https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/Recent content in Exponential Smoothing on DATATWEETSHugoenCopyright (c) 2025 DatatweetsSun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-1-simple-exponential-smoothing/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-1-simple-exponential-smoothing/Simple exponential smoothing (SES) forecasts a series with no trend and no season as a single smoothed level, updated by one parameter alpha that trades off trusting history against trusting the latest observation. Fit with statsmodels on Cyclepath, alpha optimizes to exactly 1.0, and the forecast turns out to be numerically identical to the naive baseline from Module 1: same values, same 18.99% MAPE. This is not a coincidence. It is what SES reduces to when a series has a trend it cannot represent.https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-2-holts-linear-trend/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-2-holts-linear-trend/Holt’s linear trend method adds a second smoothed component, the trend, updated alongside the level. On a toy series with a genuine linear trend and no seasonality, it recovers the true slope almost exactly. Fit on Cyclepath, both smoothing parameters are pushed to about 0.986, the model reads the steep seasonal decline from the September peak to the December trough as a real trend, and extrapolates it into a forecast that crashes from 12,468 to 461 over the test year, a 59.93% MAPE, far worse than even the naive baseline. Its AIC is nevertheless better than simple exponential smoothing’s, an extreme case of AIC and forecast accuracy disagreeing.https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-3-holt-winters-seasonal/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-3-holt-winters-seasonal/Holt-Winters seasonal smoothing adds a third smoothed component, the season, alongside the level and trend from Holt’s method. Fit on Cyclepath with an additive season, all three smoothing parameters optimize to 0, meaning a single fixed level, trend, and seasonal pattern estimated once already fits almost perfectly, and the initial level (8,973.8) and trend (89.4) closely match the series’ true generative values (9,000 and 90). The test MAPE is 1.57%, more than three and a half times better than the 5.9% seasonal-naive baseline. A multiplicative seasonal variant scores a slightly better test MAPE (1.32%) but a clearly worse AIC, echoing Module 2’s confirmation that Cyclepath is additive.https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-4-fitting-comparing-and-choosing-a-configuration/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-4-fitting-comparing-and-choosing-a-configuration/Comparing simple exponential smoothing, Holt’s linear trend, and two Holt-Winters variants side by side on Cyclepath produces the most extreme AIC-versus-test-error disagreement in the course: Holt’s method has a better AIC than simple exponential smoothing (1147.01 versus 1199.74) while forecasting more than three times worse (59.93% MAPE versus 18.99%). A damped-trend variant of Holt-Winters is also tested, with its damping parameter landing at 0.995, essentially undamped, confirming Cyclepath’s real trend never decelerates. The additive, non-damped Holt-Winters model remains the best overall choice, by AIC, by test error, and by matching the series’ known structure.https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-5-guided-project-holt-winters-vs-sarima/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/exponential-smoothing/lesson-5-guided-project-holt-winters-vs-sarima/The Module 7 capstone scores Holt-Winters against the naive baseline (19.0% MAPE), the seasonal-naive baseline (5.9%), the best non-seasonal ARIMA (16.0%), and SARIMA (1.06%). Holt-Winters lands at 1.57% MAPE, 3.76 times better than the seasonal-naive bar, though SARIMA still forecasts about 1.5 times more accurately. Retrained on the full series, Holt-Winters projects 2024 to a total of 216,824 trips, close to SARIMA’s 213,839, two independently built models agreeing on the shape of a year that has not happened yet.