The pandas wrapper in decompose() keeps track of the observation DatetimeIndex so it can be used properly in the forecast() method.

There are two potential issues here:

• while the decomposition seems to work fine without continuous data e.g. no missing data, it does require a DatetimeIndex
• though the decomposition would work with missing data, the forecast will fail awkwardly if there is a bad DatetimeIndex

A couple of possibilities:

• the decompose method could raise an exception if the input is missing a DatetimeIndex
• the forecast method could catch exceptions and raise this as a possibly helpful message

stl.py references _maybe_get_pandas_wrapper_freq which no longer exists in current and future versions of statsmodels.

It appears statsmodels has merged a PR to implement this years-old request! statsmodels/statsmodels#4044 Since this library primarily extends the existing statsmodels internals, they will probably do this more robustly and flexibly.

For those currently using this library, there are no plans to delete this library! I've still got some net new things I'd like to do with it.

Python version: 3.7.*
Statsmodel version: statsmodels==0.11.0
Stldecompose version: stldecompose==0.0.5
Error: ImportError: cannot import name '_maybe_get_pandas_wrapper_freq' from 'statsmodels.tsa.filters._utils'

Hello,

In the latest version of the stldecompose library, the import in line 5 of stl.py doesn't work with the latest versions of statsmodels as that function has already been deleted in a commit done back in July 2019

Currently, the decompose method takes a single period value for the decomposition. This could be extended to accept an iterable of frequency values which are sequentially factored out of the observation to build up the seasonal component.

Another option would be to automatically calculate the top frequencies/periodicities to factor out by calling out to a scipy FFT routine and setting some thresholding about how big (or small) a frequency component must be before we stop.

Unclear if it's an encoding issue or the link is no longer good. Consider finding another reference link.

When I perform decompose(x), the returned seasonal and residual value for each observation is NaN. Could you check this?
For the following lines of code, the output is given below.
result=stldecompose.decompose(X)
print(result.seasonal) #same output for result.resid

1998-05-06 NaN
1998-05-07 NaN
1998-05-08 NaN
1998-05-09 NaN
1998-05-10 NaN
1998-05-11 NaN
1998-05-12 NaN
1998-05-13 NaN
1998-05-14 NaN
1998-05-15 NaN
1998-05-16 NaN
1998-05-17 NaN
1998-05-18 NaN
1998-05-19 NaN
1998-05-20 NaN
1998-05-21 NaN
1998-05-22 NaN
1998-05-23 NaN
1998-05-24 NaN
1998-05-25 NaN
1998-05-26 NaN
1998-05-27 NaN
1998-05-28 NaN
1998-05-29 NaN
1998-05-30 NaN
1998-05-31 NaN

The procedure described in Cleveland et al. 1990 is very different from the code in stl.py. I would recommend changing the name to seasonal_decompose instead of STL. It is very misleading...

Dear jrmontag:
Hello! I'd like to research if there are some decomposition methods which can process embedded data.
These data's dimension is transfer from (length,) to (length, embedding). So STL can do this?

Dear Josh,

I’m working with timeseries of water supply for a city of more than 8 millions of inhabitants.
Really you project has a high interest for my purposes.

However, when install it under python 3.7 obtain an error: “cannot import name ‘factorial’ from ‘scipy.misc’.
Your .whl installed the scipy lib version 1.3.0 in which the 'scipy.misc.factorial’ was deprecated (from version 1.0.0) and now the function
factorial is under ‘scipy.special’.

May you modify this calling to scipy.special.factorial please. Otherwise, will be necessary downgrade scipy to 0.19.1 version

Please some advice on this situation, thanks on advance,

Raul

PD: the stldecomposemod project v=0.0.3 has the same error.

I have been looking at this project lately and noticed that the forecast function fails when I pass a dataframe that has a frequency offset of a minute (not exactly a minute, but any frequency that is a multiple of T/min).

I tried to look closer, but could not figure out the exact reason. Even the usage example (https://github.com/jrmontag/STLDecompose/blob/master/STL%20usage%20example.ipynb) breaks when I resample with frequency in minutes. Any help/pointers would be great.

It'd be nice to have some guidance for how others can best contribute to this project e.g. a dev requirements.txt file (twine, wheel, ?), tests to check before submitting a PR, etc.

I know that you can tune the model in the r package, but how can I tune it in python?

-Thanks!