https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/Recent content in Stationarity and Differencing on DATATWEETSHugoenCopyright (c) 2025 DatatweetsFri, 03 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-1-what-is-stationarity/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-1-what-is-stationarity/Stationarity means a series’ statistical properties — mean, variance, autocovariance — stay constant over time, so any window looks statistically like any other. This lesson defines the three requirements precisely, explains why ARIMA-family models need them, and shows informally that Cyclepath fails: its first-half mean (11,142.9) is 38% below its second-half mean (15,391.1), while its variance stays comparatively stable.https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-2-the-augmented-dickey-fuller-test/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-2-the-augmented-dickey-fuller-test/The Augmented Dickey-Fuller test’s null hypothesis is that a series has a unit root — is non-stationary. A small p-value (conventionally under 0.05) lets you reject that null and conclude stationarity; a large one means you can’t. Run with statsmodels.tsa.stattools.adfuller on raw Cyclepath, the test returns an ADF statistic of -0.920 and a p-value of 0.7815 — decisively failing to reject non-stationarity, exactly as Lesson 1’s informal check predicted.https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-3-differencing-to-remove-trend/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-3-differencing-to-remove-trend/First differencing (d=1) replaces y_t with y_t - y_{t-1}, removing a linear trend by construction. Applied to Cyclepath, the ADF p-value collapses from 0.7815 to under 0.0001 — the series passes the stationarity test. But its autocorrelation at lag 12 is a striking 0.801, proof that ADF-stationary doesn’t mean structure-free: the seasonal pattern differencing wasn’t built to remove is still fully present.https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-4-seasonal-differencing-and-transformations/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-4-seasonal-differencing-and-transformations/Seasonal differencing (y.diff(12)) removes the year-over-year relationship instead of the month-over-month one. Alone, it passes the ADF test (p=0.0001) with variance of 124,680 — over 12 times smaller than regular differencing’s 1,567,486. Combining regular and seasonal differencing still passes (p=0.0002) but variance rises to 257,769, a textbook case of overdifferencing. This lesson also covers when a log or Box-Cox transformation helps, and why Cyclepath’s confirmed additive structure means it doesn’t need one.https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-5-guided-project-making-cyclepath-stationary/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/stationarity-and-differencing/lesson-5-guided-project-making-cyclepath-stationary/The Module 3 capstone. You’ll test raw Cyclepath (fails, p=0.7815), try regular differencing (passes but leaves a lag-12 echo of 0.801), try seasonal differencing alone (passes, with variance 124,680 — the lowest of any option), and try combining both (passes but overdifferences, variance 257,768). The evidence points to one clear winner: seasonal differencing alone, the series Module 4 will read ACF and PACF from.