cross-stock relations, while TRACE mines rules over entity-relation graphs. In contrast, RICE-Alpha constructs successor relations within each issuer's validated event chronology before cross-issuer pooling, preserving temporal consistency and using the resulting evidence as a residual correction. To isolate this design, we evaluate six alternatives, including pooled-first pairing, while keeping Base Alpha and memory context fixed. As shown in Table W3, local-first pairing achieves the highest IC among all variants in both markets, and both prespecified confirmatory tests remain significant after Holm correction.

Table W3. Local-first pairing versus alternative graph-construction rules (IC; extending Table 2). ΔIC denotes the equal-weighted cross-market IC improvement over each alternative on shared dates. The last two rows report prespecified confirmatory comparisons, with Holm-adjusted p-values (p_H); individual comparisons report unadjusted p-values. Only the numerical correction is recomputed, keeping Base Alpha and memory context fixed.

Construction ruleUS ICHK ICpooled ΔIC [LB]p
Local-first (RICE-Alpha)0.03500.0283
Pooled-first: events of all issuers pooled before pairing, so a successor may come from another issuer (same edge thresholds)0.02910.0238+0.0052 [+0.0005]0.0332
Cross-issuer successors at identical per-edge support0.03030.0247+0.0041 [+0.0001]0.0474
Lifecycle-free: repeated coverage also creates occurrences0.03170.0251+0.0032 [−0.0005]0.0746
Predecessor-reversed: same pairs and returns, reversed query direction0.02990.0242+0.0046 [+0.0002]0.0424
Event types permuted within issuer (category frequencies, pair counts and return sample sizes preserved)0.02930.0240+0.0050 [+0.0006]0.0296
Succession-free: each event type keeps its own return history, no successor information0.02850.0228+0.0060 [+0.0011]0.0210 (p_H 0.0260)
Mean of the five rules other than succession-free0.03010.0244+0.0044 [+0.0012]0.0130 (p_H 0.0260)

Comments on the paper

8 comments
AuthorsOpenReview ↗

Q1 & W1 (Backbone look-ahead):

We agree that backbone look-ahead requires direct validation. We run three controls (Table F1): (1) rerunning the full system from the 2023 warm-up with Llama-3.1-70B-Instruct (cutoff: Dec 2023), without external pretrained memory encoders or live retrieval; (2) masking issuer names, tickers, and aliases for Llama and Qwen3-Max; and (3) evaluating after Llama's release and after the submitted evaluation window. RICE Delta consistently yields positive IC increments with the pre-2024 backbone across both markets and all windows (significant over the full window; wider intervals in shorter windows). Entity masking produces small, insignificant IC changes for both backbones. Although Llama's US IC and RankIC are slightly below Momentum's, its ICIR and Sharpe remain higher, with a significant ICIR improvement, confirming RICE Delta's incremental contribution beyond backbone look-ahead.

Table F1. Backbone look-ahead controls across evaluation windows. Llama-3.1-70B-Instruct has a Dec 2023 cutoff. Inc = IC gain over Base Alpha; post-sub. = after the submitted window. Masking replaces issuer names, tickers, and aliases across all LLM roles. Qwen3-Max: US predictive metrics only.

MktWindow or controlLLMICICIRRankICRankICIRARR%SharpeMDD%Base ICInc [LB]
USMomentum (12–1), fullnone0.03160.12740.03220.132418.940.88921.46
USRICE-Alpha, full (Tables 1, 2, 9)DeepSeek0.03500.29280.03600.292730.861.65612.090.0191+0.0159 [+0.0091], p 0.0002
USRICE-Alpha − momentum, full: Δ [LB]DeepSeek+0.0034 [−0.0171]+0.1654 [+0.0810]
USRICE-Alpha, full (Table 2)Qwen3-Max0.02810.19580.02910.2012
USRICE-Alpha, full, entity names maskedQwen3-Max0.02740.19020.02850.1961
USMasked − unmasked (two-sided 95% CI)Qwen3-Max−0.0007 (−0.0039, +0.0025)
USRICE-Alpha, fullLlama0.02980.22650.03090.230425.361.40214.370.0164+0.0134 [+0.0058], p 0.0018
USRICE-Alpha − momentum, full: Δ [LB]Llama−0.0018 [−0.0222]+0.0991 [+0.0124]
USRICE-Alpha, from 2024-07-24 (post-Llama)Llama0.02930.22310.03040.227024.581.37114.370.0161+0.0132 [+0.0044]
USRICE-Alpha, full, entity names maskedLlama0.02910.22100.03020.224824.711.36814.520.0158+0.0133 [+0.0057]
USMasked − unmasked (two-sided 95% CI)Llama−0.0007 (−0.0040, +0.0026)
USMomentum (12–1), 2026-03-31–2026-10-23 (post-sub.)none0.02010.08320.02140.088912.840.70213.92
USRICE-Alpha, 2026-03-31–2026-10-23 (post-sub.)Llama0.02850.21660.02960.220323.121.2979.840.0158+0.0127 [−0.0023]
HKMomentum (12–1), fullnone0.01600.06450.03270.131223.101.09813.23
HKRICE-Alpha, full (Tables 1, 2, 9)DeepSeek0.02830.19860.03950.251436.641.72511.830.0195+0.0088 [+0.0003], p 0.0446
HKRICE-Alpha − momentum, full: Δ [LB]DeepSeek+0.0123 [−0.0129]+0.1341 [+0.0219]
HKRICE-Alpha, fullLlama0.02460.17020.03460.210530.121.48812.960.0168+0.0078 [+0.0007], p 0.0360
HKRICE-Alpha − momentum, full: Δ [LB]Llama+0.0086 [−0.0167]+0.1057 [−0.0071]
HKRICE-Alpha, from 2024-07-24 (post-Llama)Llama0.02410.16680.03390.206329.051.45212.960.0166+0.0075 [−0.0007]
HKRICE-Alpha, full, entity names maskedLlama0.02400.16620.03380.205629.441.45513.100.0163+0.0077 [+0.0006]
HKMasked − unmasked (two-sided 95% CI)Llama−0.0006 (−0.0050, +0.0038)
HKMomentum (12–1), 2026-03-31–2026-10-23 (post-sub.)none0.01120.04580.02360.096115.310.81210.44
HKRICE-Alpha, 2026-03-31–2026-10-23 (post-sub.)Llama0.02210.15290.03100.189127.351.3829.120.0150+0.0071 [−0.0070]
AuthorsOpenReview ↗

