{"id":"99e17d49-573c-4525-b225-e221f504af71","arxiv_id":"2506.22055","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"An LSTM+XGBoost hybrid is reported to beat standalone models on crypto price forecasts, but the paper gives no reproducible protocol or data.","lead":"This paper applies an LSTM plus XGBoost two-stage model to predict prices of Bitcoin, Ethereum, Litecoin, and Dogecoin, reporting that the hybrid beats individual models on MAPE and MinMax RMSE. The report is a mostly descriptive application with no code, data, or experimental protocol, so the headline comparison cannot be independently checked.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The outperformance claim rests entirely on Table II, but the paper provides no dataset, split, or run-to-run variability; the 0.0045 MAPE edge over XGBoost cannot be distinguished from noise.","rationale":"The reader identified the missing experimental protocol as the load-bearing weakness, and I agree. I sharpen it: the reported margins are so small that the missing protocol is not a formality but the difference between an established result and an unsupported one. The hybrid's MAPE advantage over XGBoost is only 0.0045, and the MinMax RMSE advantage is 0.0046; without error bars or a described split, these values could easily arise from run-to-run variance or a single favorable split. The paper also suffers from internal inconsistencies—the abstract claims sentiment and macroeconomic features that the limitations section says are not used, and section numbers are duplicated—but the decisive issue is that Table II cannot be reproduced or interpreted. There is no code, no data, no hyperparameter list, and no seed or fold information. Given the absence of any independent support (no machine-checked proof, no released artifacts, no parameter-free derivation that would make the result analytic), the central claim fails as an empirical statement. The verdict of REJECT is therefore appropriate, and my read does not change it.","tokens_in":11260,"tokens_out":4633,"duration_ms":48067,"concrete_test":"Obtain the authors' code and raw data, or reimplement from a specified protocol. Set a single time-ordered split (e.g., first 80% train, last 20% test) across all six models, with no shuffling and identical features. Run each model with 10 random seeds and record MAPE and MinMax RMSE per seed. Compute the paired difference in MAPE between Hybrid and XGBoost across the 10 runs and a 95% confidence interval. If the interval includes zero, the Table II margin is not statistically significant. Also verify that every model is evaluated on exactly the same test period and that no validation data was used for early stopping. This single check would establish whether the claimed superiority is reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Hybrid LSTM+XGBoost outperforms standalone models on MAPE and MinMax RMSE (Table II). The only evidence is a single row of 'Test' numbers: Hybrid MAPE 0.0488 and MinMax RMSE 0.0659 versus the next-best XGBoost 0.0533 and 0.0705. These margins are 0.0045 and 0.0046, respectively. The manuscript never states the dataset source, date range, train/validation/test split, number of runs, or any variance estimate. Without this protocol, the small margins could be explained by random initialization, seed choice, or a favorable (or leaky) split. Time-series leakage is a particular risk: if a random split was used, future information can enter the training window. Section IV-J defines the metrics (Eqs. 13-14) but says nothing about the data or split. Section III specifies the architecture (Eqs. 1-12) but never fixes d, n_steps_in, n_steps_out, or the exact feature set. The abstract additionally promises sentiment and macroeconomic inputs, yet Section V-A.3 (the limitations section, confusingly numbered V like the conclusion) states that sentiment and geopolitical events 'are not accounted for in the current model.' This is an internal inconsistency about what was actually evaluated, but the more basic problem is that the reported comparison is not reconstructable. Therefore the central empirical claim is unsupported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage hybrid model in which an LSTM extracts temporal features from cryptocurrency price series and XGBoost performs regression on the LSTM's final hidden state. It gives the standard LSTM and XGBoost equations, reports exploratory data analysis for Bitcoin, Ethereum, Dogecoin, and Litecoin, and presents Table II comparing six models on MAPE and MinMax RMSE. The conclusion states that the LSTM+XGBoost hybrid outperforms standalone models across the four cryptocurrencies and claims the model integrates global and localized exchange data with sentiment and macroeconomic auxiliary features.","tokens_in":11528,"tokens_out":6241,"duration_ms":67468,"significance":"If the empirical claim were properly supported, the contribution would be modest and incremental: the LSTM+XGBoost two-stage architecture is already described in prior cited work, and the reported performance edge over the next-best model is small. The paper's positive aspects are its explicit metric definitions and the internally consistent ordering in Table II. However, the absence of any reproducible experimental