acceptodds
Under review as a conference paper at ICLR 2027

R3-TTA: Test-Time Adaption From Matured Outcomes for Post-event Market Prediction

Abstract

Standard Test-Time Adaptation assumes that an unlabelled query carries a useful learning signal. We show this assumption fails in market response prediction, where outcomes have not occurred at prediction time and posts arriving within the same time bar share a single future return, causing standard gradient adapters to severely degrade performance relative to the frozen baseline. We propose Regime-Routed Retrieval Test-Time Adaption (R3TTA), which grounds adaptation in historical events with matured outcomes. For each query bar, R3TTA retrieves past matured bars routed by arrival timing and lagged market state, fits a prior-shrunk Bernoulli response model with a closed-form minimizer, and fuses this local estimate with the frozen prediction in logit space. By design, zero parameters are updated, no signals are extracted from unlabelled test batches, and shared outcomes are counted strictly once. Fusing the local nonparametric estimate with the frozen prediction drives the primary performance gains, far outperforming naive bias-offset adjustments. On post-event market prediction, R3TTA consistently improves performance across diverse architectures and walk-forward evaluations achieving substantial gains on tabular models where all gradient adapters fail. Without retuning, R3TTA seamlessly transfers to broad market assets. We release the aligned post–market panel with a reproducible content tagging layer. validated adaptation for on.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Test-Time-Adaption, Quant Investment

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 3 reviews, above 27% of the others.
Reviewer_por9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on test-time adaptation under delayed label setting. At prediction time, the ground truth has not been realized yet, and multiple posts falling into the same time bar share one single future market target. The authors find that standard gradient-based TTA may perform worse than a frozen baseline here, because online test samples cannot supply reliable self-supervision signals. They propose R3-TTA, a retrieval-based and gradient-free TTA method. It restricts the memory bank to only mature historical samples with known ground truth, uses Bernoulli estimator to get local probability and fuses it with logits from the frozen base model. The evaluation is built on a self-collected social-media financial dataset, covering S&P 500 and crypto volatility prediction. Multiple backbones are tested, including tree models, MLP and LLMs. However, the empirical evidence is not sufficient in current submission, missing some key ablation, reproducibility materials, and there should be more TTA baselines to be compared with.

Critical strengths

The research gap detected by the paper author is critical and worth fixing, i.e., outcomes have not occurred at prediction time and posts arriving within the same time bar share a single future return in market response prediction. This finding shows a good scientific taste.

Critical weaknesses

The experiments do not show enough generalization and effectiveness. Both cases utilize the same social media source as input and the financial mechanism show high similarity. We cannot confirm whether the observed phenomenon and R3-TTA’s advantage can transfer to other more independent tasks. Baseline comparison is also not fair enough. The paper mainly compares with gradient TTA methods, while retrieval-style TTA baselines are omitted.

Questions for authors
  1. Add ablation to quantify how much performance drop comes from memory contamination?
  2. The current submission has no separate supplementary files, including both dataset and code to show reproducibility evidence.
  3. Argue why the core conclusion still holds on other delayed-label prediction tasks with different data generating mechanisms?
  4. Add more retrieval-style TTA baselines for comparison.
3 comments
AuthorsOpenReview ↗

Response to Reviewer por9

Thank you for the review.

W1: generalisation beyond one source. Beyond the 2026 equity test panel:

  • Time. On a forward window after the test panel (758 bars, 2026-04-16 to 2026-08-25), read by no script before the protocol hash and scored once after it, the gain is +0.0112 [+0.0043, +0.0201], Holm p = 0.003.
  • Assets. On 40 crypto perpetuals, zero-shot, the gain is positive on 40/40, significant on 16/40 and significantly negative on 0/40.
    • BTC: 0.6135 → 0.6386 (+0.0251 [+0.0146, +0.0376]).
    • ETH: 0.5309 → 0.5710 (+0.0401 [+0.0232, +0.0536]).
    • The series are correlated (about 6 effective series) and BTC/ETH had been scored for earlier arms before the hash, so crypto is supporting evidence.
  • Architectures. Gains appear on seven frozen backbones, including four language models that read the post text.

We agree that all events come from one account, and we will state this as a limitation. The tabular backbones and the routing spaces use timing and lagged market features, not text (l.205–206). The operator does not use the account's wording, but the event times come from one account; other event streams are future work.

W2 / Q4: retrieval-style baselines. Table 2 already contains eight retrieval comparators on the same matured history (l.341; App. Table 6): T3A (NeurIPS 2021), nearest centroid, kNN vote, AdaNPC vote and AdaNPC fused (ICML 2023), Nadaraya–Watson, and kNN-LM in a selected and a fixed-grid version. While checking them we found incorrect rows in Table 2; the corrected values (GBT / MLP / FT) change the comparison:

MethodTest AUCc07 − method [95% CI], GBT / MLP / FT
R3-TTA (c07)0.7848 / 0.7779 / 0.7479—
AdaNPC fused0.7855 / 0.7802 / 0.7708−0.0007 / −0.0022 / −0.0230, all intervals include 0
kNN-LM-72 (fixed grid)0.7880 / 0.7826 / 0.7554−0.0032 / −0.0047 / −0.0076 [−0.0125, −0.0029]
T3A0.7500 / 0.7401 / 0.7255above on all three
Frozen0.7679 / 0.7514 / 0.7226—
  • AdaNPC is within sampling noise of c07, but it needs 4,536 validation configurations; c07 needs none.
  • Fixed-grid kNN-LM-72 is also tuning-free and scores higher than c07 on all three tabular backbones. The difference is significant on FT-Transformer.
  • We therefore present R3-TTA and kNN-LM-72 as two instances of tuning-free nonparametric memory adaptation and withdraw "top AUC across all backbones".
  • TDA (CVPR 2024) is discussed in Table 1 but was not run, because it caches vision–language zero-shot logits. We ran no further retrieval baseline for this response.

