CODEWITHSUNDEEP
pythonnumpypandasstatsmodels
I. F.
I. F.

Reputation: 91

linear regression for timeseries python (numpy or pandas)

I am new to python and programming in general, so forgive any simple mistakes/ things that should be obvious.

What I am trying to do is quite simple, I just want to fit a linear trend (1-d polynomial) to a bunch of time-series to see whether the slopes are positive or negative. Right now I am just trying to get it to work for one time series.

The problem: it seems like both pandas and numpy can't do regressions for datetimes. My date times are not regular (generally 1 day per month but not the same day) so can't use the suggestion posed in Linear Regression from Time Series Pandas

My time series csv looks like:

StationName,    year,   month,  day,    depth,  NO3-N,  PO4-P,  TotP,   TotN,
Kvarnbacken (Savaran),  2003,   2,  25, 0.5,    46, 9,  14, 451
Kvarnbacken (Savaran),  2003,   3,  18, 0.5,    64, 15, 17, 310
Kvarnbacken (Savaran),  2003,   3,  31, 0.5,    76, 7,  19, 566

so far what i have is 

import datetime as dt
from scipy import stats
import numpy as np

# read in station csv file
data = pd.read_csv('Kvarnbacken (Savaran)_2003.csv')
data.head()
# set up dates to something python can recognize
data['date'] = pd.to_datetime(data.year*10000+data.month *
                          100+data.day, format='%Y%m%d')

I tried 

slope, intercept, r_value, p_value, std_err = stats.linregress(data.date,
                                                               data.TotP)

and got the error TypeError: ufunc add cannot use operands with types dtype('

I also tried

coefP = np.polyfit(data.date, data.TotP, 1)
polyP = np.poly1d(coefP)
ys = polyP(data.date)
print 'For P: coef, poly'
print coefP
print polyP

and got the same error.

I am guessing the easiest way around this is to do something where I just count the days since the first measurement I have and then just do a regression with days_since to the total phosphorous concentration (totP) but I am not sure of the easiest way to do that or if there was another trick.

Upvotes: 8

Views: 11336

Answers (1)

JohnE
JohnE

Reputation: 30424

You could convert the datetime to days in the following way.

data['days_since'] = (data.date - pd.to_datetime('2003-02-25') ).astype('timedelta64[D]')

        date  days_since
0 2003-02-25           0
1 2003-03-18          21
2 2003-03-31          34

Now you should be able to regress as you did above.

slope, intercept, r_value, p_value, std_err = stats.linregress(data.days_since, 
                                                               data.TotP)
slope, intercept
(0.1466591166477916, 13.977916194790488)

You might also want to consider other regression options such as the statsmodels package, especially if you'll be doing this sort of thing very often. (Note that x and y are reversed compared to linregress)

import statsmodels.formula.api as smf

smf.ols( 'TotP ~ days_since', data=data ).fit().params

Intercept     13.977916
days_since     0.146659

That's just a fraction of the statsmodels output btw (use summary() instead of params to get the extra output.

Upvotes: 12

Related Questions