{"id":"c3af29f8-be7e-4252-a6fa-8c63ec98751d","arxiv_id":"2504.17313","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A patched channel-mixing encoder reportedly beats four baselines on multi-step stock price forecast and prediction, with small margins and no released code.","lead":"This paper introduces PCIE, a transformer-style model that patches stock price and price-change channels, mixes them across channels, and directly forecasts several future prices in one step. The authors report lower forecast error than four established baselines on two U.S. stock datasets, but the supporting evidence lacks error bars and reproducible artifacts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is supported by Table 2, but Table 4 reports different MSE/MAE for the same PCIE configuration on US_14L prediction (L=20,40,60), so the evidence base is internally inconsistent.","rationale":"The reader's weakest assumption was the absence of an explicit chronological train/validation/test split. I agree that this is a serious reporting gap, but the more directly load-bearing problem is the internal inconsistency between Table 2 and Table 4 for the identical PCIE configuration on US_14L prediction. This is not a hypothetical risk; it is a concrete contradiction in the reported numbers that the paper itself does not explain. The central claim depends on Table 2, and Table 4 is used to support the tokenization contribution, so the two tables cannot both be right as printed. Resolving this inconsistency is a prerequisite for trusting either the SOTA claim or the ablation conclusion. The chronological-split issue is also important, and if the split is random the performance numbers would be inflated, but the table contradiction is stronger evidence that the empirical reporting is unreliable as it stands. The model design is coherent, with a plausible patching and channel-mixing architecture, and the paper does include an ablation study, so this is not a rejection of the approach; it is a request for verifiable, internally consistent results. The reader's conditional verdict already requires resolving this inconsistency, so my stress-test does not change the recommended outcome.","tokens_in":10503,"tokens_out":4925,"duration_ms":51165,"concrete_test":"Obtain the authors' code and data (or ask them to run the exact Section 4.3 configuration) and reproduce PCIE on US_14L prediction for L=20, 40, and 60 under a stated chronological 7:1:2 split. Compare the reproduced MSE/MAE to Table 2 and Table 4; the table that fails to reproduce identifies the erroneous numbers and determines whether the SOTA claim or the ablation claim stands. If code is unavailable, request the raw per-horizon result logs for these rows.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PCIE outperforms state-of-the-art baselines rests on Table 2. Table 4, the ablation study, reports the identical model (PCIE) on the same dataset (US_14L) and same prediction horizons, yet gives different numbers for L=20/40/60: MSE 1.5356/1.5129/1.4801 in Table 4 versus 1.5181/1.4746/1.4611 in Table 2, with corresponding MAE differences. The L=10 row matches, so this is not a different task definition. One of the two tables must be wrong, or the runs differ in an undisclosed way (e.g., seed, hyperparameters, data split). Since Table 2 is the evidence for the headline SOTA result and Table 4 is the evidence for the tokenization contribution, the paper currently gives two mutually incompatible quantitative accounts of the same model. Until this is resolved, the reported superiority cannot be treated as reliable. The related missing detail of whether the 7:1:2 split is chronological compounds the problem: with stride S=1 patching and percentage-change inputs, a random split would let test windows overlap training windows in time and inflate all results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PCIE, a Patched Channel Integration Encoder, for multi-step stock price forecasting (predicting future close prices) and prediction (predicting future price changes). The model applies univariate patching, an adaptive temporal learning block, channel mixing, and a channel-mixing self-attention encoder with direct multi-step output. The authors evaluate PCIE on two US stock datasets, US_71 and US_14L, across horizons of 10, 20, 40, and 60 days, comparing against PatchTST, D-Va, Autoformer, and Informer. Table 2 reports lower MSE and MAE for PCIE than all baselines, and the authors also claim a data-preprocessing method that mixes price and percentage-change channels improves performance across all models. An ablation study in Table 4 is presented as evidence that the tokenization process is responsible for the improvement.","tokens_in":10763,"tokens_out":6140,"duration_ms":44778,"significance":"If the quantitative claims are correct, the paper offers a practical improvement for multi-step financial forecasting, with a tokenization design that captures cross-channel correlations and a preprocessing augmentation that appears to generalize across baselines. The paper is entirely empirical, provides no code release or machine-checked derivations, and its main support is the comparative table. The claimed SOTA result