Overview
Open Source
Tests
Code
Downloads
Citation
Support

Table of Contents

1. About The Project
   • Introduction
   • Motivation
   • Built With
2. Installation
3. Getting Started
   • Scikit-Learn Integration
   • Expanding Number Split
   • Expanding Split
   • Rolling Number Split
   • Rolling Split
   • Rolling Optimized Split
   • Interval Split
   • Calendar Split
   • Period Split
   • Grouper Split
   • Custom Function Split
   • Random Split
   • Purged KFold
   • Purged Walk-Forward
   • Train Test Split
4. Contributing
5. License
6. Support

About The Project

Introduction

fold is a powerful and flexible time-series cross-validation library designed to work seamlessly with scikit-learn. Whether you're working on financial forecasting, weather prediction, or any other domain that involves time-series data, fold offers a suite of advanced cross-validation techniques that go beyond traditional methods. By integrating fold into your workflow, you can ensure that your models are robust, reliable, and free from common pitfalls like look-ahead bias and data leakage.

Key Features:

• Versatile Splits: Support for expanding, rolling, interval, and custom function splits.
• Scikit-learn Compatibility: Easily integrate with scikit-learn’s model selection framework.
• Advanced Cross-Validation Techniques: Includes purged k-fold, purged walk-forward, and temporal safe random splits.

Motivation

Traditional cross-validation methods often fall short when applied to time-series data due to the inherent temporal dependencies. Naive random splits can introduce look-ahead bias, where future data points inadvertently influence past predictions, leading to overly optimistic model performance. This is where fold steps in, providing a robust framework to handle the unique challenges of time-series data.

Why We Built Fold?

• Accuracy and Reliability: Ensuring that models are tested in a way that closely mimics real-world scenarios is crucial for accurate predictions. fold's advanced splitting techniques prevent data leakage and provide a more realistic evaluation of model performance.
• Flexibility: Different time-series problems require different cross-validation strategies. fold offers a variety of splitting methods, from simple expanding and rolling splits to more complex strategies like purged k-fold and custom function splits, allowing users to tailor their validation process to their specific needs.
• Ease of Use: Designed to be fully compatible with scikit-learn, fold integrates seamlessly into existing workflows, making it easy for users to switch to more advanced time-series cross-validation methods without a steep learning curve.

(back to top)

Built With

• numpy = "^1.26.4"
• pandas = "^2.2.2"
• scikit-learn = "^1.4.2"
• numba = "^0.59.1"
• dateparser = "^1.2.0"

(back to top)

Installation

$ pip install python-fold

(back to top)

Getting Started

To demonstrate the use of fold, we created numpy arrays X and y, along with a pandas.DatetimeIndex object with a fixed frequency.

>>> import pandas as pd
>>> import numpy as np

>>> X = np.array([[1, 2], [3, 4], [5, 6], [7, 8]])
>>> y = np.array([1, 2, 3, 4])

>>> index = pd.date_range("2010", "2024", freq="D")

Scikit-Learn Integration

Example Model

>>> from fold import WrapperSplit, ExpandingSplit
>>> cv = WrapperSplit(
...     model=ExpandingSplit,
...     min_length=2,
...     offset=1,
...     split=-1,
...     sample_labels=["IS", "OOS"]
... )

Computing Cross-Validated Metrics

>>> # Pipeline and cross_val_score
>>> from sklearn.pipeline import make_pipeline
>>> from sklearn.model_selection import cross_val_score
>>> from sklearn.preprocessing import StandardScaler
>>> from sklearn.linear_model import LinearRegression
>>> pipe = make_pipeline(StandardScaler(),  LinearRegression())
>>> cross_val_score(pipe, X, y, cv=cv)
# array([ ...

