Fast Candlestick Patterns Detection with Limited Training Samples

Using RGB Gramian Angular Field and YOLO-LITE-V1

Heyang Huang
Stanford University
726 Serra Street, CA, 94305
[email protected]

Abstract

This report offers an automatic pipeline for training

effective candlestick pattern detectors with only limited
labeled training set. The only input data our pipeline needs
is a time series data of asset prices, along with thousands
or even hundreds labelled samples of candlestick patterns
in the form of “T5-T9 is Morning Star pattern.” The
pipeline consists of four modules: native pattern
classification, ARIMA data augmentation, auto-labeling,
Figure 1: Example of candlestick charts
and Yolo model training. The RGB-GAF pattern classifier
module first converts raw time series data into RGB
version of Gramian Angular Field (An enhanced GAF
One candlestick contains the open, close, highest and
algorithm designed by me), then trains a simple CNN
lowest price information in that timestep. As shown in
classifier on the Granmian Angular field images. The
figure 2 below,
ARIMA data augmentation module then fits an ARIMA
model on original time series and extrapolates the time
series many times longer. The exact length depends on the
length of original sequence. Then, the auto-labeling
module uses a moving window to classify whether a
segment of the extrapolated time series forms a certain
candlestick pattern. In this way, the module augments the
training set many times of its original size. Last, the Yolo
module pretrains and fits a Yolo-Lite module on the
augmented dataset. The fitted model yields a mAP > 0.4 of
0.481. The result is much better than it appears, since the
testing set labels are created by human beings and are
subjective. When we manually examine each mis-classified Figure 2: Information contained in a candlestick
cases, we in fact find most of them quite reasonable.
As mentioned by Tharavanij [1], “it is very common for
1. Introduction investors to use them in conjunction with other technical
indicators. In fact, according to a survey by Menkhoff
Candlestick chart was first developed in Japan and then (2010), fund managers apply it in their shorter-term
became popular around the world. It is a visualizable forecasts. Candlestick charting is unique in the sense that
tokenization of stock prices it concurrently plots daily open, high, low, and close price
movements Morris [2]. As such, it reveals demand and
supply changing balance and also investor sentiment and
psychology. Proponents of candlestick believe that
investors could use these chart patterns to predict short-
term price movements or future turning points.”
Since candlestick patterns are most useful for high- Auto-encoders (DA) on the GASF images of four
frequency trading, it’s natural to think about applying standard and one synthesized compound dataset. The
deep learning algorithms to automate the candlestick imputation MSE on test data is reduced by 12.18% –
pattern detection. There are three biggest challenge: lack 48.02%”
of training set, high-quality labeling, and a universal
quantifiable definition of patterns. The real-world In their work, they also compare the Gramian Angular
financial data is limited. Moreover, many candlestick Summation/ Difference model with Markov Transition
patterns are defined in a way that relies on people’s field encoder, and claims that the Gramian Anguar-based
discretion. How many consecutive bars can be seen as a encoder outperforms Markov transition field encoder in
trend? Therefore, the manual labeling process can be 16 out of 20 time series benchmarks they selected.
time-consuming and error prone.