protocol and the direct contradiction between the claimed auxiliary features and the stated limitations mean the manuscript does not currently establish its central claim or its claimed novelty.","major_comments":[{"comment":"The manuscript never states the data source, date range, sampling frequency, train/validation/test split, or preprocessing for the four cryptocurrencies, nor does it specify hyperparameters (beyond k=64), n_steps_in, n_steps_out, XGBoost M/γ/λ, number of runs, or random seeds. Table II reports single-point test metrics; the hybrid's edge over XGBoost is 0.0045 in MAPE and 0.0046 in MinMax RMSE, which is well within plausible run-to-run variability for LSTM training. Consequently the central claim that the hybrid 'outperforms standalone models' is not supported as written.","section":"§IV-A, §IV-J, Table II"},{"comment":"The abstract and Section I state that XGBoost models 'auxiliary features such as sentiment scores and macroeconomic indicators' and that the model is evaluated on 'global and localized exchange data,' but Section V-A.3 explicitly says sentiment and geopolitical events 'are not accounted for in the current model,' and Section IV-A lists only OHLCV and marketcap features. The model actually evaluated is not the model described, and the claimed sentiment, macroeconomic, and localized-exchange contributions are absent from the experiments.","section":"§I, §V-A.3"},{"comment":"The conclusion claims the model 'consistently outperforms' across Bitcoin, Ethereum, Litecoin, and Dogecoin, yet Table II reports only aggregate test metrics with no per-currency breakdown, confidence intervals, or statistical significance tests. The table is therefore insufficient to support a per-currency or 'consistent' outperformance claim.","section":"Table II, §V"},{"comment":"The MinMax RMSE definition uses max(A)-min(A) over the actual values, but the paper does not state whether this range is computed on the training portion only or on the full dataset, nor does it describe the temporal split. A random split of financial time series can leak future information into training, and a full-data range mixes test information into the evaluation normalization; both choices need to be specified and justified.","section":"§IV-J, Eq. (14)"}],"minor_comments":[{"comment":"The quantities n_steps_in, n_steps_out, d, and the XGBoost regularization parameters in Eqs. (9)-(11) are introduced but never given values; please provide a full hyperparameter table.","section":"§III, Eqs. (1)-(12)"},{"comment":"The manuscript contains two sections numbered V ('Conclusion and Novelty' and 'Limitations and Future Research Work'); the second should be renumbered, for example as Section VI.","section":"§V"},{"comment":"References [23]-[28] duplicate [16]-[21], and the in-text citation [22] in Section III does not match the cited survey; the bibliography should be cleaned up and deduplicated.","section":"References"},{"comment":"The template placeholders 'Index Terms—IEEE, template, style, LaTeX' and 'Identify applicable funding agency here. If none, delete this.' remain in the text and should be removed.","section":"§I"},{"comment":"The 'Enhanced Interpretability' bullet lists a 'Potential integration' as a key contribution; either implement and report it or move it to future work.","section":"§V, Key Novel Contributions"},{"comment":"Figure captions are present for Figures 1-12 but the figures themselves are not visible in the provided text; please confirm that the final PDF includes all figures and that each is legible.","section":"Figures"}],"recommendation":"major_revision","confidential_remarks":"The core architecture closely follows Wang and Zhang [21] and Shi et al. [51], which the authors cite; the claimed novelty rests on the unsupported sentiment, macroeconomic, and localized-exchange components. If the authors cannot supply a complete experimental protocol, per-currency results, and variance estimates, or if the experiments were in fact run on a random split, the paper should not be accepted in any form. Given the current state of the manuscript, a quick reject would also be defensible; I recommend major revision only on the assumption that the missing materials can actually be provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a clean write-up of a standard two-stage LSTM+XGBoost architecture for crypto price prediction, but the central claim — that the hybrid beats standalone models — rests entirely on Table II, and the paper provides no experimental protocol anywhere. No dataset source, no date range, no train/test split, no hyperparameters, no number of runs, no error bars. The reported margins (MAPE 0.0488 vs 0.0533 for XGBoost) are small enough that they could easily be noise or a favorable split. So despite the clear math, the paper is not a credible empirical contribution.\n\nWhat the paper does well: the equations for the LSTM and XGBoost components are standard and correctly written; the metrics (MAPE, MinMax RMSE) are appropriate for scale-free comparison; and the limitations section is honest that sentiment, news, and geopolitical events are not included. That last point directly contradicts the abstract, which promises sentiment and macroeconomic inputs, but at least the author flags the gap.