Q2 (What "evaluation-period selection" refers to):

We agree that the wording in Appendix F was imprecise. The evaluation window and submitted configuration were fixed using data through 2023, before examining 2024–2026 outcomes, and no part of the evaluation period served as a validation set (Appendix D.1; Table F2). Only online states were updated under prespecified rules using matured outcomes. The original statement was intended to acknowledge the limited generalizability of a single evaluation window, not post-hoc selection based on performance. All additional rebuttal analyses are reported as robustness checks without retuning the submitted RICE-Alpha configuration. Ex-post baseline score mappings are diagnostic only and do not affect the primary comparisons. We will clarify this distinction in Appendix F.

Table F2. Configuration freezing and evaluation-period controls.

ElementFixed usingDuring 2024–2026
Prompts, parameter values, update rules and score mappings (Tables 5 and 7)Data through 2023; the same values in both marketsUnchanged
Ablation configurations, portfolio rule and trading costsData through 2023Unchanged
Online state: memory, event statistics, feedback, view weightsInitialized by the 2023 warm-upUpdated by the fixed rules, only from outcomes that have matured
Evaluation window, 2024-01-02 to 2026-03-30Set once with the data panels (Table 8)Not chosen by looking at results; extended out of sample in Table F1
BackbonesBoth that were run are reported (Tables 1 and 2)Released during or after the window: addressed in Q1
Llama replay, controls and sensitivity analyses of this responseAdded after the reviewsEvaluate the frozen configuration; RICE-Alpha is not retuned

Replacement for Appendix F: The main evaluation uses a single fixed window (2024-01-02 to 2026-03-30), which may limit generalizability across market conditions. Neither the evaluation window nor the submitted configuration was selected or tuned using evaluation-period outcomes. Additional analyses extend the evaluation out of sample without retuning the submitted system. Potential backbone knowledge leakage is separately examined through an earlier-cutoff backbone, entity masking, and later-window evaluations.

AuthorsOpenReview ↗

Q3 (Long–short spread and factor-adjusted alpha):

We construct a dollar-neutral long–short portfolio (top vs. bottom decile, including trading costs and borrow fees) and conduct FF5 + momentum regressions for both long-only and long–short portfolios using DeepSeek and Llama. Tables 3 and 12 further show that RICE-Alpha captures a smaller fraction of market declines than gains, suggesting that lower beta alone does not fully explain its performance. As shown in Table F3, the long–short portfolios maintain near-zero market beta while achieving substantially higher Sharpe ratios than Momentum in both markets. Factor-adjusted alpha is positive for every RICE-Alpha portfolio, further supporting benefits beyond lower market exposure. Alpha estimates remain statistically imprecise, with individual significance at the 5% level only for the US DeepSeek long–short portfolio.

Table F3. Long-only and long–short portfolio performance with FF5 + momentum factor regressions. Daily factors are obtained from Kenneth French's data library (Asia Pacific ex Japan for HK). Beta denotes realized market beta relative to the index. Alpha is annualized (%) with HAC t-statistics in parentheses and two-sided p-values. Short positions incur annual borrow fees of 3% (US) and 4% (HK). DeepSeek long-only results are from Table 1, with beta from Table 12.

MarketBookLLMARR (%)SharpeMDD (%)BetaAlpha (t)p
USRICE-Alpha, long-only top 10%DeepSeek30.861.65612.090.5314.20 (1.81)0.0705
USRICE-Alpha, long-only top 10%Llama25.361.40214.370.559.60 (1.22)0.2243
USRICE-Alpha, long–short 10/10DeepSeek19.621.7808.940.0515.40 (2.33)0.0200
USRICE-Alpha, long–short 10/10Llama15.841.4929.870.0612.10 (1.81)0.0709
USMomentum (12–1), long–short 10/10none3.860.28426.41−0.11−1.62 (−0.25)0.8026
HKRICE-Alpha, long-only top 10%DeepSeek36.641.72511.830.4616.90 (1.74)0.0818
HKRICE-Alpha, long-only top 10%Llama30.121.48812.960.4711.80 (1.21)0.2262
HKRICE-Alpha, long–short 10/10DeepSeek12.611.22610.37−0.0310.70 (1.54)0.1237
HKRICE-Alpha, long–short 10/10Llama10.121.00611.24−0.028.40 (1.20)0.2301
HKMomentum (12–1), long–short 10/10none2.120.20622.07−0.07−0.88 (−0.13)0.8987

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

Related papers

Discussion (0)

Sign in to comment.