My proposed pipeline can effectively augment enough I further work on their approach. The weakness of their
number of stock prices in the form as time series thanks to method is mentioned in Xiang [5]:
the RGB-GAF transformation module which will be later
discussed in methodology section. We will then have “Let us take a normalized time series X and use the GAF
enough labeled samples to train a YOLO-LITE model to method to imagine it. Despite the presence of both sine
directly locate and classify candlestick patterns on chart and cosine functions in the initial work, now themajority
pictures. The ability to classify directly on chart photo of examples I reviewed use only the cosine function to get
rather than numeric time series form is very important. It the GAF image. Note that for any θ in [0,2𝜋), we have
mimics how real traders think and grant more adaptability also (2𝜋-θ) in [0,2𝜋). Recall that the cosine function in
to classify candlestick charts in various timeframes [0,2𝜋) is symmetric with respect to θ=2𝜋 so that cos(θ)
without major change of code structure. To summarize, equals cos(2𝜋-θ). Back to the GAF matrix, we see that
my proposed pipeline tackles the following three biggest cos(2𝜋-θi-θj)=cos(θi+θj) and therefore 𝜋-θi and 𝜋-θj give
problems in deep learning-based candlestick detection: us the same value as the i-j element of the GAF matrix.
For all θ in [0,𝜋), we know that 𝜋-θ=arccos(-x) and we
1. Not enough training samples available can therefore get the fact that the time series X can give
2. Needs for manual labeling us the same matrix. To resume, if we reverse the sign of
3. Based on numerical values rather than picture, so every point on a time series, the transformation of GAF
require complete rework when the timeframe or results in the same image.”
price range changes.
My solution to this issue is simple: I stack the sine and
2. Related Work cosine GAF as well as GAD graph together and get a
RGB rather than gray scale image. This RGB scales can
show the downward or upward trend with different colors.
2.1. Related work in Gramian Angular Field
Encoding
Given a segment of time series, how do we test whether it 2.2. Related work in fast Object Detection Model
fits into a candlestick pattern? While traditional machine and Pretraining on Abstract photos.
learning and RNN model fails to do a good job, Chen [3]
proposes that we can use Gramian Angular Field to
What object detection model works best for fast object
transform the time series data into 2 dimensional pictures
detection? The financial markets require every detection
and run CNN classifier on it. Chen claims that the CNN
model to be fast, since some of the candlestick algorithms
classifier achieves a n accuracy of 92.5% on simulated
work in min or even second timeframe. As Joseph [6]
data vs 87.2 for traditional machine learning models.
mentioned in YOLO (You Only Look Once), base YOLO
Wang [4] also supports the effectiveness of Gramian
model processes images in real-time at 45 frames per
Angular encoding on time series classification problems:
second. A smaller version of the network, Fast YOLO,
processes an astounding 155 frames per second while still
“We used Tiled Convolutional Neural Networks (tiled
achieving double the mAP of other real-time detectors.
CNNs) on 20 standard datasets to learn high-level features
Compared to state-of-the-art detection systems, YOLO
from the individual and compound GASF-GADF-MTF
makes more localization errors but is less likely to predict
images. Our approaches achieve highly competitive
false positives on background. Finally, YOLO learns very
results when compared to nine of the current best time
general representations of objects. It outperforms other
series classification approaches. Inspired by the bijection
detection methods, including DPM and R-CNN, when
property of GASF on 0/1 rescaled data, we train Denoised
generalizing from natural images to other domains like
artwork. YOLO is both fast in recognition and pre-train.
Moreover, it can achieve good result on non-natural
photos.

YOLO_LITE, designed by Huang [7], which focuses on

fast classification “runs at about 21 FPS on a non-GPU
computer and 10 FPS after implemented onto a website
with only 7 layers and 482 million FLOPS. This speed is
3.8 × faster than the fastest state of art model, SSD
MobilenetvI.” YOLO-LITE doesn’t add prune methods
or special designed convolution layer to boost
performance, instead it focuses on the optimization tricks Figure 3: Model Pipeline
available that gears towards fast classification.
3.1. RGB-GAF Pattern classification
Similarly, Zhang [8], Landola [9], Howard [10] also
provides solutions to the speed up the pretraining and As mentioned in Chen [3], the Gramian Angular Field
detection phase on YOLO. Their work finds ways to Algorithm first converts time series to polar
design original convolution layer and pruning methods to representations.
cut down number of parameters in network.

So I built upon Zhang [8]’s shufflenet and incorporate the

optimization module used in Huang[7] to formulate the
updated version of YOLO-LITE-V1. However it doesn’t
seem to outperform the vanilla YOLO-LITE model,
maybe due to some inefficient implementation I made. Then, algorithm sums up degrees to calculate the cosine
So I decide to use YOLO-LITE as my object detection function(Gramian Angular Summation Field).
model.