\n\nThe deeper problem is novelty. The paper's own references [7], [21], [22], [51], [52] describe LSTM+XGBoost hybrids for crypto and stock prediction, including two-stage designs. The 'key novel contributions' listed in Section V — two-stage architecture, adaptability across multiple cryptos, interpretability potential — are not novel on the paper's own evidence. Also, the manuscript has two sections numbered V (Conclusion and Novelty, then Limitations), and several references are duplicated (e.g., [16]/[23], [17]/[24]). That is not fatal to the science but signals an unpolished submission.\n\nThis looks like a course project or a technical note, not a research paper. The EDA in Section IV is descriptive and not tied to the model. For a reader who wants a primer on how one might combine LSTM and XGBoost, the math section is readable. But for the literature, there is no new result, no reproducible evidence, and no insight that would change how someone approaches crypto forecasting. I would not send this to peer review. Desk reject.","headline":"A standard LSTM+XGBoost hybrid whose only experimental evidence is a single table with no protocol; the claimed gains are small, the architecture is prior art, and the paper should be desk-rejected.","tokens_in":12111,"tokens_out":3062,"would_cite":false,"duration_ms":30373,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims a two-stage LSTM+XGBoost model beats five standalone baselines on Bitcoin, Ethereum, Litecoin, and Dogecoin price forecasting.","keywords":["cryptocurrency price prediction","LSTM","XGBoost","hybrid model","time series forecasting","MAPE","MinMax RMSE"],"falsifier":"Run all six models on an identical, fixed train/validation/test split of the same price histories, with the same hyperparameters and multiple random seeds; if the hybrid's MAPE of 0.0488 does not beat XGBoost alone's 0.0533 by more than run-to-run variation, the reported advantage does not hold.","tokens_in":10987,"feed_emoji":"📈","tokens_out":9992,"duration_ms":88170,"temperature":0.7,"pith_summary":"This paper tries to establish that cryptocurrency price forecasting improves when a Long Short-Term Memory network and an XGBoost model are chained in two stages rather than used alone. The LSTM is asked to learn the temporal dependencies in a rolling window of price data, and XGBoost is then asked to turn the LSTM's final hidden state into a price prediction. On Bitcoin, Ethereum, Litecoin, and Dogecoin, the author reports that the hybrid achieves the lowest test MAPE (0.0488) and lowest Min-Max normalized RMSE (0.0659) among six models, ahead of LSTM, CNN, Transformer, ARIMA, and XGBoost. The practical interest is that a modest architectural pairing could give traders and analysts a more accurate tool in a notoriously volatile market.","feed_headline":"Hybrid LSTM+XGBoost model beats five rivals on crypto price forecast","feed_subtitle":"On Bitcoin, Ethereum, Litecoin and Dogecoin, the hybrid posts the lowest MAPE and MinMax RMSE in a six-model comparison.","key_machinery":"The carrying mechanism is a two-stage pipeline in which an LSTM, a recurrent neural network with gated memory, first compresses a rolling input window $X=[x_{(t-n+1)}, \\ldots, x_{(t)}]\\in\\mathbb{R}^{n\\times d}$ into a single learned vector. The cell's forget, input, output, and candidate gates update a memory state $C_t$ and hidden state $h_t$, and only the final hidden state $z=h_n\\in\\mathbb{R}^k$ (with $k=64$) is passed onward. Stage two is an XGBoost regressor that learns a mapping $\\hat{y}=f(z)=\\sum_{m=1}^M f_m(z)$ by minimizing the regularized objective $\\sum_i \\ell(\\hat{y}_i,y_i)+\\sum_m \\Omega(f_m)$, with $\\Omega(f)=\\gamma T + \\frac{1}{2}\\lambda\\sum_j w_j^2$ penalizing tree leaves and leaf weights. The point of the chain is that XGBoost never sees raw prices; it sees a learned temporal summary, so the two models are doing complementary work.","core_discovery":"The paper's central claim is that the hybrid LSTM+XGBoost architecture outperforms every standalone model it is compared with on the task of predicting next-period cryptocurrency prices. In the reported comparison, the hybrid's test MAPE of 0.0488 and MinMax RMSE of 0.0659 are the lowest of the six models, beating the closest competitor, XGBoost alone, which posts 0.0533 and 0.0705 respectively. The author presents this as evidence that the two-stage design combines LSTM's strength at modeling sequential structure with XGBoost's strength at handling nonlinear feature interactions. The claim is posed as generalizable across the four major cryptocurrencies tested and across global and localized exchange data, although the paper's own limitation section notes that overfitting, limited explainability, and dependence on historical price features remain open concerns.","pith_inferences":["An extension the paper leaves implicit: because the reported margins are small (0.0045 MAPE over XGBoost alone), a reader should treat 'outperforms' as a single-run point estimate unless repeated trials show the gap is stable.","The abstract promises sentiment and macroeconomic features, but the dataset description lists only price, volume, and market cap; feeding exogenous variables through the same two-stage pipeline is a natural test the paper does not run.","The