https://datatweets.com/courses/time-series-forecasting/capstone/Recent content in Capstone on DATATWEETSHugoenCopyright (c) 2025 DatatweetsSun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/capstone/lesson-1-meet-lantern-and-vine/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/capstone/lesson-1-meet-lantern-and-vine/Lantern & Vine is a new, seeded, weekly series: 208 weeks of a fictional home-goods shop’s unit sales, growing from a mean of 624 units a week in its first year to 1,465 in its fourth. A first exploration, summary statistics, yearly totals, and a rolling mean, mirrors Module 1’s approach to Cyclepath exactly, and immediately surfaces three differences this capstone will have to handle: a weekly rather than monthly frequency, meaning a seasonal period of 52 instead of 12, and year-over-year growth that is visibly decelerating rather than constant.https://datatweets.com/courses/time-series-forecasting/capstone/lesson-2-a-trend-that-changes-pace/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/capstone/lesson-2-a-trend-that-changes-pace/Applying STL to Lantern & Vine produces a seasonal component whose swing grows every year, 775.8, then 800.8, then 968.0, then 1,218.2, the same growing-swing signature Module 2 used to identify multiplicative structure, confirmed here by a classical multiplicative decomposition whose seasonal factor (0.63 to 1.385) closely matches the series’ true generative values. Fitting separate growth rates to the trend’s first and second halves finds 0.78% a week decelerating to 0.36%, a genuine, datable change in pace that Cyclepath’s constant trend never had to contend with.https://datatweets.com/courses/time-series-forecasting/capstone/lesson-3-stationarizing-and-reading-the-autocorrelation/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/capstone/lesson-3-stationarizing-and-reading-the-autocorrelation/Working in log space, as Lesson 2’s multiplicative finding requires, raw log sales fail the ADF test (p=0.3192). First differencing alone passes decisively (p=0.0000, variance 0.00375), the lowest variance of any option tested. Seasonal differencing alone, the transformation that won cleanly on Cyclepath, fails outright here (p=0.8451), a direct consequence of the regime change Lesson 2 found. The autocorrelation left in the first-differenced series still shows a real spike at the seasonal lag (0.154 at lag 52), evidence that a SARIMA seasonal term is still needed even though seasonal differencing itself was not the fix.https://datatweets.com/courses/time-series-forecasting/capstone/lesson-4-fitting-and-backtesting-candidate-models/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/capstone/lesson-4-fitting-and-backtesting-candidate-models/Five SARIMA specifications on log Lantern & Vine sales, tested on a single held-out 26-week block, show a stark split: only the specification with a seasonal AR and seasonal difference term, order (1,1,0) seasonal (1,1,0,52), reaches 3.74% MAPE. Every other candidate, missing the seasonal term, scores 23% or worse, far short of the 15.68% seasonal-naive bar. A multiplicative Holt-Winters model scores a close 3.85%. Backtested properly across six origins with a 13-week horizon, the ranking from Module 8 repeats: Holt-Winters (mean 3.76%, std 0.66) edges out SARIMA (mean 4.39%, std 1.10) on both accuracy and stability, a second series confirming the same lesson.https://datatweets.com/courses/time-series-forecasting/capstone/lesson-5-guided-project-the-final-report/Sun, 05 Jul 2026 09:00:00 +0200https://datatweets.com/courses/time-series-forecasting/capstone/lesson-5-guided-project-the-final-report/The course capstone retrains Holt-Winters multiplicative on every available week of Lantern & Vine and forecasts January through June 2024: a January high of 2,037.8 units easing to 1,142.6 by late June, a seasonal pattern rather than a decline, confirmed by comparing against the same half of the previous year, an honest 18.8% year-over-year growth. The report that follows walks through all seven pipeline stages this module built, and the eight modules before it, the complete arc of the course from a first line plot to a validated, honestly reported forecast.