https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/Recent content in Components and Decomposition on DATATWEETSHugoenCopyright (c) 2025 DatatweetsSun, 28 Jun 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-1-trend-seasonality-and-residual/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-1-trend-seasonality-and-residual/Decomposition splits a series y into trend (T), seasonality (S), and residual (R) — the slow drift, the repeating calendar pattern, and whatever’s left after removing both. This lesson defines each component precisely, previews the additive and multiplicative combination rules, and explains why isolating them makes forecasting, diagnosis, and modeling all easier.https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-2-classical-decomposition-by-hand/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-2-classical-decomposition-by-hand/Classical decomposition is three steps: smooth out the seasonality and noise with a centered moving average to estimate trend, subtract that trend to expose the seasonal-plus-residual shape, then average by calendar month to isolate the seasonal indices, leaving the residual. Built by hand on Cyclepath with pandas, the result matches statsmodels’ seasonal_decompose to the decimal.https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-3-additive-vs-multiplicative-models/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-3-additive-vs-multiplicative-models/The additive-vs-multiplicative choice comes down to one question: does the seasonal swing stay a constant absolute size as the trend rises, or does it scale as a constant percentage of the level? This lesson tests that directly by tracking swing-to-level ratio year over year, applies it to Cyclepath (additive) and a contrasting series where the swing genuinely grows (multiplicative), and shows the log-transform trick that turns one into the other.https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-4-stl-decomposition/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-4-stl-decomposition/STL replaces the classical method’s simple moving average with LOESS smoothing, which covers the full series with no missing edges and, in robust mode, resists being dragged around by outliers. Tested on Cyclepath with an injected spike, STL’s robust trend moves by only 17.8 at most versus 666.7 for the classical moving-average trend — a 37x difference from one bad data point.https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-5-guided-project-decomposing-cyclepath/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/components-and-decomposition/lesson-5-guided-project-decomposing-cyclepath/The Module 2 capstone. You’ll rebuild Cyclepath, decompose it classically and with STL, confirm additive is the right model with real evidence, and interpret every component: 78.4% trend growth over eight years, a July peak of +3,307 trips against a January trough of -3,400, and a residual with no leftover structure — clean noise, ready for Module 3’s stationarity tests.