I just need to find the three warmest consecutive months from a temperature dataset for my work, so I thought I add a complete example. First, create an example dataset with monthly temperature:

```python import xarray as xr import numpy as np import pandas as pd

time = pd.date_range("2000", periods=12 * 30, freq="M") temp = np.sin((time.month - 5) / 6 * np.pi) + np.random.randn(*time.shape) * 0.3 da = xr.DataArray(temp, dims=["time"], coords=dict(time=time)) print(da) ```

python <xarray.DataArray (time: 360)> array([-0.676731, -0.812742, -1.367547, ..., 0.186731, 0.237676, -0.343879]) Coordinates: * time (time) datetime64[ns] 2000-01-31 2000-02-29 ... 2029-12-31

Currently we can achieve this like:

```python n_months = 3

monthly = da.groupby("time.month").mean() padded = monthly.pad(month=n_months, mode="wrap") rolled = padded.rolling(center=True, month=n_months).mean(skipna=False) sliced = rolled.isel(month=slice(3, -3))

central_month = sliced.idxmax() ```

Implementing pad_mode in rolling would allow to do:

```python monthly = da.groupby("time.month").mean() rolled = monthly.rolling(center=True, month=n_months, pad_mode="wrap").mean(skipna=False)

central_month = rolled.idxmax() ```

I am coming back to @shoyer suggestion in #2011 - your idea would be to do first a pad and then a rolling operation as e.g.:

``` python

import numpy as np import xarray as xr

x = np.arange(1, 366) y = np.random.randn(365) ds = xr.DataArray(y, dims=dict(dayofyear=x))

ds.pad(dayofyear=15, mode='wrap').rolling(center=True, dayofyear=31).mean()

```

I think what I want is like the filter function in R with circular=True.

I found two possibilities but they are quite "hand made" and certainly not very efficient

Solution with slicing: ``` python

take the last and first elements and append/ prepend them

first = ds[:15] last = ds[-15:] extended = xr.concat([last, ds, first], 'dayofyear')

do the rolling on the extended ds and get rid of NaNs

sol1 = extended.rolling(dayofyear=31, center=True).mean().dropna('dayofyear') ```

Solution with roll: python roll1 = ds.roll(dayofyear=150).rolling(dayofyear=31, center=True).mean() roll2 = ds.rolling(dayofyear=31, center=True).mean() sol2 = xr.concat([roll1, roll2], dim='r').mean('r') Call rolling on original and rolled dataset, and put them together again.

Probably a mix of both - I want to compute a moving average, but with periodic boundaries. rolling sets window_length - 1 elements to nan. However I want to calculate these like so:

running_mean_0 = xr.concat([ds[-15:], ds[:16]]).mean()
running_mean_1 = xr.concat([ds[-14:], ds[:17]]).mean()

and so on...