Title here
Summary here
Module · 5 lessons
Exponential Smoothing
Build the exponential smoothing family from the ground up: simple smoothing, Holt's linear trend, and Holt-Winters seasonal smoothing, fit with statsmodels and compared against SARIMA.
At a glance
Level
Intermediate
Lessons
5 lessons
Time to complete
1 week
Cost
Free forever · no sign-up
Welcome to Exponential Smoothing, the seventh module of the course. Modules 5 and 6 (AR, MA, ARMA, ARIMA; Seasonality: SARIMA) built one classical forecasting family: models based on autocorrelation, differencing, and seasonal terms. This module builds a second, older, and in many settings simpler family: exponential smoothing. Instead of modeling autocorrelation directly, these models forecast with a weighted average of past values, where more recent values count for more. The weights shrink exponentially as you look further back, which is where the name comes from.
You will build the family in the same three stages every textbook does, and see exactly why each stage exists. Simple exponential smoothing handles a level with no trend and no season, and you will watch it fail on Cyclepath in a very specific way: with no trend to work with, it converges on the naive forecast from Module 1. Holt’s linear trend adds a trend component, and you will watch it fail even more dramatically on Cyclepath, because a strongly seasonal series can trick it into extrapolating a seasonal dip as if it were a real decline. Holt-Winters seasonal smoothing finally adds a seasonal component, and it recovers almost the exact trend slope and seasonal shape that classical decomposition found back in Module 2, this time through smoothing rather than decomposition.
Every model in this module is fit for real with statsmodels.tsa.holtwinters, on the same seeded Cyclepath series used in every module before it. The two failures along the way are not simplifications for teaching purposes. They are the actual, verified output of fitting those models to this series, and they explain exactly why the next component gets added. Start with Lesson 1 on simple exponential smoothing.
Lessons in this module
1
Simple Exponential Smoothing
Simple exponential smoothing forecasts with a weighted average of the past, where the weights decay exponentially. Fit it on Cyclepath and its forecast turns out to be identical to the naive baseline from Module 1.
2
Holt's Linear Trend
Holt's method adds a trend term to exponential smoothing. On a clean trending toy series it works exactly as intended. On Cyclepath's seasonal series it mistakes a seasonal decline for a real trend and produces a forecast far worse than the naive baseline.
3
Holt-Winters Seasonal
Holt-Winters adds a seasonal component alongside level and trend. Fit on Cyclepath, it recovers almost the exact trend slope and seasonal shape found by decomposition in Module 2, and its test MAPE finally beats the seasonal-naive baseline by more than three and a half times.
4
Fitting, Comparing, and Choosing a Configuration
Lay every exponential smoothing model from this module side by side on AIC and test error, add a damped trend variant, and read the starkest AIC-versus-accuracy disagreement in the whole course.
5
Guided Project: Holt-Winters vs. SARIMA vs. the Baselines
Put Holt-Winters, SARIMA, and every baseline from the course on one table, forecast a genuine future year with both models, and decide when each classical family is the right tool.
Achievement
Complete all 5 lessons to finish the Exponential Smoothing module.
