本文共 2880 字，大约阅读时间需要 9 分钟。
Because we can not use usual bootstrap sampling method to get subsamples from time series dataset, then the MBB was proposed to address this issue.
Suppose we have the time series following : X 1 X_1 X1, X 2 X_2 X2, . . . . .... ...., X 10 X_{10} X10. Note that you can extend the footnote with other number or a symbol. First step is to split the series into several blocks in which one should figure out the size of each block. Assume we let the size of each block equal to 2, then our blocked data would look like :X 1 , X 2 ⏞ b l o c k 1 , X 3 , X 4 ⏞ b l o c k 2 , . . . . . . X 9 , X 10 ⏞ b l o c k 5 \overbrace{X_1,X_2}^{block\,1},\overbrace{X_3,X_4}^{block\,2},......\overbrace{X_9,X_{10}}^{block\,5} X1,X2 block1,X3,X4 block2,......X9,X10 block5
with the above blocks we get, now we can apply the bootstrap algorithm by taking a random sample of the blocks with replacement. The order in which the blocks are drawn is the position that they are placed in the bootstrap series. Hence, one probably blocks might be
b l o c k 1 , b l o c k 3 , b l o c k 5 , b l o c k 1 , b l o c k 2 block\,1, block3\,, block \,5, block\,1,block\,2 block1,block3,block5,block1,block2 and the corresponding original time series is X 1 , X 2 , X 5 , X 6 X 9 , X 10 X 1 , X 2 , X 3 , X 4 , X_1,X_2,\,\,\,\,\,X_5,X_6\,\,\,\,\,X_9,X_{10}\,\,\,\,\,X_1,X_2,\,\,\,\,\,X_3,X_4,\,\,\,\,\, X1,X2,X5,X6X9,X10X1,X2,X3,X4,These are the basic process of MBB for time series data. This can help us to get a new sample series with similar short term dependence data structure to the original data.
In python, you can rely on the pkg of “arch”, following I present a simple toy code from the with a minor revision.
from arch.bootstrap import MovingBlockBootstrap from numpy.random import RandomState from numpy.random import standard_normalimport numpy as npy = standard_normal((6, 1))# to generate a time series with standard norm distribution.bs = MovingBlockBootstrap(2, y,random_state=RandomState(1234))# 2 is block size, y is your time series data, random_state # is for reproducibility when requiredi=0bs_x=[ ] # an empty list to store the bootstrap series#here for is to look what the bootstrap looks like for each iteration, in our demonstrated case, we do boostrap only for 2 times.for data in bs.bootstrap(2): print(data) bs_y.append(data[0][0]) print(bs_y) i=i+1# fc is function to compute the bootstrap series mean value.，you can replace it with your own definitiondef fc(a): return a.mean(0)results = bs.apply(fc,2) ###to apply a function defined by yourself to the bootstrap replicated data, "2" means apply the fc 2times to calculate the average of the boostrap data. of course , you can figure it with any number you want. print(results)
Thx for ur reading.
转载地址：http://uaet.baihongyu.com/