and the tokenization contribution are precisely what the internal inconsistencies in Tables 2 and 4 put at risk, so the significance is conditional on the reported numbers being reproducible and internally consistent.","major_comments":[{"comment":"The same PCIE configuration is reported with different values in Table 2 and Table 4. For US_14L prediction, Table 2 gives MSE/MAE for L=20, 40, and 60 as 1.5181/0.8601, 1.4746/0.8520, and 1.4611/0.8502, while Table 4 gives 1.5356/0.8683, 1.5129/0.8664, and 1.4801/0.8577 for the same PCIE model. The L=10 row matches in both tables, so the discrepancy is not a change in task definition. One of the two tables must be incorrect, or the runs differ in an undisclosed way such as seed, hyperparameters, or data split. Since Table 2 is the evidence for the headline SOTA claim and Table 4 is the evidence for the tokenization contribution, the two tables cannot both be right; please reconcile them and report which numbers are correct.","section":"Comparison of Table 2 and Table 4"},{"comment":"Section 4.1 states only that data is split into training, validation, and testing with a ratio of 7:1:2, but it does not state that the split is chronological. This matters because the tokenization uses overlapping patches with stride S=1 (Section 3.3) and the input features include percentage changes computed from adjacent days (Section 3.1), so a random split would allow test windows to overlap in time with training windows. Such leakage would inflate the reported performance for PCIE and for the baselines. The paper must state explicitly whether the split respects time order; if it does not, the experiments must be rerun with a chronological split and the corrected results reported.","section":"Sections 4.1, 3.3, and 4.3"},{"comment":"All results are single-run point estimates with no error bars, confidence intervals, or significance tests. Several margins in Table 2 are small; for example, US_71 prediction at L=60 reports PCIE MSE 0.9983 versus D-Va 1.0054, and US_14L forecast at L=40 reports PCIE 0.5570 versus PatchTST 0.5705. Without repeated runs from different seeds or a paired significance test, the claim that PCIE outperforms the baselines is not statistically supported. Please report results over multiple seeds (at least 3 to 5) as mean plus/minus standard deviation, and where appropriate apply a paired test such as the Diebold-Mariano test for forecast comparisons.","section":"Section 4.4 and Table 2"},{"comment":"Table 3 reports a single percentage improvement for each dataset and model under 'mixing data', but the paper never defines how 'overall performance' is computed. It is not stated whether the percentage is averaged over all horizons, over MSE and MAE, or over forecasting and prediction tasks, or some weighted combination. Without this definition, the contribution of the proposed preprocessing method cannot be evaluated quantitatively. Please specify the aggregation formula and provide a per-task, per-horizon breakdown of the improvement.","section":"Section 4.4 and Table 3"}],"minor_comments":[{"comment":"The statement that 'all the other parameters are tuned according to the characteristics of the dataset' is too vague; please list the final hyperparameters, including d_model, d_patch, number of attention heads, and number of encoder layers, for each dataset.","section":"Section 4.3"},{"comment":"Instance normalization is mentioned in one sentence but its exact placement in the pipeline is not described; please clarify for which components it is applied and whether it is used in both forecasting and prediction.","section":"Section 3.7"},{"comment":"The distributional claims that forecasting targets follow a log-normal distribution and prediction targets follow a normal distribution are stated without empirical support or citations; please add evidence or references.","section":"Abstract and Section 1"},{"comment":"The 'No Tokenization' condition is not described; please explain what replaces the tokenization process in the ablation, otherwise the comparison is difficult to interpret.","section":"Section 4.5 and Table 4"},{"comment":"Reference [31] appears incomplete; it lacks publication venue, year, and page numbers. Please complete the bibliographic details.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency between Tables 2 and 4 is severe enough that, if the authors cannot reconcile the numbers, the paper would not be suitable for publication in its current form. I also share the concern about the unspecified data-split scheme, because the stride-1 patching makes temporal leakage a real risk. Beyond these issues, the idea is reasonable and the experimental design is otherwise standard, so I do not recommend rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is a modest, plausible engineering combination: univariate patching in the style of PatchTST, channel mixing by flattening across series, and an adaptive temporal learning block that picks among linear, independent linear, and MLP mappings. That is a reasonable architecture and it is clearly described. The ablation, leaving the number mismatch aside for a moment, does suggest the tokenization component helps, and the preprocessing trick of feeding both price level and percentage change as separate channels is clean and easy to adopt. The paper is readable and the experimental setup is standard: two datasets, four baselines, four horizons.