mirror of
https://github.com/webclinic017/drift.git
synced 2026-08-01 05:07:47 +00:00
Mark Aron Szulyovszky
b1dcdfc09d
* chore(Documentation): updated README * chore(Docs): updated README
41 lines
3.0 KiB
Markdown
41 lines
3.0 KiB
Markdown
# Financial time series prediction
|
|
|
|
And end-to-end pipeline to train predictive Machine Learning models on financial (non-stationary, regime changing) time series. Includes feature selection and meta labelling.
|
|
|
|
## Why?
|
|
|
|
Machine learning on financial time series requires a fundamentally different approach compared to standard ML domains. The (small amount of) data is non-stationary, where the patterns frequently change, and it's extremely important to not to leak out-of-sample data into the test set.
|
|
|
|
There are very few open-source end-to-end machine learning pipelines that can be effectively used to train and evaluate ML models on financial time series. Among them are: [qlib](https://github.com/microsoft/qlib), [AlphaPy](https://github.com/ScottfreeLLC/AlphaPy).
|
|
|
|
This repo is different to them in a couple of angles:
|
|
- Feature extraction and selection is an important, pre-built step in the pipeline. Training models on irrelevant data will worsen the performance.
|
|
- Training and evaluation is done in a [walk-forward manner](https://en.wikipedia.org/wiki/Walk_forward_optimization). We argue that that one or two train/test split is not adoquate to evaluate an ML model's performance in a non-stationary, regime changing environment. The walk-forward methodology enables us to evaluate the model's performance on almost the whole time series. Combinatorial purged k-fold cross-validation can still have a place *within* a walk-forward training method - you can shuffle the past in any way you prefer, but never use future data to train the model, if you want objective evaluation of its performance.
|
|
- The walk-forward training/evaluation methodology enables "online" (ever-changing) models, that adapt to the market environment. You can specify how frequently would you like to re-train the models.
|
|
- By default this framework always train multiple models, and average their predictions. Improves performance and adds a lot of robustness that's much needed in this environment.
|
|
- Meta-labeling (training a model to evaluate a lower level model's prediction for each timestamp) is a built-in feature - it improves performance a lot, and there are no open source implementation available as far as we're aware.
|
|
|
|
This project is inspired partially by [Marcos Lopez de Prado's Advances in Financial Machine Learning](https://www.wiley.com/en-us/Advances+in+Financial+Machine+Learning-p-9781119482086) and [The Alpha Scientist's blogposts](https://alphascientist.com/).
|
|
|
|
|
|
## Installation
|
|
|
|
Use the conda environment file attached!:)
|
|
|
|
|
|
## Pipeline components
|
|
|
|
- Feature extraction
|
|
- Dimensionality reduction
|
|
- Feature selection
|
|
- Training Level-1 models
|
|
- Training Level-2 (Meta-label) model
|
|
- Evaluation of models
|
|
- Cross-sectional portfolio construction [IN PROGRESS]
|
|
|
|
## Weaknesses
|
|
|
|
Now that we assured that you should avoid lookahead bias at all costs, for practical reasons (to make this problem tractable), these are the things that are not completely lookahead-bias free:
|
|
- Dimensionality reduction
|
|
- Feature selection
|