Cross Validation Iterators

>>> model = cv.get_model(X)
>>> print(model.get_bounds(index_bounds=False))
# bound         start  end
# split sample            
# 0     IS          0    1
#       OOS         1    2
# 1     IS          0    2
#       OOS         2    3
# 2     IS          0    3
#       OOS         3    4

>>> for i, (train, test) in enumerate(cv.split(X)):
...     print("Split %d:" % i)
...     X_train, X_test = X[train], X[test]
...     print("  X:", X_train.tolist(), X_test.tolist())
...     y_train, y_test = y[train], y[test]
...     print("  y:", y_train.tolist(), y_test.tolist())
# Split 0:
#   X: [[1, 2]] [[3, 4]]
#   y: [1] [2]
# Split 1:
#   X: [[1, 2], [3, 4]] [[5, 6]]
#   y: [1, 2] [3]
# Split 2:
#   X: [[1, 2], [3, 4], [5, 6]] [[7, 8]]
#   y: [1, 2, 3] [4]

Scikit-Learn models

>>> from sklearn.model_selection import TimeSeriesSplit
>>> from fold import SklearnFold

>>> model = SklearnFold(
...     index,
...     sk_model=TimeSeriesSplit(n_splits=2),
...     sample_labels=["IS", "OOS"]
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2014-09-03
#       OOS    2014-09-03 2019-05-04
# 1     IS     2010-01-01 2019-05-04
#       OOS    2019-05-04 2024-01-02

(back to top)

Bespoke Models

Expanding Number Split

>>> from fold import ExpandingNumberSplit
>>> model = ExpandingNumberSplit(
...     index,
...     n=5,
...     min_length=360,
...     split=-180,
...     sample_labels=["IS", "OOS"],
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2010-06-30
#       OOS    2010-06-30 2010-12-27
# 1     IS     2010-01-01 2013-09-30
#       OOS    2013-09-30 2014-03-29
# 2     IS     2010-01-01 2017-01-01
#       OOS    2017-01-01 2017-06-30
# 3     IS     2010-01-01 2020-04-04
#       OOS    2020-04-04 2020-10-01
# 4     IS     2010-01-01 2023-07-06
#       OOS    2023-07-06 2024-01-02

(back to top)

Expanding Split

>>> from fold import ExpandingSplit
>>> model = ExpandingSplit(
...     index,
...     min_length=100,
...     offset=50,
...     split=-20,
...     sample_labels=["IS", "OOS"]
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2010-03-22
#       OOS    2010-03-22 2010-04-11
# 1     IS     2010-01-01 2010-05-11
#       OOS    2010-05-11 2010-05-31
# 2     IS     2010-01-01 2010-06-30
#                 ...        ...
# 98    OOS    2023-08-21 2023-09-10
# 99    IS     2010-01-01 2023-10-10
#       OOS    2023-10-10 2023-10-30
# 100   IS     2010-01-01 2023-11-29
#       OOS    2023-11-29 2023-12-19

# [202 rows x 2 columns]

(back to top)

Rolling Number Split

>>> from fold import RollingNumberSplit
>>> model = RollingNumberSplit(
...     index,
...     n=7,
...     length=360,
...     split=0.5,
...     sample_labels=["IS", "OOS"]
... )
>>> model.get_bounds(index_bounds=True)
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2010-06-30
#       OOS    2010-06-30 2010-12-27
# 1     IS     2012-03-03 2012-08-30
#       OOS    2012-08-30 2013-02-26
# 2     IS     2014-05-05 2014-11-01
#       OOS    2014-11-01 2015-04-30
# 3     IS     2016-07-05 2017-01-01
#       OOS    2017-01-01 2017-06-30
# 4     IS     2018-09-05 2019-03-04
#       OOS    2019-03-04 2019-08-31
# 5     IS     2020-11-06 2021-05-05
#       OOS    2021-05-05 2021-11-01
# 6     IS     2023-01-07 2023-07-06
#       OOS    2023-07-06 2024-01-02

(back to top)

Rolling Split

>>> from fold import RollingSplit
>>> model = RollingSplit(
...     index,
...     60,
...     split=1/2,
...     sample_labels=["IS", "OOS"]
... )
>>> model.get_bounds(index_bounds=True)
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2010-01-31
#       OOS    2010-01-31 2010-03-02
# 1     IS     2010-01-31 2010-03-02
#       OOS    2010-03-02 2010-04-01
# 2     IS     2010-03-02 2010-04-01
#                 ...        ...
# 166   OOS    2023-09-20 2023-10-20
# 167   IS     2023-09-20 2023-10-20
#       OOS    2023-10-20 2023-11-19
# 168   IS     2023-10-20 2023-11-19
#       OOS    2023-11-19 2023-12-19

# [338 rows x 2 columns]

(back to top)