\n\nThe stress-test note holds up: Table 2 and Table 4 report different MSE and MAE for the identical PCIE configuration on US_14L prediction at horizons 20, 40, and 60. That is not a rounding difference, and it matters because Table 2 is the evidence for the headline SOTA claim. One of those tables is wrong, or the runs differ in an undisclosed way. On top of that, there are no error bars or significance tests, and the absolute gains over the best baseline are small. The 7:1:2 split is never stated to be chronological; with stride-1 patching and percentage-change features computed from adjacent days, a random split would let test windows overlap training windows in time and inflate every number. No code or data is released, so none of this can be independently checked.\n\nI don't see fraud here, and I don't see a load-bearing theoretical flaw. I see an empirical paper that is under-reported. The authors need to reconcile the two tables, state the split explicitly, provide multiple-seed results or confidence intervals, and release code and data. If they do that, the result would be a solid, if modest, contribution to a crowded field. As it stands, I would not cite it, and I would not trust the SOTA claim.\n\nWho is this for? Researchers in financial machine learning who want a baseline for multi-step stock price forecasting and prediction. It deserves a serious referee: the idea is coherent and the experiments are easily replicable in principle, so a good referee can quickly tell whether the numbers survive cleanup. My recommendation is to send it to peer review with a major-revision request rather than desk-reject, but the next version needs to be an honest accounting of uncertainty, not just another table of wins.","headline":"A plausible channel-mixing patch model for stock forecasting, but the evidence base has an internal inconsistency that must be fixed before the SOTA claim is credible.","tokens_in":11300,"tokens_out":2087,"would_cite":false,"duration_ms":21098,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PCIE, a patched-channel transformer encoder, claims better multi-step stock price forecasts and predictions than four state-of-the-art baselines on two U.S. datasets, at horizons 10, 20, 40, and 60 trading days.","keywords":["stock price forecasting","stock price prediction","multi-step forecasting","tokenization","patching","channel mixing","transformer encoder","time series"],"falsifier":"Run the identical experiment with a strictly chronological split (first 70% of trading days for training, next 10% for validation, last 20% for testing) and recompute the table; if PCIE's margins over PatchTST shrink or reverse, the SOTA claim rested on leakage from overlapping patches across the split boundary.","tokens_in":10291,"feed_emoji":"📈","tokens_out":5367,"duration_ms":45342,"temperature":0.7,"pith_summary":"This paper tries to establish that a single encoder model can handle both tasks that stock-price models usually split apart: forecasting future close prices and predicting future percentage changes, over horizons of 10, 20, 40, and 60 trading days. The model, PCIE, tokenizes each input channel into overlapping patches, embeds each patch with an adaptive temporal block, and mixes all channels before self-attention; the authors report the lowest MSE and MAE among the compared models on both datasets at every horizon. They also claim that adding percentage-change channels alongside raw prices improves every model, including the baselines. A sympathetic reader takes away that tokenization plus channel mixing is a workable route to longer-horizon financial forecasts.","feed_headline":"Price tokens beat leading stock forecast models on every horizon","feed_subtitle":"PCIE mixes price and percentage-change channels to cut multi-step forecast error from 10 to 60 days.","key_machinery":"The load-bearing object is the tokenization pipeline: univariate patching divides each input channel into patches of length P=4 with stride S=1; the adaptive temporal learning block chooses between a shared linear map, per-series linear maps, or an MLP to embed each patch; channel mixing flattens the embedded patches from all channels and adds learnable position encoding, so each self-attention token carries cross-channel information; and a final flatten-plus-linear layer emits all future steps in one pass. The direct multi-step output is what avoids iterative error accumulation, and instance normalization plus the price/percentage-change input pair handles distribution shift.","core_discovery":"On the paper's own terms, the discovery is that price series can be represented as cross-channel tokens, and that this representation is what makes multi-step stock forecasting and prediction work. PCIE reports the lowest MSE and MAE among the five compared models on US_71 and US_14L, for all forecast and prediction lengths in {10, 20, 40, 60}. The ablation study shows that removing the tokenization process increases error in nearly every setting, which the authors take