Pith. sign in

REVIEW 4 major objections 6 minor 24 references

Investigating Market Strength Prediction with CNNs on Candlestick Chart Images

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims candlestick patterns detected by YOLOv8 add no predictive value to a CNN that classifies market strength from chart images alone.

desk verdict 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. read the letter →

arxiv 2501.12239 v1 pith:SX6E26FU submitted 2025-01-21 cs.CV

Thanh Nam Duong , Trung Kien Hoang , Quoc Khanh Duong , Quoc Dat Dinh , Duc Hoan Le , Huy Tuan Nguyen , Xuan Bach Nguyen , Quy Ban Tran This is my paper · ORCID
classification cs.CV
keywords candlestickpatternsmarketstrengthpredictionconvolutionalneuralnetworkYOLOv8chartimageclassificationdeeplearningforfinancetechnicalanalysistime-seriesvs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

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.

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 (4)
  1. [III.A.2]
  2. [Tables II and III; Abstract]
  3. [IV (Pure CNN, Include-Pattern)]
  4. [IV (General methodology)]
minor comments (6)
  1. [Tables I, II, IV, V] The ticker 'AAPL' is consistently misspelled as 'APPL' in the tables; please correct.
  2. [Last pages] 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.
  3. [References] 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.
  4. [II.C] The heading 'Cross-Market Anaalysis' contains a typo; it should read 'Cross-Market Analysis.'
  5. [IV (Decomposer)] 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.
  6. [III.B.2] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the study is an empirical model comparison, and the pattern-vs-non-pattern contrast is operationally defined and externally falsifiable.

full rationale

This is an empirical supervised-learning study with no derivation chain, no equations, and no fitted parameter renamed as a prediction. The central comparison — "include-pattern" vs "non-pattern" — is operationally defined: both variants use the same VGG16 backbone and the same strength labels, differing only in whether the second stream consumes pattern images (Tables II and III). The strength labels are generated by an unspecified custom TA-Lib-based function (Section III.A.2), which is a serious reproducibility and validity concern because noisy or leaky labels would affect every model equally; however, shared labels do not make the pattern-vs-non-pattern comparison circular. The "Force Teaching" procedure (using exact pattern images in training while YOLO detection is used at inference, Section IV) is a train/test mismatch, not a reduction of the result to its inputs. Citations to Deep Candlestick Predictor are used only as architectural inspiration, and the paper explicitly states that "detailed specifics on the decomposition process were not provided in the original paper" and that a custom Decomposer was implemented (Section III.B.1.a), so no load-bearing claim is imported from prior work. The abstract's negative claim is internally contradicted by Table II, which shows BTC-USD with-pattern accuracy 0.792 versus without-pattern 0.765, but internal inconsistency is a correctness and statistics issue, not circularity. No self-citation chain, uniqueness assertion, or ansatz-smuggling is present. Accordingly, no circular step can be exhibited with the required quote-and-reduction specificity; the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central empirical claims depend almost entirely on unstated choices in data labeling and model configuration, including an unspecified custom label function and a hand-picked history window. No code or data is provided, and no new physical or conceptual entities are introduced.