Q1: memory contamination (new experiment). We replace a random fraction ρ of the memory labels with random 0/1 labels (GBT, 5 seeds, re-run environment where frozen = 0.7652):

MemoryGain over frozen
true labels+0.0199 [+0.0080, +0.0294]
ρ = 0.1 / 0.25 / 0.5+0.0151 to +0.0173 / +0.0101 to +0.0186 / +0.0051 to +0.0109
ρ = 1−0.0060 to −0.0011 (all intervals include 0)
labels → frozen out-of-fold predictions (soft / hard)+0.0186 / +0.0279
newest 75% of memory removed+0.0195
  • The gain falls as contamination rises, and it disappears at ρ = 1 without a significant drop below frozen. The operator therefore degrades to the frozen model rather than below it.
  • Replacing labels with frozen predictions, or keeping only old memory, keeps the gain. Label maturity is the condition that prevents look-ahead at deployment, not the source of the gain, and we revise the framing accordingly.

Q2: supplementary material. The code and data files are at anonymous.4open.science/r/R3TTA (l.506–514). This includes the operator, the baselines, the hashed protocol, the evaluation scripts, and the post and bar files used. Before the end of the discussion period we will add the language-model scoring scripts with frozen outputs, the crypto pipeline, the rebuttal experiment scripts and the 16-candidate selection table.

Q3: other delayed-label tasks. We claim one specific result: when predictions are committed before horizon-delayed labels mature, tuning-free local smoothing over admissible memory improves discrimination, and unsupervised gradient TTA does not. The same constraint appears in other settings:

  • delayed conversion attribution (Chapelle, KDD 2014; Ktena et al., RecSys 2019);
  • stream classification with verification latency (Grzenda et al., DMKD 2019);
  • delayed-feedback online learning (Joulani et al., ICML 2013);
  • online continual learning with label delay (Csaba et al., NeurIPS 2024). There, rehearsing similar labelled memory outperforms unsupervised adaptation, which matches our finding.

The admissibility rule, the bar-level memory and the hashed protocol carry over to these settings without change. §5 will discuss them as future work, without claiming results outside our benchmark.

We ask you to reconsider the score in light of these additions.

AuthorsOpenReview ↗

summary of rebuttal revision for por9

CommentChange in the revised PDFWhere
W1: generalisation over timeApp. E.3: held-out forward window, +0.0112 [+0.0043, +0.0201], Holm p = 0.003.App. E.3
W1: generalisation over assetsThe abstract reads "transfers to crypto perpetuals". The intro scope reads "SPY, with crypto perpetuals as a transfer panel". App. E.3: 16/40 significant; BTC 0.6135 → 0.6386, ETH 0.5309 → 0.5710.Abstract; §1; App. E.3
W1: generalisation over architecturesApp. E.1 Table 9: 7 backbones, all positive with c07.App. E.1
W1: one accountLimitation (a): events come from one account; other event streams are future work.App. C
W2/Q4: retrieval baselinesApp. E.1 Table 8 gives the recomputed AdaNPC fused and kNN-LM-72 rows. App. E.3 gives the paired contrasts with a search-count column, and the forward-window contrasts.App. E.1, E.3
Q1: memory contaminationApp. E.7, at ρ = 0.1 / 0.25 / 0.5 / 1: the operator falls back to the frozen model, not below it. Also the out-of-fold label and newest-memory results.App. E.7
Q2: supplementary materialThe Reproducibility statement lists what the repository holds and what is added: LM scripts, the crypto pipeline, the rebuttal scripts and the selection table.Reproducibility
Q3: other delayed-label tasksApp. C adds a future-work sentence. It cites Chapelle 2014, Ktena et al. 2019, Grzenda et al. 2020, Joulani et al. 2013 and Csaba et al. 2024, and makes no claims outside this benchmark.App. C

Hope the revision and clarify in rebuttal can solve your concern and consider increase the rating of the paper.

AuthorsOpenReview ↗

Additional results abut how much does unreliable memory cost?

These are the means across five random-label draws at each replacement rate, using the same 525 SPY test bars and frozen GBT as in Appendix E.7.

Memory labels randomizedAUC gain over frozenAUC lost versus clean memory
0%+0.0199—
10%+0.01620.0037
25%+0.01550.0044
50%+0.00900.0108
100%−0.00310.0229

Half-random memory loses about half the gain. At full randomization all five estimates fell below frozen, although each 95% interval included zero. Our earlier reply overstated the evidence for a fallback to frozen.

