{"id":"100ef54e-c67e-4919-88a4-a8c1a2dae460","arxiv_id":"2501.12239","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A CNN trained on raw candlestick chart images predicts market strength at roughly 0.7 accuracy, and adding YOLO-detected candlestick patterns does not reliably improve that performance.","lead":"This paper tests whether AI models can predict whether a market trend is strong or weak by looking only at images of candlestick charts, with and without detected candlestick patterns. The authors report that adding pattern detection did not consistently improve accuracy over raw chart images, with best accuracy near 0.7.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The black-box labeling function in Section III.A.2 makes all accuracies uninterpretable, and Table II contradicts the abstract's claim on BTC-USD; a transparent labeling rule is needed to validate the negative result.","rationale":"The reader's weakest assumption correctly identifies the labeling function as the foundation of the entire empirical claim. Every reported accuracy measures agreement with this unstated ground truth; if the labels are arbitrary or leaky, the pattern-versus-non-pattern comparison is meaningless, and the abstract's conclusion cannot be assessed. I agree with this concern and see no reason to downgrade it. The internal contradiction between the abstract and Table II (BTC-USD with-pattern 0.792 vs non-pattern 0.765) is a second, independent strike: even if the labels were perfect, the paper's own data do not support a blanket 'patterns do not improve performance' statement without error bars or significance tests. The concrete test I propose would settle the matter by making the label function explicit and re-running the key comparison; if the pattern effect is robust to a standard labeling rule, the negative result gains credibility, and if not, the paper's claim is shown to be an artifact of the black-box label construction. The verdict should remain REJECT because the manuscript as written does not provide enough information to verify either the labels or the conclusion.","tokens_in":9166,"tokens_out":2802,"duration_ms":29056,"concrete_test":"Obtain or reconstruct the exact label function (threshold, horizon, pattern set, class balance). Then rerun the VGG16 non-pattern vs include-pattern comparison on a fixed train/test split with a plain, fully specified rule, e.g., strong if the next 30 daily returns sum greater than 0, and report per-seed mean and standard deviation across at least 5 seeds. If the include-pattern gap flips sign or overlaps zero on BTC-USD and EUR-USD, the paper's headline conclusion fails; if it remains consistently non-positive, the concern is resolved.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim — that candlestick patterns add no value over raw images — rests entirely on a binary 'strong/weak' ground truth produced by an unspecified 'custom function' (Section III.A.2). The paper never states the return threshold, the prediction horizon, the class balance, or whether the 30-candle history window overlaps the label window. Without this, every 0.7 accuracy and every pattern/no-pattern gap in Tables II–V could be an artifact of a noisy or leaky label. This is not a minor reproducibility footnote: if the label function is changed, the ordering of models can invert. Moreover, the paper's own Tables II and III show 'include-pattern' beats 'non-pattern' by +2.7 accuracy on BTC-USD (0.792 vs 0.765), while underperforming on EUR-USD and AAPL; the abstract says patterns 'do not improve performance.' With no error bars or significance tests, the negative claim is not established even under the authors' own numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether candlestick chart images alone can predict a binary 'strong' versus 'weak' market trend strength label, and whether explicitly detected candlestick patterns (via YOLOv8) improve a pure-CNN image baseline. It compares a simple CNN (VGG16) on raw images with a two-stream CNN that fuses pattern information, and also evaluates a Decomposer/DCP-style architecture. On datasets of BTC-USD, EUR-USD, AAPL, and a few other assets, the authors report accuracies around 0.6-0.79. The central claim, stated in the abstract and repeated in the discussion, is that candlestick patterns do not improve model performance over raw chart images. The paper also reports that Decomposer/DCP performs near chance, and that performance is below more complex time-series models.","tokens_in":9306,"tokens_out":2699,"duration_ms":28932,"significance":"If the negative result were established with transparent methodology and statistical rigor, it would be a useful benchmark for the vision-based technical-analysis literature, which often reports optimistic results without proper baselines. The cross-market scope (crypto, forex, stocks) is also a valuable aspect. However, the paper's current value is limited because the central finding is contradicted by its own reported tables and rests on an unspecified labeling function that makes every accuracy number uninterpretable. The manuscript does include a reasonable architecture comparison (VGG16 across multiple backbones) and honest reporting of the Decomposer's failure, which are credit-worthy, but these strengths do not compensate for the load-bearing reproducibility gaps.","major_comments":[{"comment":"","section":"III.A.2"},{"comment":"","section":"Tables II and III; Abstract"},{"comment":"","section":"IV (Pure CNN, Include-Pattern)"},{"comment":"","section":"IV (General methodology)"}],"minor_comments":[{"comment":"The ticker 'AAPL' is consistently misspelled as 'APPL' in the tables; please correct.","section":"Tables I, II, IV, V"},{"comment":"The IEEE conference template instructions and the 'Authors background' form are still present in the submitted manuscript and should be removed before any final version.","section":"Last pages"},{"comment":"References [5] and [7] are identical (both are the Deep Candlestick Predictor paper), and the in-text citation 'Abdellah EL ZAAR et al [14]' does not match the reference list entry [14], which is a different work. Please reconcile the bibliography.","section":"References"},{"comment":"The heading 'Cross-Market Anaalysis' contains a typo; it should read 'Cross-Market Analysis.'","section":"II.C"},{"comment":"Figures 3 and 4 are referenced only in the caption and are not discussed in the text; please either integrate them into the discussion or remove them.","section":"IV (Decomposer)"},{"comment":"The description of the two-stream CNN does not specify the architecture of the pattern stream (number of layers, filter sizes, whether it is also VGG16), and the fusion mechanism is described only as 'combined in a fully connected layer.' Please provide a precise architectural summary or a reference to code.","section":"III.B.2"}],"recommendation":"reject","confidential_remarks":"The two load-bearing issues—the completely unspecified strength-labeling function and the direct contradiction between the abstract's negative claim and the BTC-USD results in Tables II–III—mean that the central contribution is not currently supported. These are not local presentation problems; they require either a substantial rewrite of the method and results or new experiments. Given the paper's current state, including leftover template text and duplicated references, I cannot recommend major revision with confidence that the authors would address the scientific core rather than the surface."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper asks a sensible question — does explicit YOLO-based candlestick pattern detection add anything to a CNN classifying market strength from chart images? The comparison across stocks, forex, and crypto is new as a combination, and the negative finding would be useful if it held. It doesn't, as written.\n\nWhat's genuinely new: no one else has fused YOLOv8-detected patterns with a VGG16 two-stream architecture for this exact strength-labeling task across three asset classes. The paper is honest that DCP provides limited signal and reports the DCP model near chance. That is a useful caution.\n\nWhat it does well: the choice of VGG16 as backbone is justified with a small ablation in Table I; the authors include three metrics; the writing is straightforward and the limitations section admits data dependence.\n\nSoft spots, in order of severity. First, the strength labels come from an undescribed 'custom function' in Section III.A.2. No threshold, no horizon, no class balance, no mention of whether the 30-candle history overlaps the label window. Without that, every accuracy number is uninterpretable — the model could be learning the label function's quirks rather than market strength. Second, the abstract's central claim is contradicted by the paper's own Table II vs III: on BTC-USD, include-pattern accuracy is 0.792 vs 0.765 non-pattern, while on EUR-USD and AAPL it is worse. With no error bars or repeated runs, the claim 'patterns do not improve performance' is not established even under the authors' own numbers. Third, there is no train/test protocol, no data split details, no code or data release, and the time-series comparison is based on citations, not experiments. Minor issues: leftover IEEE template text at the end and typos like 'Anaalysis' suggest light proofreading, but those are not scientific flaws.\n\nThe paper is probably salvageable: a transparent labeling rule, error bars, and a clearer statement of the pattern effect direction could make this a modest negative result worth having. As is, the load-bearing gaps prevent it from being a reliable data point.\n\nI'd send it out if it came to a journal, because the question is real and reviewers could force the needed details. But I wouldn't cite it yet.","headline":"Asking whether YOLO-detected candlestick patterns help CNN market-strength classifiers is a fair question, but the black-box label function and a table that contradicts the abstract's main claim make the negative result unestablished.","tokens_in":9917,"tokens_out":2069,"would_cite":false,"duration_ms":20243,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims candlestick patterns detected by YOLOv8 add no predictive value to a CNN that classifies market strength from chart images alone.","keywords":["candlestick patterns","market strength prediction","convolutional neural network","YOLOv8","chart image classification","deep learning for finance","technical analysis","time-series vs image"],"falsifier":"Reproduce the comparison with a transparent, fixed label rule such as the sign of the forward 30-day return over a fixed threshold, on a balanced test set, and check whether including YOLOv8-detected patterns changes accuracy. Also verify that YOLOv8's pattern detections actually match TA-Lib ground truth on the same charts; if detection is often wrong, the null result is explained by misdetection rather than by patterns being uninformative.","tokens_in":8933,"feed_emoji":"📈","tokens_out":3358,"duration_ms":31883,"temperature":0.7,"pith_summary":"This paper tries to establish that a convolutional neural network looking only at candlestick chart images can predict whether an upcoming trend is strong or weak only to about 0.7 accuracy, and that explicitly detecting classical candlestick patterns with YOLOv8 and feeding them into the model does not improve that result. The authors compare a plain VGG16 baseline against a two-stream CNN that incorporates detected patterns, across stock, forex, and cryptocurrency datasets, and find the pattern-aware version never consistently beats the raw-image version. A third architecture adapted from the Deep Candlestick Predictor framework stalls at roughly 0.5 accuracy, equivalent to random guessing. If the result is right, it clarifies a practical ceiling for image-only technical analysis and argues for