https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/Recent content in Seasonality: SARIMA on DATATWEETSHugoenCopyright (c) 2025 DatatweetsFri, 03 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-1-from-arima-to-sarima/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-1-from-arima-to-sarima/A SARIMA model is written SARIMA(p,d,q)(P,D,Q)[s]: the familiar non-seasonal orders plus a seasonal set that applies the same AR, differencing, and MA ideas at the seasonal lag s (12 for monthly-yearly data). This lesson introduces the full notation, maps each seasonal parameter to its non-seasonal twin, and explains why the seasonal terms are exactly the mechanism a non-seasonal ARIMA lacked for capturing ’this July resembles last July.'https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-2-seasonal-ar-and-ma-terms/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-2-seasonal-ar-and-ma-terms/Seasonal AR and MA terms are the ordinary AR and MA ideas applied at the seasonal lag: a seasonal-AR(1) says this period is a fraction of the same period one cycle ago. This lesson builds a pure seasonal AR(1) process with a known coefficient of 0.6, recovers it as 0.598, and shows its distinctive fingerprint — ACF spikes at lags 12, 24, 36 decaying geometrically while short lags stay near zero — the seasonal echo of Module 4’s AR signature.https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-3-fitting-sarima-with-statsmodels/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-3-fitting-sarima-with-statsmodels/The SARIMAX class fits a full SARIMA from an order and a seasonal_order tuple. This lesson fits SARIMA(1,1,0)(1,1,0)[12] to Cyclepath’s training set, reads a summary showing ar.L1 = -0.499 and ar.S.L12 = -0.509 (both significant at p < 0.001), and forecasts the test year — with predictions that track the actual seasonal rise and fall closely, unlike the flat non-seasonal forecast of Module 5.https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-4-diagnostics-is-the-model-adequate/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-4-diagnostics-is-the-model-adequate/Model diagnostics ask whether the residuals are white noise — zero mean, no leftover autocorrelation, roughly normal, constant variance. This lesson runs the Ljung-Box test on the SARIMA(1,1,0)(1,1,0)[12] residuals (p = 0.209 at lag 12, passing) and contrasts it with the Module 4 preview model that failed the same test at p = 0.000. It reads the summary’s built-in normality and heteroskedasticity checks, and honestly examines a small residual-ACF wrinkle that the joint test tolerates.https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-5-guided-project-a-sarima-forecast-for-cyclepath/Fri, 10 Apr 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/seasonality-sarima/lesson-5-guided-project-a-sarima-forecast-for-cyclepath/The Module 6 capstone and the course’s turning point. You’ll fit SARIMA(1,1,0)(1,1,0)[12] to Cyclepath, confirm it passes diagnostics, and score its forecast against every baseline from the course: naive (19%), seasonal-naive (5.9%), and the best non-seasonal ARIMA (16%). The SARIMA lands at 1.06% MAPE — beating the seasonal-naive bar decisively for the first time — then you retrain on the full series to forecast 2024, turning a validated model into a real prediction.