Transfer. Excluding BTC and ETH, which earlier variants had already been tested on, leaves 38 of 38 crypto series with positive gains (median +0.0108). The series share one account’s event times and are correlated. Transfer to another event source remains open.

Baselines and code. Fixed-grid kNN-LM also improves on frozen and exceeds R3-TTA on FT-Transformer by 0.0076 [0.0029, 0.0125] (E.1, E.3). A gradient-only comparison would miss this strong baseline. Tabular evaluation code is public. The promised language-model outputs, crypto pipeline and rebuttal scripts remain outstanding. Anonymous repository

Do these results address your concerns about contamination and retrieval baselines? we would be grateful for any further comments, and we would ask you to consider whether the revised evidence solved your concerns also increase the overall score.

Best

The authors

Reviewer_YLS5OpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a simple way to adjust a frozen market predictor using historical events whose outcomes are already known. For each query, it retrieves similar historical bars, estimates a local event probability with shrinkage toward a prior, and combines this with the frozen prediction. That said, I am not sure if I learned much from this paper.

Critical strengths

The method is simple enough to understand and inspect, and the reported improvements over the frozen predictors are encouraging.

Critical weaknesses

I also could not reconcile the backbone used in Table 4 with the preceding description. This matters because the ablations are used to justify the individual design choices.

Questions for authors

No

3 comments
AuthorsOpenReview ↗

Response to Reviewer YLS5

Thank you for the review. Your concern about Table 4 is correct, and the error was ours. This comment answers that weakness first, then the ablations you refer to and the main results they affect.

Table 4. The submitted Table 4 mixed GBT ablation rows with a Seed-OSS reference row (0.8023), so its Δ column is invalid. The corrected table uses GBT and the final operator c07 in every row, with every Δ taken against GBT c07 (0.7848 test AUC). Intervals are paired 5-day block-bootstrap 95% CIs.

ConfigurationWalk-forwardTestΔ vs c07 [95% CI]
Frozen0.69700.7679−0.0169 [−0.0254, −0.0070]
+ gate ρ_b (Eq. 10)0.70120.7749−0.0100 [−0.0166, −0.0025]
market space only0.69400.7750−0.0099 [−0.0247, +0.0052]
timing space only0.71160.7824−0.0025 [−0.0106, +0.0082]
joint space only0.72010.7831−0.0018 [−0.0070, +0.0024]
uniform spaces0.71310.7852+0.0004 [−0.0017, +0.0025]
raw vote0.71770.7881+0.0033 [−0.0041, +0.0090]
probability mixture0.71960.7847−0.0001 [−0.0032, +0.0018]
c07 (final)0.71540.7848—

What the ablations support:

  • Adding the reliability gate lowers AUC significantly.
  • The market space alone is the weakest single space; its interval includes 0.
  • The remaining design choices are not individually significant.
  • Single grid members score 0.7837–0.7862. The 12-member grid exists to remove the hyperparameter search over T, λ and k, not to add accuracy.

We withdraw the claim that every ablation degrades performance (l.424–431), and §4.2 will state the narrower reading above.

Main results after correcting Tables 2 and 3 to the final operator:

  • Gains over frozen. +0.0169 (GBT), +0.0265 (MLP), +0.0252 (FT-Transformer), and +0.0138 to +0.0215 on the four language models. Six of seven intervals exclude 0; the MoE interval is [−0.0003, +0.0260].
  • Forward window. 758 bars after the test panel, read by no script before the protocol hash and scored once after it: +0.0112 [+0.0043, +0.0201], Holm p = 0.003.
  • Crypto. Zero-shot on 40 crypto perpetuals: positive on 40/40, 16/40 significant, 0/40 significantly negative. The series are correlated (about 6 effective series), so the count is descriptive.

What there is to learn. In a delayed-label stream, predictions are committed before outcomes exist. In this setting:

  • no gradient TTA method exceeds the frozen model by more than 0.003 on any backbone;
  • a fixed nonparametric operator over admissible memory adds +0.017 to +0.027, with no per-backbone tuning and no backward pass;
  • a fixed-grid kNN-LM over the same memory gives similar gains and scores slightly higher than R3-TTA (significant on FT-Transformer); we now present it as a second instance of the same method class.

New controls added for the rebuttal, including the negative results:

  • Placebo times. Gains at bars without posts are positive but smaller (mean +0.0076, against +0.0199 at post times; all 10 placebo gains are below it).
  • Trading hours. Within regular trading hours alone, the gain is not significant (+0.0026).
  • Memory labels. Replacing them with the frozen model's out-of-fold predictions keeps the gain (+0.0186 soft, +0.0279 hard). Label maturity is therefore the condition that prevents look-ahead, not the source of the gain, and we revise the "matured outcomes" framing accordingly.
  • Other anchors. Gains persist on bagged and shallower GBT anchors (+0.0146 to +0.0212).

We ask you to reconsider the rating in light of the corrected table and these controls.

