PaperlayerView exact revision on arXiv

Source-linked AI summary

Explainable Deep Convolutional Candlestick Learner

Jun-Hao Chen, Samuel Yen-Chi Chen, Yun-Cheng Tsai, Chih-Shiang Shur

arXiv:2001.02767v4cs.LGcs.CVstat.ML

TL;DR

Traders cannot verify what deep models learn when recognizing candlestick patterns. The paper uses GASF encoding and local-search attacks to inspect a CNN’s reasoning, and reports that the learned patterns resemble human perception.

  • Problem

    Traders lack explanations for how deep models recognize candlestick patterns, making it difficult to trust their precise recognition.

    Source: §1 · p. 1

  • Method

    The authors encode OHLC time series as GASF matrices for a CNN and adapt local-search attacks that perturb diagonal elements to probe classification.

    Source: §2.1 · p. 3, §2.2 · p. 3

  • Results

    64.36% average attack success; the authors interpret the attacks as showing that the GASF-CNN recognizes diagonal patterns similarly to human perception.

    Source: §3 · p. 4, §4 · p. 4

  • Takeaways & Limitations

    The authors argue that the explainable GASF-CNN is more trustworthy and reliable for traders than models whose learned reasoning is unknown.

    Source: §4 · p. 4

Abstract

from arXiv · show

Candlesticks are graphical representations of price movements for a given period. The traders can discovery the trend of the asset by looking at the candlestick patterns. Although deep convolutional neural networks have achieved great success for recognizing the candlestick patterns, their reasoning hides inside a black box. The traders cannot make sure what the model has learned. In this contribution, we provide a framework which is to explain the reasoning of the learned model determining the specific candlestick patterns of time series. Based on the local search adversarial attacks, we show that the learned model perceives the pattern of the candlesticks in a way similar to the human trader.

1 Introduction

Candlestick recognition is central to technical analysis, but traders cannot verify what deep models learn without explanations. The paper introduces an adversarial-attack framework to inspect the model’s pattern reasoning.

1 Introduction

  • Traders need candlestick recognition tools, but deep models’ lack of explainability leaves them unable to trust what the models recognize precisely.The paper presents its attack-based framework as a way to examine model reasoning.Source: §1 · p. 1

1.1 Candlestick Pattern

  • Candlesticks encode opening, highest, lowest, and closing prices; body placement and color indicate whether the close is above or below the open.A short candle has similar opening and closing prices.Source: §1.1 · pp. 1–2

1.2 Time Series System

  • Time-series charts place time on the horizontal axis and price on the vertical axis; chart cycles define each candle’s interval, from minutes to weeks.A five-minute candle represents the price change over five minutes.Source: §1.2 · p. 2

1.3 The 8 Most Powerful Candlestick Patterns

  • The paper examines eight common candlestick patterns, including the Morning Star and Evening Star as well as six other named formations.The Morning Star and Evening Star are three-candle patterns.Source: §1.3 · p. 2

1.4 Explain our Model

  • GASF encoding feeds an eight-pattern CNN that reached 90.0% accuracy; local search perturbs matrix diagonals, with 1,500 matrices attacked per label.The attack covers 10 bars: the last three form the OHLC patterns and the remaining bars represent the trend.Source: §1.4 · p. 3

2 Methods

The method encodes OHLC time series as GASF matrices for CNN classification, then uses diagonal-only local search perturbations to probe the model’s decision-relevant regions.

2.1 GASF-CNN

  • The GASF-CNN pipeline converts OHLC time series into symmetric matrices using a polar-coordinate encoding, then feeds the transformed data to a CNN.The series is normalized before angular conversion and the GASF construction.Source: §2.1 · p. 3
  • The CNN architecture uses two convolutional layers with 16 kernels and one fully connected layer with 128 units.Source: §2.1 · p. 3

2.2 Local Search Attack

  • The local search attack perturbs only diagonal GASF elements, recalculates corresponding non-diagonal values, and tests the resulting matrix with the CNN until misclassification.Restricting perturbations to the diagonal is intended to maintain consistency between the time series and its GASF matrix.Source: §2.2 · p. 3

3 Results

On EUR/USD one-minute data, the modified local search attack achieved a 64.36% average success rate. The results support diagonal attack regions and indicate that the model uses the final three bars for patterns and earlier bars for trend.

3 Results

  • The evaluation used EUR/USD one-minute OHLC data, with training from January 1, 2010 to January 1, 2016 and testing from January 2, 2016 to January 1, 2018.The passage also reports eight pattern labels with 1,500 samples each and 3,000 samples for label 0.Source: §3 · p. 3
  • 64.36% average attack success rate was reached with the modified local search attack, and the full results include attack rates of at least 40.0%.Table 2 presents attack ratios by label.Source: §3 · p. 4, Table 2 · p. 4
  • The results suggest diagonal attack regions are plausible, with the model recognizing patterns in the last three bars and trend in the remaining bars.Source: §3 · p. 4

4 Discussion

Attacks on morning star and evening star examples altered bars enough to make each pattern invisible or violate its rules. The authors conclude that the attacks reveal GASF-CNN features resembling human pattern perception.

  • In the morning star examples, the charts changed little overall, but the 8th and final bars became small enough to make the pattern invisible.Figure 5 compares the original pattern on the left with the attacked pattern on the right.Source: §4 · p. 4, Figure 5 · p. 4
  • In the evening star example, the 9th bar became too large and the final bar too small, violating the three-bar pattern rules and making the pattern invisible.Source: §4 · p. 4
  • The two empirical examples are presented as evidence that local search attacks reveal GASF-CNN learning in a way resembling human perception.The authors describe the resulting model as more trustworthy and reliable for traders than models whose learning experience is unknown.Source: §4 · p. 4

5 Conclusion

The conclusion presents GASF-CNN for candlestick recognition and modified local-search adversarial attacks as a way to explain its pattern decisions. The authors interpret attack results as human-like pattern perception, while noting that confirming this interpretation requires further analysis.

  • The paper proposes GASF-CNN for candlestick recognition and modified local-search adversarial attacks to explain its pattern decisions.Source: §5 · p. 4
  • The GASF-CNN identifies eight candlestick types, and attack results suggest it perceives patterns as a human does.Source: §5 · pp. 4–5
  • Figure 6 presents an attack-result example for the evening-star pattern.Source: Figure 6 · p. 5
  • The authors provide an open-source implementation of the local-search adversarial attack approach.Source: §5 · p. 5
Loading 2001.02767v4…