Skip to main content

Invariant Gradient Boosted Decision Tree Package - Era Splitting.

Project description

This is the official code base for Era Splitting. Using this repository you can install and run the EraHistGradientBoostingRegressor with the new era splitting, directional era splitting, or original criterion implemented via simple arguments.

Era Splitting Paper: https://arxiv.org/abs/2309.14496

Installation via Pip

pip install erasplit

Example Implementation w/ Numerai Data

from pathlib import Path
from numerapi import NumerAPI #pip install numerapi
import json

"""Era Split Model"""
from erasplit.ensemble import EraHistGradientBoostingRegressor

napi = NumerAPI()
Path("./v4").mkdir(parents=False, exist_ok=True)
napi.download_dataset("v4/train.parquet")
napi.download_dataset("v4/features.json")

with open("v4/features.json", "r") as f:
    feature_metadata = json.load(f)
features = feature_metadata["feature_sets"]['small']
TARGET_COL="target_cyrus_v4_20"

training_data = pd.read_parquet('v4/train.parquet')
training_data['era'] = training_data['era'].astype('int')

model = EraHistGradientBoostingRegressor( 
    early_stopping=False, 
    boltzmann_alpha=0, 
    max_iter=5000, 
    max_depth=5, 
    learning_rate=.01, 
    colsample_bytree=.1, 
    max_leaf_nodes=32, 
    gamma=1, #for era splitting
    #blama=1,  #for directional era splitting
    #vanna=1,  #for original splitting criterion
)
model.fit(training_data[ features ], training_data[ TARGET_COL ], training_data['era'].values)

Explanation of Parameters

Boltzmann Alpha

The Boltzmann alpha parameter varies from -infinity to +infinity. A value of zero recovers the mean, -infinity recovers the minumum and +infinity recovers the maximum. This smooth min/max function is applied to the era-wise impurity scores when evaluating a data split. Negative values here will build more invariant trees.

Read more: https://en.wikipedia.org/wiki/Smooth_maximum

Gamma

Varies over the interval [0,1]. Indicates weight placed on the era splitting criterion.

Blama

Varies over the interval [0,1]. Indicates weight placed on the directional era splitting criterion.

Vanna

Varies over the interval [0,1]. Indicates weight placed on the original splitting criterion.

Behind the scenes, this is for formula which creates a linear combination of the split criteria. Usually we just set one of these to 1 and leave the other at zero.

gain = gamma * era_split_gain + blama * directional_era_split_gain + vanna * original_gain

Complete (New Updated) Code Notebook Examples Available here:

https://github.com/jefferythewind/era-splitting-notebook-examples

Citations:

@misc{delise2023era,
      title={Era Splitting}, 
      author={Timothy DeLise},
      year={2023},
      eprint={2309.14496},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

This code was forked from the official scikit-learn repository and is currently a stand-alone version. All community help is welcome for getting these ideas part of the official scikit learn code base or even better, incorporated in the LightGBM code base.

https://scikit-learn.org/stable/about.html#citing-scikit-learn

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

erasplit-1.0.5.tar.gz (2.4 MB view details)

Uploaded Source

File details

Details for the file erasplit-1.0.5.tar.gz.

File metadata

  • Download URL: erasplit-1.0.5.tar.gz
  • Upload date:
  • Size: 2.4 MB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/5.1.1 CPython/3.10.14

File hashes

Hashes for erasplit-1.0.5.tar.gz
Algorithm Hash digest
SHA256 234cecde09c30a20e992d183afaf321944fc830ec8e4afff53618cd4ddd275d3
MD5 f7cc28e3da6191c8aaaafad0cefd331d
BLAKE2b-256 691a51e36d8ad21ce99ba8641999cb05ca2eea14e4c119fde2db1a2e5c8828cd

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page