as evidence that patching, adaptive temporal learning, and channel mixing together capture correlations between channels that univariate models miss.","pith_inferences":["A testable extension would ablate channel mixing while keeping patching fixed: the paper's ablation removes the whole tokenization process, so it does not isolate which component drives the gain.","The price-plus-percentage-change input is effectively a differencing scheme; the same preprocessing could be applied to volatility, volume, or other financial series, and may benefit models beyond these baselines.","Because the paper reports point estimates without variance across random seeds, a natural next check is whether the margins over PatchTST persist across repeated runs and across different market regimes in the test period."],"forward_implications":["Multi-step forecasts of 10 to 60 trading days can be produced in a single forward pass, so early errors cannot compound through iterative decoding.","Adding percentage-change channels alongside raw prices improved every model tested, including the baselines, which suggests the preprocessing trick transfers beyond PCIE.","The ablation ties the performance gain to tokenization itself, so future models can adopt patching and channel mixing even if they keep a different backbone.","The reported table gives a concrete target: on US_71 and US_14L, any new method should beat PCIE's MSE and MAE at horizons 10, 20, 40, and 60 before claiming an improvement."],"supporting_citations":[{"why":"Baseline PatchTST; supplies the patching concept and the univariate transformer that PCIE extends with channel mixing.","marker":"[16]"},{"why":"D-Va baseline and the multi-step regression stock prediction task setting that PCIE must beat.","marker":"[11]"},{"why":"Autoformer baseline, representing decomposition/auto-correlation transformers for long-series forecasting.","marker":"[25]"},{"why":"Informer baseline, the efficient transformer whose distilling convolution causes information loss in comparison.","marker":"[32]"},{"why":"Source for the claim that direct multi-step forecasting avoids the cumulative error of iterative decoding.","marker":"[28]"},{"why":"Reversible instance normalization, used to handle distribution shift in price data.","marker":"[10]"}],"fun_headline_variants":["Tokenized prices enhance multi-step stock forecasting and prediction","PCIE model beats SOTA on stock forecast and prediction horizons","Price tokens cut multi-step stock forecast error across horizons","Cross-channel tokenization improves stock prediction up to 60 days","New token method tops stock forecast benchmarks on all horizons"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the unstated premise that the 7:1:2 train/validation/test split is chronological and leakage-free, since the paper does not say the split respects time order and overlapping patches could let training and test samples share days.","fun_headline_variants_meta":{"raw":{"variants":["Tokenized prices enhance multi-step stock forecasting and prediction","PCIE model beats SOTA on stock forecast and prediction horizons","Price tokens cut multi-step stock forecast error across horizons","Cross-channel tokenization improves stock prediction up to 60 days","New token method tops stock forecast benchmarks on all horizons"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000405,"raw_usage":{"total_tokens":2102,"prompt_tokens":933,"completion_tokens":1169,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":1099}},"tokens_in":549,"tokens_out":1169,"duration_ms":8844,"temperature":1.0,"reasoning_tokens":1099,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:42:36.454876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical experiment with a strictly chronological split (first 70% of trading days for training, next 10% for validation, last 20% for testing) and recompute the table; if PCIE's margins over PatchTST shrink or reverse, the SOTA claim rested on leakage from overlapping patches across the split boundary.","supporting_citations":[{"cited_title":"In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management","cited_arxiv_id":null,"evidence_quote":"D-Va baseline and the multi-step regression stock prediction task setting that PCIE must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline PatchTST; supplies the patching concept and the univariate transformer that PCIE extends with channel mixing."},{"cited_title":"Advances in Neural Information Processing Systems34, 22419–22430 (2021)","cited_arxiv_id":null,"evidence_quote":"Autoformer baseline, representing decomposition/auto-correlation transformers for long-series forecasting."},{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"Informer baseline, the efficient transformer whose distilling convolution causes information loss in comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source for the claim that direct multi-step forecasting avoids the cumulative error of iterative decoding."},{"cited_title":"In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=cGDAkQo1C0p","cited_arxiv_id":null,"evidence_quote":"Reversible instance normalization, used to handle distribution shift in price data."}],"review_version":1}