AuthorsOpenReview ↗

summary of rebuttal revision for YLS5

CommentChange in the revised PDFWhere
Table 4 Δ invalidApp. E.1 Table 10: GBT throughout, Δ against R3-TTA (0.7848). The last row reads 0.7154 walk-forward / 0.7848 test.App. E.1
Ablation readingApp. E.1: the gate is the only significant drop; the market space alone is weakest; the other variants and single members (0.7837–0.7862) are within noise; the 12-member grid removes the search over T, λ and k.App. E.1
Main results with the final operatorApp. E.1 Table 9: +0.0169 / +0.0265 / +0.0252 on the tabular backbones and +0.0138 to +0.0215 on the language models; 6 of 7 intervals exclude 0.App. E.1
Forward windowApp. E.3: +0.0112 [+0.0043, +0.0201], Holm p = 0.003. Both tuning-free memory operators outperform every gradient adapter.App. E.3
CryptoApp. E.3: 40/40 positive, 16/40 significant, 0 significantly negative; about 6 effective series; BTC and ETH numbers.App. E.3
kNN-LM-72App. E.1 and E.3: slightly above R3-TTA; paired contrasts and search counts (0 vs 4,536).App. E.1, E.3
New controlsPlacebo times in App. E.6; trading hours in App. E.5 and Limitation (b); memory labels and other anchors in App. E.7.App. E.5–E.7; App. C

Hope the revision and clarify in rebuttal can solve your concern and consider increase the rating of the paper.

AuthorsOpenReview ↗

Additional results about what the corrected ablation shows

You were right about Table 4. It used a Seed-OSS-36B reference for gradient-boosted tree (GBT) ablations. Appendix E.1, Table 10 gives the comparison on one backbone and the same 525 SPY test bars.

GBT configurationAUCGain over preceding row (paired 95% CI)
Frozen0.7679—
Add retrieval with the neighbor-agreement gate0.7749+0.0070 [+0.0039, +0.0096]
Keep retrieval and remove the gate (final R3-TTA)0.7848+0.0100 [+0.0025, +0.0166]

Retrieval helps even with the gate. Removing it helps further. The remaining component comparisons have intervals spanning zero, so we withdraw the claim that every design choice improves AUC.

There is also a gain in probability accuracy. The Brier score falls from 0.1857 to 0.1796 (paired reduction 0.0061 [0.0008, 0.0105]). Counting each completed market bar once gives the frozen predictor a useful update while the new posts still await their shared outcome.

Does the corrected comparison resolve your concern about the ablation? we would be grateful for any further comments, and we would ask you to consider whether the revised evidence solved your concerns also increase the overall score.

Best

The authors

Reviewer_1gqyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies test-time adaptation (TTA) for predicting market responses to social media posts, they use Trump's Truth Social posts as the case study. The authors argue that standard TTA is ill-posed since the outcome has not happened at prediction time, and all posts in the same 15-minute bar share one label. Therefore they propose R3-TTA, which retrieves historical bars with matured labels in three routing spaces and computes a prior-shrunk Bernoulli estimate in closed form, and then fuses it with the frozen logit at a fixed weight β=1/2\beta = 1/2, eventually, the fused predictions is averaged over a fixed 12-member grid. During the process, no parameter is updated and nothing is estimated from the test batch. Overall, the method improves bar-level AUC over the frozen model on seven backbones over a 2026 test panel. The gains also extend to a walk-forward split, a forward window and a transfer to 40 crypto perpetuals, in contrast, the gradient TTA baselines mostly hurt or stay flat under this setup.