free parameters (3)
  • strength-label threshold = not reported
    The custom function that labels future trends as strong or weak must use some threshold on future price movement; the threshold is not given, so every accuracy figure depends on an unstated choice.
  • history window length = 30 candles
    Section III.A.2 states that 30 past candles are selected whenever a pattern appears; this hand-picked window defines the input images and the label context.
  • training hyperparameters = not reported
    Optimizer, learning rate, epochs, batch size, image size, and YOLO confidence thresholds are not reported, so the exact trained models cannot be reproduced.
assumptions (4)
  • domain assumption TA-Lib candlestick pattern definitions are a valid ground truth for strong and weak upcoming trends.
    Section III.A.2 uses TA-Lib to identify patterns and then labels strength; if these rule-based patterns do not predict future movement, the learning task is arbitrary.
  • domain assumption Daily OHLC data from Yahoo Finance is accurate and sufficient for the task.
    Section III.A.1 collects daily prices via Yfinance; any data quality or adjustment issues propagate into the images and labels.
  • domain assumption Converting charts to images preserves the information needed for strength prediction.
    The whole method feeds pixel representations rather than time series; the paper itself notes in the Discussion that image conversion sacrifices temporal context, so this assumption is load-bearing.
  • domain assumption The train/test split is free of temporal leakage and class imbalance was handled or accounted for.
    No split, shuffle policy, or class-balance information is reported; without this, the accuracy numbers could reflect leakage or majority-class bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Investigating Market Strength Prediction with CNNs on Candlestick Chart Images." pith.science (2026). https://pith.science/paper/SX6E26FU

@misc{pith2026250112239,
  author       = {Pith},
  title        = {Pith review of: Investigating Market Strength Prediction with CNNs on Candlestick Chart Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SX6E26FU}},
  note         = {Machine review of arXiv:2501.12239}
}
read the original abstract

This paper investigates predicting market strength solely from candlestick chart images to assist investment decisions. The core research problem is developing an effective computer vision-based model using raw candlestick visuals without time-series data. We specifically analyze the impact of incorporating candlestick patterns that were detected by YOLOv8. The study implements two approaches: pure CNN on chart images and a Decomposer architecture detecting patterns. Experiments utilize diverse financial datasets spanning stocks, cryptocurrencies, and forex assets. Key findings demonstrate candlestick patterns do not improve model performance over only image data in our research. The significance is illuminating limitations in candlestick image signals. Performance peaked at approximately 0.7 accuracy, below more complex time-series models. Outcomes reveal challenges in distilling sufficient predictive power from visual shapes alone, motivating the incorporation of other data modalities. This research clarifies how purely image-based models can inform trading while confirming patterns add little value over raw charts. Our content is endeavored to be delineated into distinct sections, each autonomously furnishing a unique contribution while maintaining cohesive linkage. Note that, the examples discussed herein are not limited to the scope, applicability, or knowledge outlined in the paper.

Figures

Figures reproduced from arXiv: 2501.12239 by the authors.

Figure 1
Figure 1. Two stream CNN architecture [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Simple CNN architecture In the main stream, we employ a CNN architecture to extract features directly from the candlestick chart images. This stream focuses on capturing information from the candlestick charts, allowing us to analyze the primary data source comprehensively. The features extracted from this stream provide essential insights into the underlying market trends The pattern stream is dedicated to extracti… view at source ↗
Figure 3
Figure 3. Loss and accuracy of DCP model [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: F1-Score of DCP model Furthermore, there exists a necessity to engage in systematic experimentation with and refinement of diverse methodological approaches, aimed at attaining heightened accuracy scores. Such approaches encompass potential variations in model architec…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 18 canonical work pages

  1. [1]

    We selected stocks(AAPL), exchange(EUR_USD), and crypto(Bitcoin) as our trading instruments of interest to get data for our train model and perform p redictions

    Data collection To gather necessary historical data, we accessed the financial market’s API provided by Yfinance Exchange. We selected stocks(AAPL), exchange(EUR_USD), and crypto(Bitcoin) as our trading instruments of interest to get data for our train model and perform p redictions. The time data we collected is from 2017 up until 30 October 2023. The re...

    work page 2017

  2. [2]

    These patterns are then labeled as either strong or weak, providing a vital ground truth for the model's learning process

    Data processing Subsequently, a specialized custom function is employed to identify specific candlestick patterns that effectively define the strength of impending trends within the cryptocurrency market. These patterns are then labeled as either strong or weak, providing a vital ground truth for the model's learning process. We use Talib library[19] to d...

    work page

  3. [3]