normalized metrics make the same pipeline applicable to other high-volatility assets; testing the model on altcoins outside the top four or on commodities would show whether the claimed adaptability is a property of the architecture or of these four coins.","The paper notes potential overfitting in its limitation section; a concrete hedge would be to evaluate on a walk-forward split that re-trains the model each month, which would also clarify whether the hybrid advantage survives regime changes."],"forward_implications":["If the reported comparison is taken at face value, anyone building a crypto price forecaster from price history alone would do better to chain LSTM with XGBoost than to tune either model alone: the hybrid's MAPE is 0.0488 versus 0.0533 for the best standalone model.","The ordering of baselines (Transformer at 0.0594 MAPE ahead of CNN at 0.0612 but behind XGBoost at 0.0533) suggests that on this dataset, tree-based nonlinear regression on tabular features is a stronger baseline than deep sequence models.","Because the metrics are scale-independent (MAPE and MinMax RMSE), the claimed improvement is intended to transfer across the four coins' very different absolute price levels.","The two-stage design keeps the output interpretable at the XGBoost stage, so feature-importance tools can be applied to the latent representation produced by the LSTM.","The paper's own forward-looking section expects the same hybrid pattern to work with Transformer encoders and with external sentiment features, which is a direct extension of the claimed architecture."],"supporting_citations":[{"why":"Supplies the LSTM baseline and the evidence that LSTM captures long-term dependencies in financial sequences.","marker":"[3]"},{"why":"Defines XGBoost and the regularized tree-boosting objective the second stage uses.","marker":"[5]"},{"why":"Earlier hybrid LSTM+XGBoost result on Bitcoin and Ethereum that this paper extends to four coins.","marker":"[7]"},{"why":"Reported the earlier hybrid LSTM+XGBoost RMSE improvement on Bitcoin and Ethereum that the paper presents as motivation.","marker":"[21]"},{"why":"Provides ARIMA, the classical statistical baseline in the six-model comparison.","marker":"[11]"},{"why":"Provides the Transformer/Temporal Fusion Transformer baseline in the comparison.","marker":"[15]"},{"why":"Supplies the CNN model used as a comparison baseline.","marker":"[4]"}],"fun_headline_variants":["LSTM+XGBoost hybrid tops six-model crypto forecast test","Hybrid LSTM+XGBoost wins crypto price prediction race","Best crypto forecast: LSTM+XGBoost beats standalone models","LSTM+XGBoost hybrid posts lowest error on Bitcoin, Ethereum, more","Hybrid AI model beats five rivals at crypto price prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison stands on the assumption that all six models were evaluated on the same unseen test data with no leakage from the test period into training or model selection, a protocol the paper does not actually specify.","fun_headline_variants_meta":{"raw":{"variants":["LSTM+XGBoost hybrid tops six-model crypto forecast test","Hybrid LSTM+XGBoost wins crypto price prediction race","Best crypto forecast: LSTM+XGBoost beats standalone models","LSTM+XGBoost hybrid posts lowest error on Bitcoin, Ethereum, more","Hybrid AI model beats five rivals at crypto price prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000457,"raw_usage":{"total_tokens":2262,"prompt_tokens":884,"completion_tokens":1378,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1296}},"tokens_in":500,"tokens_out":1378,"duration_ms":10018,"temperature":1.0,"reasoning_tokens":1296,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:11:57.665854+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run all six models on an identical, fixed train/validation/test split of the same price histories, with the same hyperparameters and multiple random seeds; if the hybrid's MAPE of 0.0488 does not beat XGBoost alone's 0.0533 by more than run-to-run variation, the reported advantage does not hold.","supporting_citations":[{"cited_title":"Deep learning with long short-term memory networks for financial market predictions,","cited_arxiv_id":null,"evidence_quote":"Supplies the LSTM baseline and the evidence that LSTM captures long-term dependencies in financial sequences."},{"cited_title":"XGBoost: A scalable tree boosting system,","cited_arxiv_id":null,"evidence_quote":"Defines XGBoost and the regularized tree-boosting objective the second stage uses."},{"cited_title":"Hybrid deep learning and machine learning model for cryptocurrency prediction,","cited_arxiv_id":null,"evidence_quote":"Earlier hybrid LSTM+XGBoost result on Bitcoin and Ethereum that this paper extends to four coins."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides ARIMA, the classical statistical baseline in the six-model comparison."},{"cited_title":"Temporal Fusion Transformers for interpretable multi-horizon time series forecasting,","cited_arxiv_id":null,"evidence_quote":"Provides the Transformer/Temporal Fusion Transformer baseline in the comparison."},{"cited_title":"A CNN–LSTM model for gold price time-series forecasting,","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN model used as a comparison baseline."}],"review_version":1}