Rolling Optimized Split

>>> from fold import RollingOptimizeSplit
>>> model = RollingOptimizeSplit(
...     index,
...     n=7,
...     split=0.5,
...     sample_labels=["IS", "OOS"]
... )
>>> model.get_bounds(index_bounds=True)
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2011-10-02
#       OOS    2011-10-02 2013-07-02
# 1     IS     2011-10-02 2013-07-02
#       OOS    2013-07-02 2015-04-02
# 2     IS     2013-07-02 2015-04-02
#       OOS    2015-04-02 2016-12-31
# 3     IS     2015-04-02 2016-12-31
#       OOS    2016-12-31 2018-10-01
# 4     IS     2016-12-31 2018-10-01
#       OOS    2018-10-01 2020-07-01
# 5     IS     2018-10-01 2020-07-01
#       OOS    2020-07-01 2022-04-01
# 6     IS     2020-07-01 2022-04-01
#       OOS    2022-04-01 2023-12-31

(back to top)

Interval Split

>>> from fold import IntervalSplit
>>> model = IntervalSplit(
...     index,
...     every="YS",
...     closed_end=True,
...     split=0.5,
...     sample_labels=["IS", "OOS"]
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2010-07-03
#       OOS    2010-07-03 2011-01-02
# 1     IS     2011-01-01 2011-07-03
#       OOS    2011-07-03 2012-01-02
# 2     IS     2012-01-01 2012-07-02
#       OOS    2012-07-02 2013-01-02
# 3     IS     2013-01-01 2013-07-03
#       OOS    2013-07-03 2014-01-02
# 4     IS     2014-01-01 2014-07-03
#       OOS    2014-07-03 2015-01-02
# 5     IS     2015-01-01 2015-07-03
#       OOS    2015-07-03 2016-01-02
# 6     IS     2016-01-01 2016-07-02
#       OOS    2016-07-02 2017-01-02
# 7     IS     2017-01-01 2017-07-03
#       OOS    2017-07-03 2018-01-02
# 8     IS     2018-01-01 2018-07-03
#       OOS    2018-07-03 2019-01-02
# 9     IS     2019-01-01 2019-07-03
#       OOS    2019-07-03 2020-01-02
# 10    IS     2020-01-01 2020-07-02
#       OOS    2020-07-02 2021-01-02
# 11    IS     2021-01-01 2021-07-03
#       OOS    2021-07-03 2022-01-02
# 12    IS     2022-01-01 2022-07-03
#       OOS    2022-07-03 2023-01-02
# 13    IS     2023-01-01 2023-07-03
#       OOS    2023-07-03 2024-01-02

(back to top)

Calendar Split

>>> from fold import CalendarSplit
>>> model = CalendarSplit(
...     index,
...     n_train=2,
...     n_test=1,
...     every='YS',
...     sample_labels=["IS", "OOS"]
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2010-07-01
#       OOS    2010-04-01 2010-07-01
# 1     IS     2010-04-01 2010-10-01
#       OOS    2010-07-01 2010-10-01
# 2     IS     2010-07-01 2011-01-01
#                 ...        ...
# 52    OOS    2023-04-01 2023-07-01
# 53    IS     2023-04-01 2023-10-01
#       OOS    2023-07-01 2023-10-01
# 54    IS     2023-07-01 2024-01-01
#       OOS    2023-10-01 2024-01-01

# [110 rows x 2 columns]

(back to top)

Period Split

>>> from fold import PeriodSplit
>>> model = PeriodSplit(
...     index,
...     n_train=(10, 'Y'),
...     n_test=(2, 'Q'),
...     sample_labels=["IS", "OOS"]
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     IS     2010-07-01 2020-07-01
#       OOS    2020-07-01 2021-01-01
# 1     IS     2011-01-01 2021-01-01
#       OOS    2021-01-01 2021-07-01
# 2     IS     2011-07-01 2021-07-01
#       OOS    2021-07-01 2022-01-01
# 3     IS     2012-01-01 2022-01-01
#       OOS    2022-01-01 2022-07-01
# 4     IS     2012-07-01 2022-07-01
#       OOS    2022-07-01 2023-01-01
# 5     IS     2013-01-01 2023-01-01
#       OOS    2023-01-01 2023-07-01
# 6     IS     2013-07-01 2023-07-01
#       OOS    2023-07-01 2024-01-01

(back to top)

Grouper Split

>>> from fold import GrouperSplit
>>> model = GrouperSplit(
...     index,
...     split=0.5,  
...     sample_labels=['IS', 'OOS']
... )
>>> print(model.get_bounds(index_bounds=True))
# bound            start        end
#      sample                      
# 2010 IS     2010-10-13 2010-11-22
#      OOS    2010-11-22 2011-01-01
# 2011 IS     2011-01-01 2011-07-02
#      OOS    2011-07-02 2012-01-01
# 2012 IS     2012-01-01 2012-07-02
#      OOS    2012-07-02 2013-01-01
# 2013 IS     2013-01-01 2013-07-02
#      OOS    2013-07-02 2014-01-01
# 2014 IS     2014-01-01 2014-07-02
#      OOS    2014-07-02 2015-01-01
# 2015 IS     2015-01-01 2015-07-02
#      OOS    2015-07-02 2016-01-01
# 2016 IS     2016-01-01 2016-07-02
#      OOS    2016-07-02 2017-01-01
# 2017 IS     2017-01-01 2017-07-02
#      OOS    2017-07-02 2018-01-01
# 2018 IS     2018-01-01 2018-07-02
#      OOS    2018-07-02 2019-01-01
# 2019 IS     2019-01-01 2019-07-02
#      OOS    2019-07-02 2020-01-01
# 2020 IS     2020-01-01 2020-07-02
#      OOS    2020-07-02 2021-01-01
# 2021 IS     2021-01-01 2021-07-02
#      OOS    2021-07-02 2022-01-01
# 2022 IS     2022-01-01 2022-07-02
#      OOS    2022-07-02 2023-01-01
# 2023 IS     2023-01-01 2023-07-02
#      OOS    2023-07-02 2024-01-01
# 2024 IS     2024-01-01 2024-03-27

(back to top)

Custom Function Split

>>> from fold import FunctionSplit, Key
>>> def func(index, prev_start):
...     if prev_start is None:
...         prev_start = index[0]
...     new_start = prev_start + pd.offsets.MonthBegin(1)
...     new_end = new_start + pd.DateOffset(years=1)
...     if new_end > index[-1] + index.freq:
...         return None
...     return [
...         slice(new_start, new_start + pd.offsets.MonthBegin(9)),
...         slice(new_start + pd.offsets.MonthBegin(9), new_end)
...     ]
>>> 
>>> model = FunctionSplit(
...     index,
...     split_func=func,
...     split_args=(Key("index"), Key("prev_start")),
...     index_bounds=True,
...     sample_labels=["IS", "OOS"],
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     IS     2010-02-01 2010-11-01
#       OOS    2010-11-01 2011-02-01
# 1     IS     2010-03-01 2010-12-01
#       OOS    2010-12-01 2011-03-01
# 2     IS     2010-04-01 2011-01-01
#                 ...        ...
# 153   OOS    2023-08-01 2023-11-01
# 154   IS     2022-12-01 2023-09-01
#       OOS    2023-09-01 2023-12-01
# 155   IS     2023-01-01 2023-10-01
#       OOS    2023-10-01 2024-01-01

# [312 rows x 2 columns]

(back to top)

Random Split

>>> from fold import RandomNumberSplit
>>> model = RandomNumberSplit(
...     index,
...     50,
...     min_length=360,
...     split=0.5,
...     sample_labels=["train", "test"]
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     train  2017-08-31 2018-02-27
#       test   2018-02-27 2018-08-26
# 1     train  2020-02-07 2020-08-05
#       test   2020-08-05 2021-02-01
# 2     train  2013-06-25 2013-12-22
#                 ...        ...
# 47    test   2017-10-20 2018-04-18
# 48    train  2017-10-20 2018-04-18
#       test   2018-04-18 2018-10-15
# 49    train  2014-08-29 2015-02-25
#       test   2015-02-25 2015-08-24

# [100 rows x 2 columns]

(back to top)

Purged KFold

>>> from fold import PurgedKFold
>>> model = PurgedKFold(
...     index,
...     n_folds=10,
...     n_test_folds=2,
...     sample_labels=["IS", "OOS"]
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2021-03-16
#       OOS    2021-03-16 2024-01-02
# 1     IS     2010-01-01 2022-08-09
#       OOS    2019-10-22 2024-01-02
# 2     IS     2010-01-01 2024-01-02
#                 ...        ...
# 42    OOS    2010-01-01 2015-08-11
# 43    IS     2011-05-28 2024-01-02
#       OOS    2010-01-01 2014-03-17
# 44    IS     2012-10-21 2024-01-02
#       OOS    2010-01-01 2012-10-21