    Real -Time Flying Object Detection with YOLOv8,

    D. Reis, J. Kupec, J. Hong, and A. Daoudi, “Real -Time Flying Object Detection with YOLOv8,” May 2023, doi: https://doi.org/10.48550/arxiv.2305.09972

    arXiv doi:10.48550/arxiv.2305.09972 2023

  4. [4]

    (Region -based Convolutional Neural Networks). Additionally, we experiment with techniques that convert candlestick chart images into time series data, allowing us to leverage specialized libraries and tools to further enhance pattern recognition In terms of predicting the strength of market trends, we employ a combination of deep learning models, includi...

    work page

  5. [5]

    Deep Candlestick Predictor,

    DCP Our approach builds upon the framework introduced in the paper "Deep Candlestick Predictor," with tailored adaptations and modifications to address the specific requirements of our work. In this section, we present the key components of our methodology, in cluding the Decomposer, the CNN - autoencoder, and the CNN1D, emphasizing the distinctive change...

    work page

  6. [6]

    Non-Pattern

    CNN a) Two-stream CNN Furthermore, in our research, we strategically incorporated an alternative methodology known as the Two -Stream Convolutional Neural Network (CNN). This distinctive approach is characterized by the integration of two distinct streams, each uniquely designed to fulfill specific purposes within the framework of our study. Fig. 1. Two s...

    work page

  7. [7]

    An Introduction to Convolutional Neural Networks

    Keiron O'Shea and Ryan Nash, “An Introduction to Convolutional Neural Networks”, arXiv (Cornell University), Nov. 2015, doi: https://doi.org/10.48550/arxiv.1511.08458

    arXiv doi:10.48550/arxiv.1511.08458 2015

  8. [8]

    Deep Feature Fusion via Two -Stream Convolutional Neural Network for Hyperspectral Image Classification,

    X. Li, M. Ding, and A. Pizurica, “Deep Feature Fusion via Two -Stream Convolutional Neural Network for Hyperspectral Image Classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 4, pp. 2615–2629, Apr. 2020, doi: https://doi.org/10.1109/tgrs.2019.2952758

    work page arXiv 2020

Show all 24 references
  1. [9]

    Faster R -CNN: Towards Real- Time Object Detection with Region Proposal Networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster R -CNN: Towards Real- Time Object Detection with Region Proposal Networks,” arXiv (Cornell University), Jun. 2015, doi: https://doi.org/10.48550/arxiv.1506.01497

    2015 doi arXiv

  2. [11]

    Dynamic Deep Convolutional Candlestick Learner,

    J. Chen and Y. Tsai, “Dynamic Deep Convolutional Candlestick Learner,” arXiv (Cornell University), Jan. 2022, doi: https://doi.org/10.48550/arxiv.2201.08669

    2022 doi arXiv

  3. [12]

    Deep Candlestick Predictor: A Framework toward Forecasting the Price Movement from Candlestick Charts,

    S.-J. Guo, F. Hsu, and C. Hung, “Deep Candlestick Predictor: A Framework toward Forecasting the Price Movement from Candlestick Charts,” Dec. 2018, doi: https://doi.org/10.1109/paap.2018.00044

    2018

  4. [13]

    DPP: Deep predictor for price movement from candlestick charts,

    C.-C. Hung and Y. -J. Chen, “DPP: Deep predictor for price movement from candlestick charts,” PLOS ONE, vol. 16, no. 6, p. e0252404, Jun. 2021, doi: https://doi.org/10.1371/journal.pone.0252404

    2021 doi

  5. [14]

    Predicting the price movement from candlestick charts: a CNN -based approach,

    C. C. Hung, Y. J. Chen, S. J. Guo, and F. C. Hsu, “Predicting the price movement from candlestick charts: a CNN -based approach,” International Journal of Ad Hoc and Ubiquitous Computing, vol. 34, no. 2, p. 111, 2020, doi: https://doi.org/10.1504/ijahuc.2020.107821

