ValueError: freq T not understood. Please report if you think this in error. (seasonal_decompose)

Question

I was trying out seasonal_decompose to decompose my time series. The data is perfect time series with frequency of '2T' i.e. 2 Minutes. From file tsatools.py (site-pkgs\statsmodels\tsa\tsatools.py), in line 655, I added the following. _ elif freq == 'T': return 6024752_ I added this from the following inference: Freq A means 1 year hence it returns 1. Q means quaterly and hence returns 4 M means monthly and hence returns 12 and so on. Therefore, T means per minute, hence 60247*365

When I do the above, I get following error: ValueError: Inferred frequency of index and frequency don't match. This function does not re-sample From line 70 in seasonal.py (statsmodel\tsa\seasonal.py) Because : variable freq is : <2 * Minutes> And variable pfreq is 2T 524160.

I mean seasonal decompose should be able to decompose timeseries of 1min frequency, and something seems to have changed. Please have a look at it, and let me know if I'm missing anything.

Mendive

This has been fixed in statsmodels master and will be changed in the upcoming release so it will be possible to override the default period length freq. – Sholokhov

Answer 1

I got the same problem, I think it's an internal problem since I test all the index data type. At the end, I tried this and it works.

import statsmodels.api as sm
decompfreq = 6*12
decomposition = sm.tsa.seasonal_decompose(ts3_log.values,freq=decompfreq)
trend = decomposition.trend

decompfreq is calculated based on the time window, which is 10 mins, so the freqency actually is half a day.

Hope this would help you

Cure

If I understand correctly, does freq need to be the # of samples in a season? So if I had daily seasonality and my samples are every 5 minutes, then freq = ((24*60)//5)? – Outright

yup, I think so – Cure

What is that supposed to mean? Number of samples in a season? Am I supposed to know the season/seasons beforehand? – Selvage

I don't understand. Then it's simply the len of my series? – Chevalier

Answer 2

"You need to explicitly set the period when using data that does not have an obvious seasonality." - User bashtage.

If your data is in seconds, then try:

sm.tsa.seasonal_decompose(series, period = 60*60*24) to represent a daily period.

Reference: the statsmodels.tsa.tsatools documentation for freq_to_period.

Crotchety