Hello! Welcome to the 14th Tobig's Conference.
Our project title is Bigscoin.
Bigscoin provides users with personalized Bitcoin pattern notifications and interpretable prediction services.
To add explainability and reliability to the prediction of deep learning models which are considered black boxes,
we developed a visualizing system that allows you to see Bitcoin charts patterns and price predictions at a glance using XAI techniques.
14th Tobig's TimeSeries and Explainable AI Conference
Coin investors often set up investment strategies by referring to the candlestick chart patterns.
Among the chart patterns, five most observed patterns in coin charts are as follows.
But the problem is that monitoring chart patterns all day long is time consuming and labor intensive.
In addition, traditional chart pattern recognition services are not clear and trustworthy
because the decision making process is impenetrable.
To address this inconvenience, we used Conv2d-based model combined with Grad-Cam in our real-time chart pattern alert service.
- Crawling high, low, market and closing price in 5 minute candlestick chart using "pyupbit" package [August 2017 ~ May 2022]
- Transforming price data into candlestick chart images as model input
- Labeling image data into 5 different chart patterns
We utilized Grad-Cam to visually interpret the Conv2d-based model output.
You can see the why the model chose to make the prediction by looking at the Grad-Cam results.
Users can easily figure out whether the pattern classified by the model is trustworthy.
Users who receive pattern notifications start to think about their investment strategies,
especially when they get triangle patterns where ups and downs are unpredictable.
N-BEATS, an explainable time series model, can help users establish their strategies in a reliable way.
- Crawling high, low, market and closing price in 5 minute candlestick chart using "pyupbit" package
We use N-BEATS model, a deep learning architecture that decomposes its forecast into two distinct components, trend and seasonality, to provide explanation in the model.
Drop-out is added to the model to visualize confidence interval for population mean of the predictions.
- Members of ToBig's (Big Data Analysis & Artificial Intelligence Organization) participated in this project.
| Year | Name | Team | Contribution |
|---|---|---|---|
| 16th year | Gwonho Kim | Web Serving | Project Leader, Backend |
| 16th year | Hanna Park | Web Serving | Frontend & Backend |
| 16th year | Yoone Kim | Web Serving | Frontend & Backend |
| 16th year | Jooho Kim | Regression | N-BEATS Visualization & Model Tuning, Data Augmentation |
| 17th year | Hyuntai Kim | Regression | DeepAR, N-BEATS, Informer Modeling & Experiments |
| 17th year | Sanyoon Kim | Regression | ARIMA Modeling & DeepAR Tuning |
| 16th year | Yerim Lee | Classification | Classification Modeling & Grad-CAM Visualization, Presentation |
| 17th year | Seyeon Rha | Classification | Experiments |
| 17th year | Heonwoo Yoo | Classification | Domain Knowledge & Data Inspection |