    2020

  6. [15]

    Predicting Short Trend of Stocks by Using Convolutional Neural Network and Candlestick Patterns,

    Kietikul Jearanaitanakij and Bundit Passaya, “Predicting Short Trend of Stocks by Using Convolutional Neural Network and Candlestick Patterns,” Oct. 2019, doi: https://doi.org/10.1109/incit.2019.8912115

    2019

  7. [16]

    Forecasting stock prices with a feature fusion LSTM-CNN model using different representations of the same data,

    T. Kim and H. Y. Kim, “Forecasting stock prices with a feature fusion LSTM-CNN model using different representations of the same data,” PLOS ONE, vol. 14, no. 2, p. e0212320, Feb. 2019, doi: https://doi.org/10.1371/journal.pone.0212320

    2019 doi

  8. [17]

    Using Deep Learning Neural Networks and Candlestick Chart Representation to Predict Stock Market (Feb

    Rosdyana Mangir Irawan Kusuma, Trang -Thi Ho, Wei -Chun Kao, Yu - Yen Ou, Kai -Lung Hua, 2019. Using Deep Learning Neural Networks and Candlestick Chart Representation to Predict Stock Market (Feb

    2019

  9. [18]

    Deep Learning for Price Movement Prediction Using Convolutional Neural Network and Long Short -Term Memory,

    C. Yang, J. Zhai, and G. Tao, “Deep Learning for Price Movement Prediction Using Convolutional Neural Network and Long Short -Term Memory,” Mathematical Problems in Engineering, vol. 2020, pp. 1 –13, Jul. 2020, doi: https://doi.org/10.1155/2020/2746845

    2020 doi

  10. [19]

    Stock Price Movement Prediction from Financial News with Deep Learning and Knowledge Graph Embedding,

    Y. Liu, Q. Zeng, and H. Yang, “Stock Price Movement Prediction from Financial News with Deep Learning and Knowledge Graph Embedding,” Lecture Notes in Computer Science, vol. 11016, Jul. 2018

    2018

  11. [20]

    Stock Price Prediction Using CNN and LSTM - Based Deep Learning Models,

    S. Mehtab and J. Sen, “Stock Price Prediction Using CNN and LSTM - Based Deep Learning Models,” 2020 International Conference on Decision Aid Sciences and Application (DASA), Nov. 2020

    2020

  12. [21]

    A Stock Price Prediction Model Based on Investor Sentiment and Optimized Deep Learning,

    G. Mu, N. Gao, Y. Wang, and D. Li, “A Stock Price Prediction Model Based on Investor Sentiment and Optimized Deep Learning,” vol. 11, pp. 51353–51367, Jan. 2023, doi: https://doi.org/10.1109/access.2023.3278790

    2023

  13. [22]

    Stock Chart Pattern recognition with Deep Learning,

    M. Velay and F. Daniel, “Stock Chart Pattern recognition with Deep Learning,” arXiv (Cornell University), Aug. 2018, doi: https://doi.org/10.48550/arxiv.1808.00418

    2018 doi arXiv

  14. [23]

    A Stock Closing Price Prediction Model Based on CNN -BiSLSTM,

    H. Wang, J. Wang, L. Cao, Y. Li, Q. Sun, and J. Wang, “A Stock Closing Price Prediction Model Based on CNN -BiSLSTM,” Complexity, vol. 2021, pp. 1–12, Sep. 2021, doi: https://doi.org/10.1155/2021/5360828

    2021 doi

  15. [25]

    Please ensure that all template text is removed from your conference paper prior to submission to the conference

    https://github.com/TA-Lib/ta-lib-python IEEE conference templates contain guidance text for composing and formatting conference papers. Please ensure that all template text is removed from your conference paper prior to submission to the conference. Failure to remove template ...

  16. [2019]

    https://doi.org/10.48550/arXiv.1903.12258

    1903 doi arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.