Merge pull request #7572 from jaimefrio/rolling_var_bug

BUG: Error in rolling_var if window is larger than array, fixes #7297

Author: jreback

doc/source/v0.14.1.txt

@@ -271,6 +271,8 @@ Bug Fixes
 - Bug in non-monotonic ``Index.union`` may preserve ``name`` incorrectly (:issue:`7458`)
 - Bug in ``DatetimeIndex.intersection`` doesn't preserve timezone (:issue:`4690`)
 
+- Bug in ``rolling_var`` where a window larger than the array would raise an error(:issue:`7297`)
+
 - Bug with last plotted timeseries dictating ``xlim`` (:issue:`2960`)
 - Bug with ``secondary_y`` axis not being considered for timeseries ``xlim`` (:issue:`3490`)
 

pandas/algos.pyx

@@ -1173,6 +1173,10 @@ def roll_var(ndarray[double_t] input, int win, int minp, int ddof=1):
 
     minp = _check_minp(win, minp, N)
 
+    # Check for windows larger than array, addresses #7297
+    win = min(win, N)
+
+    # Over the first window, observations can only be added, never removed
     for i from 0 <= i < win:
         val = input[i]
 
@@ -1196,23 +1200,27 @@ def roll_var(ndarray[double_t] input, int win, int minp, int ddof=1):
 
         output[i] = val
 
+    # After the first window, observations can both be added and removed
     for i from win <= i < N:
         val = input[i]
         prev = input[i - win]
 
         if val == val:
             if prev == prev:
+                # Adding one observation and removing another one
                 delta = val - prev
                 prev -= mean_x
                 mean_x += delta / nobs
                 val -= mean_x
                 ssqdm_x += (val + prev) * delta
             else:
+                # Adding one observation and not removing any
                 nobs += 1
                 delta = (val - mean_x)
                 mean_x += delta / nobs
                 ssqdm_x += delta * (val - mean_x)
         elif prev == prev:
+            # Adding no new observation, but removing one
             nobs -= 1
             if nobs:
                 delta = (prev - mean_x)
@@ -1221,6 +1229,7 @@ def roll_var(ndarray[double_t] input, int win, int minp, int ddof=1):
             else:
                 mean_x = 0
                 ssqdm_x = 0
+        # Variance is unchanged if no observation is added or removed
 
         if nobs >= minp:
             #pathological case

pandas/stats/tests/test_moments.py

@@ -366,7 +366,8 @@ def _check_ndarray(self, func, static_comp, window=50,
                        preserve_nan=True,
                        has_center=True,
                        fill_value=None,
-                       test_stable=False):
+                       test_stable=False,
+                       test_window=True):
 
         result = func(self.arr, window)
         assert_almost_equal(result[-1],
@@ -429,6 +430,27 @@ def _check_ndarray(self, func, static_comp, window=50,
             assert_almost_equal(result[-1],
                                 static_comp(self.arr[-50:] + 1e9))
 
+        # Test window larger than array, #7297
+        if test_window:
+            if has_min_periods:
+                for minp in (0, len(self.arr)-1, len(self.arr)):
+                    result = func(self.arr, len(self.arr)+1, min_periods=minp)
+                    expected = func(self.arr, len(self.arr), min_periods=minp)
+                    nan_mask = np.isnan(result)
+                    self.assertTrue(np.array_equal(nan_mask,
+                                                   np.isnan(expected)))
+                    nan_mask = ~nan_mask
+                    assert_almost_equal(result[nan_mask], expected[nan_mask])
+            else:
+                result = func(self.arr, len(self.arr)+1)
+                expected = func(self.arr, len(self.arr))
+                nan_mask = np.isnan(result)
+                self.assertTrue(np.array_equal(nan_mask, np.isnan(expected)))
+                nan_mask = ~nan_mask
+                assert_almost_equal(result[nan_mask], expected[nan_mask])
+
+
+
 
     def _check_structures(self, func, static_comp,
                           has_min_periods=True, has_time_rule=True,