Implement Rep-Holdout

[georgeblck]

Thank you for this repository and the implemented CV-methods; especially GapRollForward. I was looking for exactly this package.

I was wondering if you are interested in implementing another CV-Method for time series, called Rep-Holdout. It is used in this evaluation paper (https://arxiv.org/abs/1905.11744) and has good performance compared to all other CV-methods - some of which you have implemented here.

As I understand it, it is somewhat like sklearn.model_selection.TimeSeriesSplit but with a randomized selection of all possible folds. Here is the description from the paper as an image:

---

The authors provided code in R but it is written very differently than how it needs to look in Python. I adapted your functions to implement it in python but I am not the best coder and it really only serves my purpose of tuning a specific model. Seeing as the performance of Rep-Holdout is good and -to me at least - it makes sense for time series cross validation, maybe you are interested in adding this function to your package?

[WenjieZ]

The so-called Rep-Holdout seems lame to me. Use the following code instead:

n = LENGTH_OF_DATA
m = NUMBER_OF_FOLDS
window = (a, b)
cv = GapRollForward(min_train_size=a, min_test_size=n-b, roll_size=(b-a)//(m-1))

[WenjieZ]

Oops! The last line should have been:

cv = GapRollForward(min_train_size=a, 
                    min_test_size=n-b, 
                    max_test_size=np.inf, 
                    roll_size=(b-a)//(m-1))

[georgeblck]

Thank you for your fast reply. My problem with the GapRollForward is that - without adjusting/tuning the various input arguments - it lead to overfitting when finetuning my time series forecasting task. E.g. I have 2.5 years and choose the first 2 years as the min_train_size and then do rolling cross validation from that point onwards. Although this rolling prediction closely resembles the actual prediction procedure, when I use it as part of hyperparameter tuning for the underlying method it naturally overfits to the data outside of the min_train_size.

That is why I liked the RepHoldout; because if used as part of a tuning procedure it removes the bias induced by min_train_size and results in a more even crossvalidation performance.

I'm not sure if I understand your suggestion correctly but I compared a small example here with the Rep-Holdout:

import numpy as np
from tscv import GapRollForward

# Number of Samples
n = 30
# Number of Folds
m = 5
# Windowsize
window = (1, 5)

cv = GapRollForward(min_train_size=window[0], 
                    min_test_size=n-window[1], 
                    max_test_size=np.inf, 
                    roll_size=(window[1]-window[0])//(m-1))
for train, test in cv.split(range(n)):
    print("train:", train, "test:", test)

train: [0] test: [ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29] train: [0 1] test: [ 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29] train: [0 1 2] test: [ 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29] train: [0 1 2 3] test: [ 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29] train: [0 1 2 3 4] test: [ 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29]

and here is the output from my basic RepHoldout implementation:

cv =RepHoldout(nreps=5, train_size=4,
                 test_size=1, gap_size=0)

for train, test in cv.split(range(30)):
    print("train:", train, "test:", test)

train: [23 24 25 26] test: [27] train: [3 4 5 6] test: [7] train: [5 6 7 8] test: [9] train: [6 7 8 9] test: [10] train: [14 15 16 17] test: [18]

[WenjieZ]

That is what the max_train_size parameter is for.

# Number of Samples
n = 30
# Number of Folds
m = 5
# Windowsize
window = (5, 25)

cv = GapRollForward(min_train_size=window[0], 
                    min_test_size=n-window[1], 
                    max_train_size=4, 
                    roll_size=(window[1]-window[0])//(m-1))
for train, test in cv.split(range(n)):
    print("train:", train, "test:", test)

train: [1 2 3 4] test: [5] train: [6 7 8 9] test: [10] train: [11 12 13 14] test: [15] train: [16 17 18 19] test: [20] train: [21 22 23 24] test: [25]

[WenjieZ]

A few comments:

1. The variable window refers to the "available window" in your Rep-Holdout.
2. It's a good idea to use balanced training and test set across all folds of cross-validation. You are doing it right.

[georgeblck]

Thank you for your comments. I like your code and see that it is very similar to the Rep-Holdout; just not randomized and instead truly rolling forward.

[WenjieZ]

Yeah, randomization is generally to be avoided if they are not an essential part of an algorithm. In particular, in your example, Rep-Holdout can be seen as simple sampling with replacement, and my code systematic sampling, which often results lower variance and is thus preferred.

[georgeblck]

Thank you for the insights! What are your thoughts on the paper I linked to then and the performance of the different methods?

[fengchi863]

thanks for your link, I also read this paper yesterday. I have some questions. I think the rep-holdout is not like your demo. It gives an available window to split, so I think the train or test is not a fixed length.

it may be this:

cv.split(range(10)): train:[1,2,3,4] test:[5,6,7,8,9,10] train:[1,2,3,4,5] test:[6,7,8,9,10] train:[1,2,3] test:[4,5,6,7,8,9,10] train:[1,2,3,4,5,6,7] test:[8,9,10]

[chasse20]

thanks for your link, I also read this paper yesterday. I have some questions. I think the rep-holdout is not like your demo. It gives an available window to split, so I think the train or test is not a fixed length.

it may be this:

cv.split(range(10)): train:[1,2,3,4] test:[5,6,7,8,9,10] train:[1,2,3,4,5] test:[6,7,8,9,10] train:[1,2,3] test:[4,5,6,7,8,9,10] train:[1,2,3,4,5,6,7] test:[8,9,10]

The method described in that paper is very ambiguous, but what you have seems most likely what the author meant. The last one WenjieZ posted was just a rolling block method. I do agree with him though on forgoing randomness and would rather just exhaust each fold as you have shown starting from [1] to [1,2,3,4,5,6,7,8,9] training versus test.

[WenjieZ]

thanks for your link, I also read this paper yesterday. I have some questions. I think the rep-holdout is not like your demo. It gives an available window to split, so I think the train or test is not a fixed length.

it may be this:

cv.split(range(10)):

train:[1,2,3,4] test:[5,6,7,8,9,10]

train:[1,2,3,4,5] test:[6,7,8,9,10]

train:[1,2,3] test:[4,5,6,7,8,9,10]

train:[1,2,3,4,5,6,7] test:[8,9,10]

Increase the max_test_size parameter (default to 1) if this split is what you are looking for.