firmai/deltapy — explained in plain English
Analysis updated 2026-08-05 · repo last pushed 2023-09-19
Turn basic price-and-volume data into hundreds of engineered features for stock prediction models.
Apply feature generation pipelines to financial time series to reduce prediction error.
Generate new columns from sales, sensor, or customer tabular data to boost model accuracy.
Run the hosted Colab notebook to experiment with feature engineering without any local setup.
| firmai/deltapy | karpathy/deep-vector-quantization | nvidia/cuopt-examples | |
|---|---|---|---|
| Stars | 556 | 647 | 452 |
| Language | Jupyter Notebook | Jupyter Notebook | Jupyter Notebook |
| Last pushed | 2023-09-19 | 2021-11-20 | — |
| Maintenance | Dormant | Dormant | — |
| Setup difficulty | easy | hard | moderate |
| Complexity | 3/5 | 4/5 | 3/5 |
| Audience | data | researcher | developer |
Figures from each repo's GitHub metadata at analysis time.
Hosted Google Colab notebook lets you try it with zero setup, though local install via pip is also straightforward.
DeltaPy is a toolkit that helps you squeeze more predictive power out of your existing data. Instead of going out and collecting new data, it generates new columns and rows from what you already have, giving your machine learning models more signal to work with. It was built with finance use cases in mind, but the approach works for any data-rich prediction task. The core idea is a five-step pipeline: transforming, interacting, mapping, extracting, and synthesising. Transforming takes a single column and creates new variants of it, like smoothing out noise or applying fractional differencing. Interacting combines multiple columns together, such as multiplying or dividing them, or finding relationships through regression. Mapping compresses or reprojects your data into fewer dimensions. Extraction pulls out dozens of statistical properties from your data, things like entropy, volatility, or fractal dimensions. Synthesising generates entirely new data points. The order matters, and following it is meant to maximize how much your model improves. Someone working on a stock price prediction model might use this to turn a basic price-and-volume dataset into hundreds of engineered features without manually writing each transformation. A quant researcher could apply it to financial time series, but anyone with tabular data, whether that's sales figures, sensor readings, or customer records, could benefit. The README mentions an example where combining multiple methods halved the prediction error. The project is delivered as a Python package you install with pip, but most of the documentation and examples live in Jupyter notebooks, including a hosted Colab version you can run in a browser without setting anything up locally. One thing to note: the README is upfront that it doesn't fully handle the common pitfall of "peeking ahead" in time-series data, where information from the future accidentally leaks into training. Some functions would need restructuring to avoid that, so users working with time-sensitive data should be cautious.
DeltaPy is a Python toolkit that automatically generates new data columns and rows from your existing dataset, giving machine learning models more signal to work with. It runs a five-step pipeline (transform, interact, map, extract, synthesize) to create engineered features without manual effort.
Mainly Jupyter Notebook. The stack also includes Python, Jupyter Notebook, Google Colab.
Dormant — no commits in 2+ years (last push 2023-09-19).
The explanation does not mention a specific license, so the permissions for using this code are unclear.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly data.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.