Search historical market data for similar time ranges using Dynamic Time Warping distance.
Given an end of range timestamp, a start time of day, and a lookback range in days, we will search for similar time ranges, and return the top N matches. The practical use of this would be to run this at the start of day, using now() for the end of time range.
Market data is provided by the user. It should be in the form of OHLCV with a UTC timestamp. The data used here came from databento.com.
The matches are scored using one of several variations of DTW. The default method will run DTW on the highs, lows, and close, and return the average distance, with double weight on the close (hlc4). Lower distance is better.
The matches can be plotted with Tradingview lightweight charts by using the widget. Projection bands from the target window can be also be plotted.
These main entrypoint is the Jupyter notebook notebooks\search.ipynb.
This project was initially generated using the Cookiecutter Data Science template, then trimmed down.
├── LICENSE
├── Makefile <- Makefile for convenience commands like `make data` or `make train`
├── README.md
├── docs
├── data
│ ├── processed <- The final, canonical data sets for modeling.
│ └── raw <- The original, immutable data dump.
├── market_span_cluster <- Source code for use in this project
├── notebooks
│ └── search.py <- search.py the main app.
├── pyproject.toml
└── requirements.in
└── requirements.txt
python -m venv venv
source venv/bin/activate # On Windows: venv\Scripts\activate
pip install -e To run the webapp:
streamlit run market_span_cluster/app.pyWhen adding a new dependency, update the requirements.in file and run:
pip-compile requirements.in
pip-sync # this doesn't update the packages in the .venv like it should
pip install -r requirements.txt # this should be handled by pip-sync- Jade Koskela
- Claude.ai 3.5 Sonnet
- ChatGPT 4.o