# [90 rows x 2 columns]

(back to top)

Purged Walk-Forward

>>> from fold import PurgedWalkForwardSplit
>>> model = PurgedWalkForwardSplit(
...     index,
...     n_folds=10,
...     n_test_folds=1,
...     min_train_folds=2,
...     max_train_folds=None,
...     sample_labels=["IS", "OOS"]
... )
>>> print(model.get_bounds(index_bounds=True))
# bound             start        end
# split sample                      
# 0     IS     2010-01-01 2012-10-21
#       OOS    2012-10-21 2014-03-17
# 1     IS     2010-01-01 2014-03-17
#       OOS    2014-03-17 2015-08-11
# 2     IS     2010-01-01 2015-08-11
#       OOS    2015-08-11 2017-01-03
# 3     IS     2010-01-01 2017-01-03
#       OOS    2017-01-03 2018-05-29
# 4     IS     2010-01-01 2018-05-29
#       OOS    2018-05-29 2019-10-22
# 5     IS     2010-01-01 2019-10-22
#       OOS    2019-10-22 2021-03-16
# 6     IS     2010-01-01 2021-03-16
#       OOS    2021-03-16 2022-08-09
# 7     IS     2010-01-01 2022-08-09
#       OOS    2022-08-09 2024-01-02

(back to top)

Train Test Split

>>> from fold import train_test_split

Create a 2-d dataset

>>> data = pd.DataFrame(
...     np.random.normal(size=(len(index), 10)),
...     index=index
... )

Select model

>>> model = PurgedWalkForwardSplit(index, sample_labels=["IS", "OOS"])

Get a pandas DataFrame object

>>> splitter = train_test_split(data, model)

Last valid split number

>>> last_split = splitter.last_valid_index()[0]

Train and test-set

>>> X_train = splitter[last_split, "IS"]
>>> y_test = splitter[last_split, 'OOS']

Apply transform

>>> from sklearn.preprocessing import StandardScaler
>>> transformer = StandardScaler().set_output(transform="pandas")
>>> transformer.fit(X_train).transform(y_test)
#                   x0        x1        x2  ...        x7        x8        x9
# 2022-08-09 -0.576242 -0.225575  0.042182  ...  0.520885 -0.610266 -0.194724
# 2022-08-10  0.802063  0.434773  1.828127  ... -1.606077 -1.558970 -0.376618
# 2022-08-11  0.391401  0.134264  2.245485  ...  0.896719  0.684971 -0.846399
# 2022-08-12  1.504788 -0.540543 -1.633605  ... -2.186249  1.044936 -0.853112
# 2022-08-13 -0.749084  1.271873 -1.019302  ... -0.577152 -0.061101  0.526385
#              ...       ...       ...  ...       ...       ...       ...
# 2023-12-28 -1.143657 -0.639147 -2.203081  ... -0.028623 -0.756192 -0.558058
# 2023-12-29 -0.183911  0.440218  0.670539  ...  1.014517 -0.620626  0.175182
# 2023-12-30 -0.416487 -0.570313  0.215304  ... -1.022257  0.240412  1.375879
# 2023-12-31  1.897310  0.405948 -0.346925  ... -1.120385 -0.382428  0.381911
# 2024-01-01  1.369621  1.100124  1.282244  ...  1.018545  1.021911  1.326567

# [511 rows x 10 columns]

(back to top)

Contributing

Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated.

If you have a suggestion that would make this better, please fork the repo and create a pull request. You can also simply open an issue with the tag "enhancement". Don't forget to give the project a star! Thanks again!

1. Fork the Project
2. Create your Feature Branch (git checkout -b feature/AmazingFeature)
3. Commit your Changes (git commit -m 'Add some AmazingFeature')
4. Push to the Branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

(back to top)

Support

If you found fold useful, you can support us with a donation. Your contribution will help to continue developing and improving this project. Many thanks!

License

Distributed under the BSD-3 License. See LICENSE.txt for more information.

(back to top)