3. Methods
The pipeline consists of four modules: GAF pattern
classification, ARIMA data augmentation, auto-labeling,
and Yolo model training, as shown below in figure 3. The
GAF pattern classifier module first converts candlesticks
into an enhanced version of Gramian Angular Field, then
trains a simple CNN classifier on the Granmian Angular
field images. The ARIMA data augmentation module then
fits an ARIMA model on original time series and Figure 4: Vanilla Gramian Angular Field Conversion
extrapolates the time series many times longer, given the
length of original sequence. Then, the auto-labeling Following the step above, I reproduce the following
module uses a moving window to classify whether a GASF representation as mentioned by Chen [3] below in
segment of the extrapolated time series forms a certain Figure 5 and 6. If we compare them, we can find the
candlestick pattern. In this way, the module augments the angular field representation does capture the fundamental
training set many times of its original size. Last, the Yolo difference in the two patterns and show it in the form of
module fits a Yolo-Lite module on the augmented dataset. color gradient shifts.

Figure 5 Vanilla Representaion of Bearish Engulfing

 layer shape. We use those CNNS to classify the RGB-
GAF with the respective resolution

Figure 6: Vanilla Representation of Morning Star

However, the weakness of this method is mentioned in

Xiang [5]:
“Note that for any θ in [0,2𝜋), we have also (2𝜋-θ) in
[0,2𝜋). Recall that the cosine function in [0,2𝜋) is
symmetric with respect to θ=2𝜋 so that cos(θ) equals
cos(2𝜋-θ). Back to the GAF matrix, we see that cos(2𝜋- Figure 8: Simple CNN Used for RGB-GAF Classification
θi-θj)=cos(θi+θj) and therefore 𝜋-θi and 𝜋-θj give us the of size 10*10*3
same value as the i - j elements of the GAF matrix. For all
θ in [0,𝜋), we know that 𝜋-θ=arccos(-x) and we can
therefore get the fact that the time series X can give us the 3.2. ARIMA data augmentation
same matrix. To resume, if we reverse the sign of every For the time series, we apply the ARIMA algorithm as
point on a time series, the transformation of GAF results mentioned below in Figure 10
in the same image.”

To resolve this issue, I propose the RGB-GAF algorithm:

stacking the sine version, or Gramian Angular Difference
Field, and cosine GASF graphs together and get a RGB
rather than gray scale image. This RGB scales can show
the downward or upward trend with different colors.

Figure 9: ARIMA

After we fit the ARIMA model on the original low price

series. We train another three ARIMA models on the
differences between low and open, close, high
respectively.
Figure 7: RGB GAF Representation of Morning Star
We then extrapolate on all of those four ARIMA series
and then add the three-difference series with the low price
series to derive the augmented time series. This way of
implementation is used to ensure high price will always
be the high price in each timestamp.
As shown in Figure 7, we can now differentiate the
downward trend from the upward trend of same
magnitude by the red/green color gradient. 3.3. Auto-labeling
We then transform the augmented data into RGB-GAF
Then, we apply the following simple CNN models to and run the fitted simple CNN , which is shown in Figure
classify the candlestick patterns as shown in Figure 8. The 9, to predict the label on augmented data, and thereby
network structure is inspired by Chen [3] and selected creates augmented data.
using K-fold validation. Noticed than the input shape is
not limited to (10,10,3), for different moving window
size, we use different input shape and corresponding input
3.4. YOLO-LITE Detection 4. Dataset and features
A summary of YOLO-LITE algorithm: As shown in figure 3, there are 2 datasets that we derive
from online source: training input data and testing data.
For each bounding box that the algorithm loops over and The augmented dataset is generated by ARIMA Simulator
propose, we calculate the confidence for that bounding and labeled by trained CNN model.
box and each object class, or the probability of an object
exists in a bounding box: 4.1. Training Dataset
The training dataset is used to train a simple CNN model
to label time series data with respective candlestick
patterns.
Then we transform C into class-specific conditional
probability: The timeseries data is derived from WRDS with 4h, 1h,
30min, and 5min intervals on currency and SPY futures.
It is roughly 800M in total.
An example is shown below in figure 4:

YOLO-LITE defines loss as

Figure 4: Example of Currency time series in 4h bin

For data preprocessing, I use a moving window of size 5,

10, 20, respectively and normalized the high, low, open
close price in range [0,1] for each segment, which is
mandate by Gramian Angular Field transformation.