Critical strengths
  • Overall the evaluation protocol is careful. The bar-level AUC would count each shared outcome once, meanwhile, the designs such as label maturity audit, day-clustered block bootstrap, and the extra panels make the gain over the frozen model credible.
  • The paper gives a clear diagnosis of why query-driven TTA breaks in this setting, in particular, no forward-looking signal, correlated copies within a bar. The proposed fix is simple, deterministic and cheap, as reported in the paper, it takes about 0.107 ms per bar, and its order invariance in proposition 2 makes predictions easy to audit. I consider this mainly as a practical strength for deployed event-driven forecasters.
Critical weaknesses
  • The technical novelty seems limited and the TTA framing is a bit of a stretch. In my opinion, this method is very close to kNN-LM style interpolation or Nadaraya-Watson voting, with Beta prior smoothing and logit fusion. Prop 1 is the standard Beta-Bernoulli posterior mean, meanwhile, Prop 2 follows directly from the construction. Since the adaptation signals comes from labeled history instead of from the input, this makes it more like retrieval-augmented prediction or stacking.
  • It is not fully clear that the task tests post-event prediction. The label is magnitude-only with a median threshold, the clean model uses only five timing and four lagged market features (no text), meanwhile, the label window can cross overnight or weekend gaps. So a large part of the AUC may come from predicting volatility from time of day and recent market state, which might be similar for any timestamp that has post or not post at all. Also, it is not explained why local label averages over the same pre-deployment data add information beyond a GBT trained on the same nine features.
  • There are several consistency and clarity issues. Table 4 is described as the evaluation on Seed-OSS-36B, but it seems that the frozen row (0.7679) matches the GBT frozen model in Table 3 (also 0.7679), not Seed-OSS (0.7909). Also, it seems that "gate restored" in Fig. 4b is not defined. The authors introduce 39-instrument universe and asset-level targets, but I could not find asset-level results.
  • The Trump dataset is listed as the first contribution, but its construction is described only briefly in Appendix D. I could not find the source of the posts or the essential details, such as how they were collected, the market data vendor, or how the 8847 records were filtered to 8678. Based on the code provided in the reproducibility statement, the "content-tagging layer" seems to consist of eight keyword-based regex lists (common_protocol_clean.py), several of them lack word boundaries, therefore it may produce false positives, for example, "xi" may match words like "Mexico" or "maximum", and "import" may match "important". After checking the data file data/posts.csv, I found out about 1/4 of the posts are empty text, and close to half have fewer than five words among the posts in the study window. Another issue is that the overnight posts seem to be aligned to the next morning's bar. If the released data is the actual one used and reported in the paper, I would suggest performing more sophisticated preprocessing/filtering.
  • Based on the docstrings in the released scripts exp_candidates_0923.py and exp_declare_0923.py, the model-selection process seems to differ from the paper's description. In particular, the docstrings in the 2 files appear to indicate a few things, first, the final candidate screen was started after a previously declared variant had already been scored on the 2026 test panel; second, the selection rule was revised after the walk-forward and validation results of earlier rounds had been read; third. the original rule selects a different method (to be specific, c10_evid_prob_joint2 in the code), which is designated as a secondary arm, this is not reported in the paper; finally, the revised rule ranks candidates first by the number of walk-forward folds below frozen, so in this case, the statement "no fold falls below the frozen model" is more of the selection criterion itself rather than independent evidence to me. And I could not find this disclosed in the paper. It might because these docstrings are outdated or do not reflect the final pipeline, if so, it would be helpful to clarify them. Otherwise, I would suggest reporting the full candidate set, the selection rules, as well as the results of the originally selected method, since this affects the paper's claims that the protocol was declared before the test read.
Questions for authors
  • Could you clarify which backbone is used in each row of table 4?
  • Could you run the same pipeline on bars that contain no post, or with shuffled post times? If the AUC and the gain stay similar, the results would be more convincing about volatility timing than about posts to me.
  • How do the main results change if you exclude bars whose label window spans an overnight or weekend gap, or restrict to regular trading hours?
  • The abstract and section 3.3 say R3-TTA far outperforms “bias-offset” adjustments, but I could not find this defined or reported in table 2, table 4 or any part of the appendix. This is central to the variance reduction explanation, so could define and report it? Also, could you please report the results for the 39-instrument targets in the paper?
  • Could you please provide more details regarding how the dataset is constructed? It also would be helpful to clarify if redistribution of the posts and intraday market data is allowed.
  • Could you clarify the inconsistency of the code/data and the paper mentioned in the weakness?
  • The released repository does not seem to include the scripts or data for the LLM backbones and the crypto-perpetual panel, are those planned for release?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 1gqy (1/2): W1, W2, W3/Q1

Thank you for carefully auditing our code and data. We checked each point against our run records; all are correct. This comment follows the order of your review through W3/Q1; the second answers W4/Q5, W5/Q6 and Q2–Q7. l. = line in the submitted PDF; AUC = test AUC on the 2026 panel (525 bars); intervals = paired 5-day block-bootstrap 95% CIs; c07 = final operator; respec = c07 with uniform space weights (printed in Table 2); kNN-LM-72 = fixed-grid kNN-LM, no validation search. E1–E6 are new experiments in our re-run environment (frozen 0.7652, c07 0.7851; paper environment 0.7679 / 0.7848).

W1. Novelty: "close to kNN-LM / Nadaraya–Watson with Beta smoothing"; Prop. 1 standard, Prop. 2 immediate. We agree the estimator combines classical parts and claim no new estimator; Prop. 1 will be cited as known, Prop. 2 becomes a remark. What we claim:

  1. A delayed-label setting where gradient TTA gives no gain (§3.1): labels mature after a fixed horizon, posts in one bar share one label, predictions are committed before any label of the bar exists. No gradient adapter beats frozen by more than 0.003; nonparametric memory operators add +0.017 to +0.027. This matches Csaba et al. (NeurIPS 2024): under label delay, unsupervised adaptation does not close the gap and rehearsing similar labelled memory does.
  2. No per-backbone search: c07 uses zero validation configurations (AdaNPC 4,536; tuned kNN-LM 3,402) and no backward pass, and is within noise of AdaNPC. kNN-LM-72 is also tuning-free and scores higher than c07 on 7/7 backbones (significant on FT, and on GBT in our re-run environment); we present it as a second instance of the class, not a beaten baseline.
  3. An auditable protocol: hashed protocol, full selection log (in our second comment), forward window no script read before the hash.

Not RAG or stacking: no meta-parameters are fit, and β, the spaces and the scales are fixed in advance. Unlike TabR (ICLR 2024) and ModernNCA (ICLR 2025), it needs no joint training and attaches to any frozen model, including 27B–36B language models.

