Statsmodels seasonal_decompose - what is naive about it?

Question

Have been working with time series in Python, and using sm.tsa.seasonal_decompose. In the docs they introduce the function like this:

We added a naive seasonal decomposition tool in the same vein as R’s decompose.

Here is a copy of the code from the docs and its output:

import statsmodels.api as sm

dta = sm.datasets.co2.load_pandas().data
# deal with missing values. see issue
dta.co2.interpolate(inplace=True)

res = sm.tsa.seasonal_decompose(dta.co2)
res.plot()

They say it is naive but there is no disclaimer about what is wrong with it. Does anyone know?

Vinnie

There is nothing wrong with it. I guess it's "naive" because it uses simple convolution and averages to separate out trend and seasonal structure, i.e. it's not a fancy algorithm. – Sharisharia

Thanks, @user333700 , I wonder if R's decompose is also "naive" and if and when it is ever necessary to go fancy. Get the impression given that .seasonal_decompose worked on my data that it's probably all you typically need. – Vinnie

One possibility where it is most likely too "naive", i.e. doesn't capture the features of the data appropriately, is shifting seasonal patterns, for example because of a trend or break in the pattern. Sometimes those can also be removed by a box-cox or similar transformation. Seasonal adjustments for macro economic data (e.g. by X11) also take shifts in holidays and similar calendar effects into account. – Sharisharia

Answer 1

I made some (aehm...naive) researches, and, according to the reference, it seems that StatsModels uses the classic moving average method to detect trends and apply seasonal decomposition (you can check more here, specifically about Moving Average and Classical Decomposition).

However, other advanced seasonal decomposition techniques are available, such as STL decomposition, which also has some Python implementations. (UPDATE - 11/04/2019 as pointed out in the comments by @squarespiral, such implementations appear to have been merged in the master branch of StatsModels).

At the above links, you can find a complete reference on the advantages and disadvantages of each one of the proposed methods.

Hope it helps!

Richela

The "Python implementation" of STL decomposition is apparently already merged into statsmodels: github.com/statsmodels/statsmodels/issues/4044 - the documentation for the function is here: statsmodels.org/dev/examples/notebooks/generated/… – Lovesome

Thanks @squarespiral, I updated the answer according to your comment. – Richela

Great answer. Still, the +1 was mostly for the pun. – Mining

Answer 2

According to the source codes, statsmodels.tsa.seasonal() de-constructs a time series by:

1. Trend is computed by a convolutional window, decided by the argument period=12, for example;
2. The detrended time-series is obtained by either Raw-Trend or Raw/Trend, controlled by the argument model;
3. The seasonal is finally computed by averaging all temporal segments within a single period (e.g., 12 here);
4. Finally, the residual is the noise left.

I would say, it is somehow conventional and like a feature-engineering work for machine learning tasks. Not sure whether it will improve the performance for some advanced deep learning architectures (e.g. Transformer).

Soni