Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
The dataset viewer is not available for this split.
Server error while post-processing the rows. This occured on row 1. Please report the issue.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Synthetic Chart Patterns

A synthetic dataset of 50,000 price-chart images with object-detection labels for 11 classic chart patterns. It is intended for experiments with detecting chart patterns in images; it is not real historical market data and is not evidence that these patterns predict future prices.

Example charts

These are actual generated training images. The images themselves do not have boxes drawn on them; their matching YOLO label files contain the bounding-box coordinates.

Double top Ascending triangle
Synthetic double top chart Synthetic ascending triangle chart
Flag Pennant
Synthetic flag chart Synthetic pennant chart

Annotated overview

The yellow boxes below are drawn for inspection only; they are not part of the training images.

Sample charts with bounding boxes

Contents

  • 45,000 training images and 5,000 validation images, rendered at 1280 × 720 pixels.
  • Each image has one automatically generated bounding box and a corresponding YOLO-format text label.
  • Images vary in pattern geometry, noise, simulated gaps and wicks, price scale, candle count, and chart presentation, including dark and light backgrounds.
  • Classes are distributed almost evenly: 4,090–4,091 training examples and 454–455 validation examples per class.
  • train.zip and val.zip contain the images and labels. data.yaml contains the class mapping for Ultralytics; the paths are relative to the extracted dataset directory.
ID Class
0 head_and_shoulders
1 inverse_head_and_shoulders
2 double_top
3 double_bottom
4 ascending_triangle
5 descending_triangle
6 symmetrical_triangle
7 rising_wedge
8 falling_wedge
9 flag
10 pennant

How the dataset was made

A Python generator starts with template turning points for each pattern and randomly varies their placement and height. It interpolates a price path, applies a small slope and mean-reverting noise, and adds surrounding price movement. Simulated open, high, low and close prices provide the candle bodies and wicks; occasional gaps and longer wicks add variation. Matplotlib renders each chart in a TradingView-inspired style, and the generator calculates a bounding box from the pattern's candle and price range.

The noise standard deviation is sampled from 2–12% of a per-image pattern amplitude, with an additional cap of 5% of that pattern's own price range. This is not a percentage of the asset price. Training and validation use different random seed streams derived from seed 42. Validation is synthetic and generated by the same method as training.

Label format

Each image has a matching .txt file containing one line:

class_id x_center y_center width height

Coordinates and dimensions are normalized relative to the image size. For example, class 0 is head_and_shoulders. Bounding boxes are generated from the known pattern interval, not manually reviewed.

Download and use

Download train.zip, val.zip, and data.yaml from this repository. Extract both ZIP files into the same directory as data.yaml:

unzip train.zip
unzip val.zip

This creates images/train, images/val, labels/train, and labels/val. The resulting directory is compatible with the YOLO detection format and can be used with Ultralytics. The archives contain PNG images and text labels, not raw OHLC time series.

Checks and limitations

File counts and image/label pairing were checked for the entire dataset; label bounds and class balance were checked for all labels. Image dimensions were checked on a sample of 300 images per split. A sample of generated price paths was compared with the generator's own noise-free templates; this is a consistency check, not an independent assessment of pattern quality or realism.

Patterns are based on a limited set of synthetic templates, so real-world charts may look very different. Every image has one positive label: there are no intentionally generated negative images with no pattern, and no multiple-pattern annotations. A model could learn generator-specific visual cues rather than generalizable patterns. Evaluate on independently sourced, properly labeled real-world charts before relying on it. This dataset is for research and experimentation, not financial advice.

License

CC0 1.0 Universal: the dataset is dedicated to the public domain to the extent possible under applicable law. See https://creativecommons.org/publicdomain/zero/1.0/.

Downloads last month
155