W2. Post-event prediction

(a) Volatility timing — you are right about the frozen tabular model. Same operator and memory, query times moved to bars without posts (±7 d, ±14 d, random no-post bars, clock-matched random dates; 10 settings, E2): frozen GBT scores 0.7269–0.7760, against 0.7652 at post times. We will state that the frozen tabular AUC mainly reflects volatility timing. The c07 gain at post times (+0.0199) exceeds all 10 placebo gains (+0.0002 to +0.0121, mean +0.0076; rank p ≈ 1/11), so about +0.012 is specific to post times and the rest is local smoothing present at any time.

(b) Text. Of the four frozen language models that read the post, only Seed-OSS-36B is significantly above frozen GBT: +0.0256 [+0.0034, +0.0471]; the others give +0.0014 to +0.0166. None differs significantly from GBT + R3-TTA (0.7851). The tabular backbones and routing spaces use timing and lagged market features, not text (l.205–206).

(c) Why local averages add. Not through newer labels (E3): removing the newest 75% of memory keeps the gain (+0.0195), and so does replacing every memory label with the GBT's out-of-fold predictions (+0.0186 [+0.0112, +0.0261] soft, +0.0279 [+0.0146, +0.0413] hard). The gain is averaging the GBT's scores over routing-space neighbourhoods; bagging and shallower trees do not remove it (E4: +0.0146 to +0.0212 on five anchors, all intervals exclude 0). We will revise "matured outcomes" (l.069–075, §4.3): maturity is the admissibility condition that prevents look-ahead, not the source of the gain.

W3 / Q1. Table 4 backbone, "gate restored", 39-instrument universe. Table 4 mixed GBT ablation rows with a Seed-OSS reference row (0.8023), so its Δ column is invalid. Table R1 replaces it: GBT throughout, Δ against c07 (0.7848); walk-forward = 9 folds, 1,594 bars.

ConfigurationWalk-fwdTestΔ vs c07 [95% CI]
Frozen0.69700.7679−0.0169 [−0.0254, −0.0070]
+ gate ρ_b (Eq. 10)0.70120.7749−0.0100 [−0.0166, −0.0025]
market space only0.69400.7750−0.0099 [−0.0247, +0.0052]
timing space only0.71160.7824−0.0025 [−0.0106, +0.0082]
joint space only0.72010.7831−0.0018 [−0.0070, +0.0024]
uniform spaces (respec)0.71310.7852+0.0004 [−0.0017, +0.0025]
raw vote0.71770.7881+0.0033 [−0.0041, +0.0090]
probability mixture0.71960.7847−0.0001 [−0.0032, +0.0018]
c07 (final)0.71540.7848—

We withdraw "every ablation strictly degrading" (l.424–431). Single members score 0.7837–0.7862 (all intervals include 0); the 12-member grid removes the search over T, λ, k rather than adding accuracy. "Gate restored" (Fig. 4b) is shipped72, the earlier 72-member operator with half its members gated by ρ_b; c07 is above it on 9/9 walk-forward folds; the caption will define it.

AuthorsOpenReview ↗

Response to Reviewer 1gqy (2/2)

W4 / Q5. Dataset construction, filtering, redistribution.

  • Sources. Posts: CNN Truth Social archive (ix.cnn.io/data/truth-social/truth_archive.json), HTML stripped. Bars: unadjusted 15-minute OHLCV, 04:00–20:00 New York time.
  • Filtering (App. D.1). 32,436 records → 8,847 in 2024-11-01 to 2026-04-10 → 8,691 from 2024-11-06 → 8,678 aligned to a bar; each step and the vendor will appear in D.1.
  • Overnight posts. As you note, a post between 20:00 and 04:00 or on a weekend maps to the next 04:00 bar, the first at which it can be traded (App. D.2); Q3 quantifies it.
  • Text length. Your estimate is correct: 24.1% of posts are empty after HTML stripping and 15.5% more contain only a link (39.5% have no words besides a URL); 46.8% have fewer than 5 words.
  • Filtering test (E6, new). Scoring only test posts with words (70% of rows): gain +0.0206 [+0.0082, +0.0312]; ≥5 words: +0.0195 [+0.0076, +0.0294]. Re-running after removing those posts from training, memory and test: +0.0114 [+0.0020, +0.0201] and +0.0099 [−0.0011, +0.0203]. Removing the same number of random posts per split keeps +0.0146 to +0.0262 (5 seeds, all CIs exclude 0), so the drop comes from which posts are removed: link-only posts still mark event times and add memory rows. We keep all posts; E6 goes in App. D.
  • Regex. \b was missing; with it matches drop (import 254 → 6, tax 602 → 92). Tags are not model inputs; no AUC changes.
  • Redistribution (Q5, Q7). The anonymous repository holds data/posts.csv and data/market_15min.csv, the files used.