The label data is stored in a separate txt file to indicate the

timestamp range corresponding to a candlestick pattern.
While the loss function is used to find the sweet spot of Candlestick patterns that I used are Morning Star, Bearish
center of bounding box, the mAP (for our case >0.4 is Reversal, Inverted Hammer, Bullish Reversal, Evening
used) is used to measure the prediction accuracy which Star, Bullish Abandoned Baby, Bearish Abandoned Baby,
defined as: piercing, which are encoded as 1,2,3,4,5,6,7, 8
respectively. Noticed that, all other unlabeled intervals
have label 0 has default value, indicating that there’s no
candlestick pattern in this interval. I manually labelled
200-300 for each pattern. An example can be found below
in Figure 5:
The most difficult part for YOLO-LITE implementation
is pretraining. I additionally generated 3500 segments of
the augmented data auto-labeled in previous module to
candlestick charts and change their background to the
same texture as in Yahoo-finance. Notice that those
pictures contain different number of candlestick charts but
has the same resolution 64*64*3. I then train the model Figure 5: Training dataset label
listed in the YOLO-LITE open-source repo with those
3500 labeled photos. 4.2. Testing Dataset
Testing dataset consists of 200 pictures that I cropped
from different online broker websites. All of them are
candlestick patterns that differ in size, scale, underlying
assets, date. I then change their background to the same
white-gray grid line as in Yahoo-finance and compress
their resolution to 64*64*3. I annotate the pictures
manually with online bounding box annotator.

5. Experiments and Results

5.1. Evaluation Metrics and baseline

The metric I use is mAP>0.4:

Figure: Confusion Matrix for YOLO-LITE

For baseline, there’s no existing benchmark to evaluate
how good a bounding box/ classification for candlestick
pattern is. Since the definition of candlestick pattern is
itself subject. So instead, I just compare whether the
detected pattern in picture is similar to my own
judgments.

However, I did implement a random forest classifier as

the metrics baseline to compare non-CNN model
performance with my RGB-GAF+CNN method. The
result is shown below in key result section.

5.2. Hyperparameters
After k-fold validation (k=10), the best setup of
hyperparameter for YOLO is listed below: Figure: validation loss for YOLO-LITE

lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)

lrf: 0.01 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.937 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.1 # warmup initial bias lr

The hyperparameter for Simple CNN families doesn’t

matter that much based on cross-validation.

5.3. Key results

Figure: Average training/validation loss for all
The mAP>0.4 for YOLO-LITE is 0.481. Simple CNNs with different moving window size
The average classification AUC_ROC for simple CNNs
are 0.71. I choose to use AUC_ROC because there are
much more case 0 than all others, I need to add the F-
score component to deal with imbalanced dataset.
 texture I use and bad quality labeling. So, the classifier
overfits to background.

If we examine the confusion matrix, we can realize that

the class 8: piercing pattern is most misclassified. This is
understandable because it is the only non-reversal
indicator we include. All other indicators serve to indicate
trend reversal while piercing indicates continuation of
previous trend.

The convergence performance of both YOLO-LITE

and simple CNNs are reasonable. Notice that loss of
simple CNNs are much better than random forester
Figure: Top YOLO_LITE Classification Result classifier and it also converges faster. Given the training
loss/validation loss plot, I don’t think there exists any
overfitting issues. However, the misclassification
examples clearly shows that the complexity of models
need to be increased or more training samples are needed.

