rolling_mean and rolling_sum produce negative output from positive input

[mrjbq7]

I have an array of non-negative numbers, that when used with rolling_sum or rolling_mean produce an output array that has a small negative number in it.

The test looks like this:

import numpy as np
import pandas

data = np.load('data.npy')
assert all(data >= 0)

sums = pandas.rolling_sum(data, 2, min_periods=1)
zero = np.where(sums < 0)[0]
assert len(zero) == 0, zero

mean = pandas.rolling_mean(data, 2, min_periods=1)
zero = np.where(mean < 0)[0]
assert len(zero) == 0, zero

It requires a small binary array to reproduce, because of the floating point numbers (so I created a gist: https://gist.github.com/3948013).

You can run the test case:

$ git clone git://gist.github.com/3948013.git
$ cd 3948013
$ python test.py

I made sure this bug affects the most current version of Pandas:

>>> import pandas
>>> pandas.__version__
'0.9.1.dev-8cd93d3'

>>> import numpy
>>> numpy.__version__
'1.7.0b2'

[mrjbq7]

It looks like bottleneck doesn't have this problem...

[mrjbq7]

I used a similar approach to bottleneck where the new value is added first before the previous value is subtracted and it fixes this particular bug...

diff --git a/pandas/src/moments.pyx b/pandas/src/moments.pyx
index 503a63c..9b5b621 100644
--- a/pandas/src/moments.pyx
+++ b/pandas/src/moments.pyx
@@ -175,16 +175,16 @@ def roll_sum(ndarray[double_t] input, int win, int minp):
     for i from minp - 1 <= i < N:
         val = input[i]

+        if val == val:
+            nobs += 1
+            sum_x += val
+
         if i > win - 1:
             prev = input[i - win]
             if prev == prev:
                 sum_x -= prev
                 nobs -= 1

-        if val == val:
-            nobs += 1
-            sum_x += val
-
         if nobs >= minp:
             output[i] = sum_x
         else:
@@ -218,16 +218,16 @@ def roll_mean(ndarray[double_t] input,
     for i from minp - 1 <= i < N:
         val = input[i]

+        if val == val:
+            nobs += 1
+            sum_x += val
+
         if i > win - 1:
             prev = input[i - win]
             if prev == prev:
                 sum_x -= prev
                 nobs -= 1

-        if val == val:
-            nobs += 1
-            sum_x += val
-
         if nobs >= minp:
             output[i] = sum_x / nobs
         else: