#StackBounty: #time-series #arima #trend #change-point Working with Time Series: What drives a trend change?

Bounty: 50

I’m working with a time series that had a clear trend changepoint where it went from having an upward trend to a downward trend (After accounting for seasonality).

The problem that I’m running into is that many of the techniques I know assume a stationary time series – if I go and make this time series stationary than I cant go and find out whats impacting the trend because the trend is just gone.

What techniques exist to find what other features/regressors could be driving that trend change? Can I just simply do a time series decomposition and regress some features onto the trend component? Is there anything more robust?

edit: Here’s a fake generated time series with a similar pattern to what I’m working with (I manually added in seasonality), just note I didn’t incorporate any regressors into this dataset despite that being a key part of my problem (since I’m not quite sure how to model it, I probably shouldnt make a generative model with it):

x1 = np.random.normal(0.06, 0.0025, size=106)
time = np.arange(106)
changept = 60
time2 = (time > changept)
trend = time2*time - changept*time2
y = 0.0015*time + x1 + trend*-0.00225
dates = pd.date_range('2012-06-10', '2014-06-15', freq='W')
seasonality = np.array([-0.0109,-0.0077,-0.0037,-0.0032,

y += seasonality

df = pd.DataFrame({'y':y, 'changepoint':time2*1}, index=dates)

Get this bounty!!!

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.