BUG: pandas rolling_quantile does not use interpolation

[leo4183]

i recently bumped an unexpected issue with pandas rolling funcs. rolling_quantile for example:

>> row = 10
>> col = 5
>> idx = pd.date_range(20100101,periods=row,freq='B')
>> a = pd.DataFrame(np.random.rand(row*col).reshape((row,-1)),index=idx)
>> a
                   0           1           2           3           4
2010-01-01  0.341434    0.497274    0.596341    0.259909    0.872207
2010-01-04  0.222653    0.056723    0.064019    0.936307    0.785647
2010-01-05  0.179067    0.647165    0.931266    0.557698    0.713282
2010-01-06  0.049766    0.259756    0.945736    0.380948    0.282667
2010-01-07  0.385036    0.517609    0.575958    0.050758    0.850735
2010-01-08  0.628169    0.510453    0.325973    0.263361    0.444959
2010-01-11  0.099133    0.976571    0.602235    0.181185    0.506316
2010-01-12  0.987344    0.902289    0.080000    0.254695    0.753325
2010-01-13  0.759198    0.014548    0.139858    0.822900    0.251972
2010-01-14  0.404149    0.349788    0.038714    0.280568    0.197865

>> a.quantile([0.25,0.5,0.75],axis=0)
               0           1           2           3           4
0.25    0.189963    0.282264    0.094964    0.255999    0.323240
0.50    0.363235    0.503864    0.450966    0.271964    0.609799
0.75    0.572164    0.614776    0.600761    0.513510    0.777567

>> np.percentile(a,[25,50,75],axis=0)
[array([ 0.18996316,  0.28226404,  0.09496441,  0.25599853,  0.32323997]),
 array([ 0.36323529,  0.50386356,  0.45096554,  0.27196429,  0.60979881]),
 array([ 0.57216415,  0.61477607,  0.6007611 ,  0.51351021,  0.7775667 ])]

>> pd.rolling_quantile(a,row,0.25).tail(1)
                   0           1       2           3           4
2010-01-14  0.179067    0.259756    0.08    0.254695    0.282667

looks like pandas.DataFrame.quantile member func is consistent with the numpy.percentile func. however the pandas.rolling_quantile func returns diff results. reduce the row number to 5, the problem will be gone (all three methods return the same results). any thoughts?

ps: i also tested rolling_std func which will "randomly" generate errors with 10^-7 ~ 10^-8 scales (compared to pandas.DataFrame std member func or numpy/scipy std funcs which could limit the error close to np.spacing(1) level) for long (row-wise) pandas.DataFrames

python environment:

python 3.4.2
cython 0.21.1
numpy 1.8.2
scipy 0.14.0
pandas 0.15.1
statsmodels 0.6.0

[shoyer]

It looks like the difference here is that quantile and percentile take the weighted average of the nearest points, whereas rolling_quantile simply uses one the nearest point (no averaging):

In [35]: a = pd.DataFrame(np.arange(row*col).reshape((row,-1)))

In [36]: a
Out[36]:
    0   1   2   3   4
0   0   1   2   3   4
1   5   6   7   8   9
2  10  11  12  13  14
3  15  16  17  18  19
4  20  21  22  23  24
5  25  26  27  28  29
6  30  31  32  33  34
7  35  36  37  38  39
8  40  41  42  43  44
9  45  46  47  48  49

In [37]: a.quantile([0.25,0.5,0.75],axis=0)
Out[37]:
          0      1      2      3      4
0.25  11.25  12.25  13.25  14.25  15.25
0.50  22.50  23.50  24.50  25.50  26.50
0.75  33.75  34.75  35.75  36.75  37.75

In [38]: pd.rolling_quantile(a,row,0.5).tail(1)
Out[38]:
    0   1   2   3   4
9  20  21  22  23  24

So yes, I think we can consider this a bug. rolling_median for example, gets this right:

In [40]: pd.rolling_median(a,row).tail(1)
Out[40]:
      0     1     2     3     4
9  22.5  23.5  24.5  25.5  26.5

@leo4183 Thanks for the report! Please raise another issue for rolling_std (though I suspect the precision issues are known, if perhaps not well documented).

[jreback]

I have xref #8659 there are a number of related issued w.r.t. rolling functions. Pull requests are welcome!

[leo4183]

thx @jreback . as suggested by @shoyer , i created another ticket for rolling_std #9420

[queise]

This issue still happens with the new rolling() objects (pandas 0.18.0).
Following the same example used above by @leo4183 , the quantile() function does not interpolate:
a.rolling(row).quantile(0.25).tail(1)
and thus gives different results than:
np.percentile(a,25,axis=0)
(except for example for row=5 or row=25, when no interpolation is needed).

A workaround is to apply the numpy function, which works for any number in 'row'.
a.rolling(row).apply(func=np.percentile, args=(25,)).tail(1)

On the other hand, median always works:
a.rolling(row).median().tail(1)
that is, it gives the same results as:
np.percentile(a,50,axis=0)

A similar issue may be also happening with the functions std() and var(), which also give different results than the numpy equivalents.

[jreback]

@queise this issue is still open and the impl has not changed

pull requests to fix are welcome

[queise]

thank you @jreback for your quick reply. I'll take a look and try my best.