W5 / Q6. Selection history; code–paper inconsistency. The docstrings are accurate, the paper did not disclose this:

  1. respec was scored on test on 2026-09-21, two days before the candidate screen.
  2. Sixteen candidates (four rounds) were scored only on walk-forward and 2025-Q4 validation (407 bars); neither overlaps test or the forward window.
  3. Rule v1, fixed before any candidate was scored (best pooled walk-forward AUC; P(Δ ≤ 0) < 0.05 and validation Δ ≥ −0.002 against respec), selects c10_evid_prob_joint2.
  4. Rule v2, written after reading rounds 1–3 and before round 4 (walk-forward ≥ respec and ≥ kNN-LM-72, validation Δ vs frozen ≥ 0; ranked by fewest folds below frozen, then worst fold), selects c07_joint2. The protocol was hashed under v2 (f4063cad…8002, 2026-09-23 11:30:40 CST).
  5. "No fold falls below the frozen model" (l.084–089) is therefore the v2 ranking criterion, not evidence; it is removed from the findings.
Test AUCc10 (v1)c07 (v2)
GBT / MLP / FT0.7849 / 0.7786 / 0.75470.7848 / 0.7779 / 0.7479
crypto significant14/4016/40

c10 − kNN-LM-72: GBT −0.0032 [−0.0080, −0.0003], MLP −0.0040 [−0.0090, +0.0022], FT −0.0007 [−0.0058, +0.0033]. Under v1 the deficit to kNN-LM-72 is significant on GBT instead of FT; nothing else changes. The hashed forward-window script scored only c07. The appendix will list all 16 candidates, both rules, timestamps and c10 results. The forward window does not depend on the selection; crypto is supporting evidence only, since BTC and ETH had been scored for earlier arms.

Q2. Bars with no post, or shuffled post times. E2 in W2: at 10 placebo settings without posts the frozen GBT still reaches 0.7269–0.7760, and c07 gains +0.0002 to +0.0121 there against +0.0199 at post times. We report both: about +0.012 post-specific, the rest generic local smoothing.

Q3. Gaps & trading hours (E1).

Test subsetBarsGain [95% CI]
all525+0.0199 [+0.0080, +0.0294]
label window contiguous437+0.0247 [+0.0090, +0.0394]
window crosses a gap88+0.0193 [−0.010, +0.0529]
regular trading hours only178+0.0026 [−0.016, +0.0198]

Gap-crossing windows do not drive the gain. Within regular trading hours alone it is not significant; we state this as a limitation.

Q4. Bias-offset (E5). δ_b = logit(q_b) − logit(q̄_b), the same kernel and shrinkage applied to frozen memory probabilities; p = σ(ℓ + β·δ_b), β = 1/2, no gate. Test: +0.0021 [−0.0022, +0.0040] over frozen, −0.0153 [−0.0240, −0.0079] against respec; walk-forward +0.0051 / −0.0109. These replace "far outperforming". The 39-instrument target is withdrawn (W3).

Q7. Language-model and crypto materials. Both enter the release before the discussion ends: scoring scripts with frozen outputs, and the crypto pipeline with per-asset results.

Corrections (confirmed against run records).

ItemSubmittedCorrected (GBT / MLP / FT)
Tab. 2 R3-TTArespec 0.7852 / 0.7788 / 0.7483c07 0.7848 / 0.7779 / 0.7479
Tab. 2 AdaNPC fused0.7655 / 0.7802 / 0.74080.7855 / 0.7802 / 0.7708
Tab. 2 kNN-LM-720.7808 / 0.7726 / 0.75540.7880 / 0.7826 / 0.7554
Tab. 3 LM rowsgated shipped72c07, +0.0138 to +0.0215
Fig. 4c15/40 significant16/40
AuthorsOpenReview ↗