6. Conclusion and future works

While there still exists cases that YOLO_LITE model
labels far-off pictures for reasons that I cannot
understand, the overall performance of both bounding box
precision and classification precision is already
reasonably good with a mAP>0.4 at 0.481. Part of the
error is due to imperfectness in manual labelled test sets.
My design of RGB Granian Angular Field, data ARIMA-
Figure: Top YOLO_LITE Classification Result CNN-based data augmentation, and model choice of
YOLO-LITE contributes to the performance to a great
extent. If more time, computational resource, and people
When examines misclassified test cases, some of them are available in future, we will increase the number of
are caused by including extra candlesticks in the pattern, precisions of testing labels and explore more completed
which shouldn’t be considered wrong. However, for other setup for object detection model.
misclassified cases, the YOLO algorithm often just bound
box one candlestick and randomly claims it to be a
pattern, like the figure below:
7. Contributions and acknowledgement
Thanks to friends who help me manually label candle
charts: Shijie Guo, Jenifer Lu, and Ren Li
Thanks to the open-source code provider:

1. Yolo-Lite Team https://reu2018dl.github.io/

2. GAF python codebase:

https://pyts.readthedocs.io/en/stable/generated/py
ts.image.GramianAngularField.html

3. WRDS Database https://wrds-

www.wharton.upenn.edu/
Figure: Misclassified YOLO_LITE Classification
. 4. Yahoo Finance for candlestick charts picture

I fail to find a solid explanation for these errors, since 5. TD_Ameritrade for candlestick charts picture
there are barely any samples in training set that consist of
only one bar. I guess it may be due to the background
8. References imperative-style-high-performance-deep-learning-
library.pdf.
[1] Tharavanij, P., Siraprapasiri, V. & Rajchamaha, K.
Performance of technical trading rules: evidence from
Southeast Asian stock markets. SpringerPlus 4, 552 (2015).
https://doi.org/10.1186/s40064-015-1334-7
[2] Morris, Greg L. Candlestick Charting Explained: Timeless
Techniques for Trading stocks and Sutures. McGraw Hill
Professional, 2006. FirstName Alpher and FirstName
Gamow. Can a computer frobnicate? In CVPR, pages 234–
778, 2005.
[3] Chen, Jun-Hao, και Yun-Cheng Tsai. ‘Encoding
Candlesticks as Images for Patterns Classification Using
Convolutional Neural Networks’. arXiv, 2019.
https://doi.org/10.48550/ARXIV.1901.05237.
[4] Wang, Zhiguang, and Tim Oates. "Imaging time-series to
improve classification and imputation." In Twenty-Fourth
International Joint Conference on Artificial Intelligence.
2015.
[5] Xiang, RGB GAF image: A possible solution to one weak
point of Gramian Angular Field Imaging, 2022

[6] Joseph Redmon, Santosh Divvala, Ross Girshick, Ali

Farhadi; Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 2016, pp. 779-788

[7] R. Huang, J. Pedoeem and C. Chen, "YOLO-LITE: A Real-

Time Object Detection Algorithm Optimized for Non-GPU
Computers," 2018 IEEE International Conference on Big
Data (Big Data), 2018, pp. 2503-2510, doi:
10.1109/BigData.2018.8621865.

[8] X. Zhang, X. Zhou, M. Lin and J. Sun, "Shufflenet: An

extremely efficient convolutional neural network for
mobile devices", 2017.

[9] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W.

J. Dally and K. Keutzer, "Squeezenet: Alexnet-level
accuracy with 50x fewer parameters and¡ 0.5 mb
model size", 2016.

[10] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W.

Wang, T. Weyand, et al., "Mobilenets: Efficient convo-
lutional neural networks for mobile vision applications",
2017.

[11] Virtanen, Pauli, Ralf Gommers, Travis E. Oliphant, Matt

Haberland, Tyler Reddy, David Cournapeau, Evgeni
Burovski, κ.ά. ‘SciPy 1.0: Fundamental Algorithms for
Scientific Computing in Python’. Nature Methods 17
(2020): 261–72. https://doi.org/10.1038/s41592-019-0686-
2.

[12] Paszke, Adam, Sam Gross, Francisco Massa, Adam Lerer,

James Bradbury, Gregory Chanan, Trevor Killeen, κ.ά.
‘PyTorch: An Imperative Style, High-Performance Deep
Learning Library’. Στο Advances in Neural Information
Processing Systems 32, επιμέλεια H. Wallach, H.
Larochelle, A. Beygelzimer, F. d\textquotesingle Alché-
Buc, E. Fox, και R. Garnett, 8024–35. Curran Associates,
Inc., 2019. http://papers.neurips.cc/paper/9015-pytorch-an-