adding other data modalities or returning to time-series representations.","feed_headline":"Candlestick patterns add no edge to CNN market call","feed_subtitle":"Image-only model peaks near 0.7 accuracy; detected patterns fail to beat raw charts","key_machinery":"The central machinery is the comparison between two CNN pipelines: a simple VGG16 that classifies raw chart images, and a two-stream CNN that fuses features from the raw chart with features extracted from candidate candlestick patterns detected by YOLOv8. The labels for 'strong' versus 'weak' market strength come from a custom function applied to TA-Lib pattern detections and the following 30 candles. A third pipeline, the Decomposer, segments charts into sub-charts before classification and is evaluated only on raw images.","core_discovery":"The central claim is that candlestick pattern information, as detected by YOLOv8, does not improve market-strength prediction from chart images. On the AAPL test set, the pattern-inclusive model reaches 0.653 accuracy while the raw-image model reaches 0.701; on EUR-USD the pattern version scores 0.623 versus 0.753 without patterns, and only on BTC-USD does the pattern version slightly edge ahead (0.792 versus 0.765). The authors interpret these results as showing that explicit pattern detection adds little or negative value over raw candlestick visuals. They also report that the Decomposer-based DCP approach, which segments charts into sub-charts, performs near chance, suggesting that sub-chart decomposition does not help either.","pith_inferences":["The 0.7 accuracy ceiling could partly reflect noise in the custom labeling function rather than a fundamental limit of image inputs; cleaner labels might change the pattern-versus-non-pattern comparison.","A testable extension is to feed candlestick patterns as soft attention masks or continuous feature maps instead of binary detected boxes, which could reverse the null result.","The paper compares against 'more complex time-series models' only by reference; a matched experiment with identical labels and splits would directly settle whether the image representation is the bottleneck.","The near-random Decomposer performance suggests the architectural adaptations (RGB autoencoder, halved CNN1D layers) may be faulty, so the negative result should not be generalized to all chart-segmentation approaches."],"forward_implications":["If the central claim is correct, adding object-detected candlestick patterns to a strength-prediction CNN is not worth the extra complexity.","Image-only models appear to plateau at roughly 0.7 accuracy, suggesting chart visuals carry only weak predictive signal for trend strength.","Time-series representations, which preserve temporal order, likely outperform image conversions for this task, as the authors argue.","Sub-chart decomposition as implemented in the DCP adaptation does not help and may hurt, with performance near random.","Future work should shift toward richer inputs such as volume, technical indicators, or multimodal data rather than refining pattern detection alone."],"supporting_citations":[{"why":"Supplies the Deep Candlestick Predictor framework that the Decomposer, CNN-autoencoder, and CNN1D components are adapted from.","marker":"[5]"},{"why":"YOLOv8 is the object detector used to locate candlestick patterns in the Include-Pattern experiments.","marker":"[3]"},{"why":"TA-Lib provides the rule-based pattern detections that feed the custom function generating the strong/weak ground-truth labels.","marker":"[19]"}],"fun_headline_variants":["Image-only CNN beats pattern-enhanced model on market strength","Candlestick patterns fail to boost CNN market prediction","Chart images alone predict market strength at 0.7 accuracy","Pattern detection yields no edge for candlestick CNN","Raw charts outperform YOLO patterns in market strength test"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The strong-versus-weak labels come from an unspecified custom function applied to TA-Lib pattern detections and the next 30 candles; if that function is noisy, threshold-sensitive, or leaks future information, every accuracy number measures that labeling scheme rather than market strength.","fun_headline_variants_meta":{"raw":{"variants":["Image-only CNN beats pattern-enhanced model on market strength","Candlestick patterns fail to boost CNN market prediction","Chart images alone predict market strength at 0.7 accuracy","Pattern detection yields no edge for candlestick CNN","Raw charts outperform YOLO patterns in market strength test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000468,"raw_usage":{"total_tokens":2313,"prompt_tokens":906,"completion_tokens":1407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":1327}},"tokens_in":522,"tokens_out":1407,"duration_ms":8242,"temperature":1.0,"reasoning_tokens":1327,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:21:28.196913+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the comparison with a transparent, fixed label rule such as the sign of the forward 30-day return over a fixed threshold, on a balanced test set, and check whether including YOLOv8-detected patterns changes accuracy. Also verify that YOLOv8's pattern detections actually match TA-Lib ground truth on the same charts; if detection is often wrong, the null result is explained by misdetection rather than by patterns being uninformative.","supporting_citations":[{"cited_title":"Deep Candlestick Predictor,","cited_arxiv_id":null,"evidence_quote":"Supplies the Deep Candlestick Predictor framework that the Decomposer, CNN-autoencoder, and CNN1D components are adapted from."},{"cited_title":"Stock Price Movement Prediction from Financial News with Deep Learning and Knowledge Graph Embedding,","cited_arxiv_id":null,"evidence_quote":"TA-Lib provides the rule-based pattern detections that feed the custom function generating the strong/weak ground-truth labels."}],"review_version":1}