Summary of rebuttal revision for 1gqy

CommentChange in the revised PDFWhere
W3/Q1: Table 4 mixed backbonesApp. E.1 Table 10 recomputes Table 4 with GBT in every row and Δ against R3-TTA (0.7848). Only adding the gate ρ_b lowers AUC significantly (−0.0100 [−0.0166, −0.0025]). Single members score 0.7837–0.7862. The body's Table 4 is unchanged.App. E.1
W3/Q1: "gate restored" undefinedDefined in App. E.3 as shipped72, the earlier 72-member operator with half its members gated by ρ_b. R3-TTA is above it on 9/9 walk-forward folds (0.7154 vs 0.7073; frozen 0.6970).App. E.3
Q1/Q4: 39-instrument universe"|A| = 39" is replaced by "the evaluated equity instrument is SPY".§3 Problem Definition
Corrections: Tables 2 and 3, Fig. 4cRecomputed with c07 in App. E.1: Table 8 (Table 2 rows: c07, AdaNPC fused, kNN-LM-72) and Table 9 (all 7 backbones; 6 of 7 intervals exclude 0). E.1 and E.3 state 16/40 significant on crypto.App. E.1, E.3
"Top AUC" and "every ablation" readingsApp. E.1 states that under c07 kNN-LM-72 scores slightly above R3-TTA on the tabular backbones, and that only the gate ablation is significant. App. E.3 gives the paired contrasts and the search counts (0 vs 4,536).App. E.1, E.3
Gradient adapters on FTApp. E.3: ETA 0.7250, EATA 0.7240 and TAFAS 0.7227 stay within 0.003 of frozen (0.7226).App. E.3
W5/Q6: selection historyApp. E.2: respec scored on 2026-09-21; 16 candidates scored on walk-forward and validation only; rule v1 → c10, rule v2 → c07; hash and timestamp; the c10 vs c07 table; the c10 − kNN-LM-72 intervals. It states that "no fold below frozen" is a rule-v2 ranking criterion. App. B.1 has a pointer.App. E.2; App. B.1
W1: novelty, Prop. 1/2Prop. 1's title adds "standard Beta–Bernoulli conjugacy". Prop. 2 becomes Remark 1. App. E.3 names the classical components and contrasts the method with TabR, ModernNCA and stacking: no fitted parameters, no joint training.§3; App. A.6; App. E.3
W1: delayed-label settingCsaba et al. (NeurIPS 2024) is cited in Related Work. App. E.3 states the agreement: rehearsing labelled memory beats unsupervised adaptation.§2; App. E.3
Q4: bias-offsetApp. E.8 gives the definition and the numbers: test +0.0021 over frozen, −0.0153 against respec; walk-forward +0.0051 / −0.0109; validation 0.7403. The abstract sentence is as submitted.App. E.8
W2(a)/Q2: volatility timing, placeboApp. E.6: the post-time gain (+0.0199) exceeds all 10 placebo gains (rank p ≈ 1/11); about +0.012 is post-specific. Frozen GBT scores 0.7269–0.7760 at placebo times. Limitation (c) is added.App. E.6; App. C
W2(b): textApp. E.6: only Seed-OSS-36B is significantly above frozen GBT (+0.0256).App. E.6
W2(c): why local averages addApp. E.7: the gain survives two changes: memory labels replaced by out-of-fold predictions (+0.0186 soft / +0.0279 hard), and the newest 75% of memory removed (+0.0195). Label maturity is stated as the admissibility condition. Five regularised anchors gain +0.0146 to +0.0212; the section notes they were chosen after the test set was seen.App. E.7
Q3: overnight gaps, trading hours§3 points to App. E.5, whose gap/session table shows that gap-crossing windows do not drive the gain. Limitation (b) covers regular trading hours (+0.0026, not significant). App. D.2 adds the 04:00–20:00 bars and the next-04:00 mapping.§3; App. E.5; App. D.2; App. C
W4/Q5: data constructionThe l.336 sentence is fixed. App. D.1 adds: the source URL and mirror; HTML stripping; the filtering chain 32,436 → 8,847 → 8,691 → 8,678; Polygon as the bar vendor. App. E.9 adds the text statistics, the E6 filtering test and the regex \b correction.§4; App. D.1; App. E.9
Q5/Q7: redistribution, LM/crypto materialsThe abstract says "We release alignment code and labels for the post–market panel". The Reproducibility statement lists the additions and states that Polygon bars and archive text are not redistributed.Abstract; Reproducibility
Re-run environmentApp. E.4: frozen 0.7652 → c07 0.7851; Brier −0.0076; +0.0058 over shipped72.App. E.4

Hope the revision and clarify in rebuttal can solve your concern and consider increase the rating of the paper.

AuthorsOpenReview ↗

Additional results about does the gain survive market closures?

You asked whether overnight gaps explain the SPY result. Appendix E.5 restricted the bars we scored. We also restricted which historical bars R3-TTA could retrieve.

Windows allowed in both memory and evaluationTest barsAUC gain over frozen GBT (paired 95% CI)
All windows525+0.0199 [+0.0080, +0.0294]
Contiguous windows only437+0.0211 [+0.0086, +0.0314]
Regular trading hours only178+0.0084 [−0.0157, +0.0276]

Retrieval still helps when overnight and weekend gaps are excluded. The regular-hours estimate remains uncertain. As the placebo results in E.6 show, timing already predicts volatility at bars without posts. Our forecast ranks post-time bars by the probability that the absolute SPY move will exceed 17.15 basis points.

One clarification on selection (E.2–E.3). The original rule selected c10 and the revised rule selected c07. Their SPY test AUCs were 0.7849 and 0.7848. Avoiding losses across folds helped select c07. The September 23 protocol fixed its later 758-bar comparison with frozen GBT (+0.0112 AUC) before scoring that comparison. Other analyses had already used those dates.

Does this address your concerns about market timing and selection? we would be grateful for any further comments, and we would ask you to consider whether the revised evidence solved your concerns also increase the overall score.

Best

The authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Friendly Reminder to Reviewers

Dear Reviewers,

Thanks for your attention to this email. During the rebuttal phase, we extensively answered the questions raised by the reviewers and believe we have addressed the reviewers' concerns also make this interesting work better. This is a friendly reminder that the author-reviewer discussion period is coming to an end. We understand you are busy, but we would appreciate any additional feedback or suggestions you might have on our work. Your insights are valuable to us and R3TTA.

Sincerely,

The Authors

AuthorsOpenReview ↗

Friendly Reminder to Reviewers

Dear Reviewers,

Thanks for your attention to this email also your time and effort to review task. During the rebuttal phase, we extensively answered the questions raised by the reviewers and believe we have addressed the reviewers' concerns also make this interesting work better. This is a friendly reminder that the author-reviewer discussion period is coming to an end. We understand you are busy, we would be grateful for any further comments, and we would ask you to consider whether the revised evidence solved your concerns. Your insights are valuable to us and R3TTA in domain adaption task.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.