Time Series
by Ryan Holbrook · Kaggle
Our Verdict
Worth it — with caveatsKaggle Learn's Time Series is a free, hands-on micro-course (six interactive lessons, roughly five hours) taught by Kaggle's Ryan Holbrook that teaches forecasting through a feature-engineering-plus-machine-learning lens rather than classical statistics. Working entirely in the browser with pandas, statsmodels and scikit-learn, you build forecasting models from time-based features: a time dummy and lags, trend via moving averages, seasonality with Fourier features, then hybrid models (linear regression + XGBoost residuals) and four ML forecasting strategies (MultiOutput, Direct, Recursive, DirRec). It is best understood as a practical, competition-flavored on-ramp: the lessons are derived from winning Kaggle solutions and feed directly into the Store Sales companion competition. Notably, it deliberately skips ARIMA, exponential smoothing/ETS, Prophet and deep-learning sequence models, so it is a complement to a classical-forecasting course, not a replacement. We could not verify a credible aggregate star rating (Class Central lists 5.0 from a single review), so treat any headline rating with caution and judge it on curriculum fit.
It is an excellent, free, and unusually practical introduction IF you already know Python/pandas and want the ML-feature-engineering approach to forecasting; it is a poor fit if you need classical statistical methods (ARIMA/ETS/Prophet) or deep learning, since the course intentionally omits them.
Best for: Data scientists and analysts who already write Python and pandas comfortably and want a fast, applied path to ML-based forecasting and Kaggle tabular competitions. It is ideal for people who learn by doing, want to feed into the Store Sales competition, and are interested in feature engineering (lags, Fourier seasonality) and hybrid linear+gradient-boosting models rather than statistical theory.
Skip if: Complete programming beginners (it assumes Python/pandas and jumps straight into scikit-learn LinearRegression and XGBRegressor without teaching ML basics), and anyone whose goal is classical statistical forecasting (ARIMA, SARIMA, exponential smoothing/ETS, Prophet) or deep-learning sequence models (RNN/LSTM/Transformers), none of which are covered. Learners who want rigorous coverage of stationarity, autocorrelation/ACF-PACF, or probabilistic/uncertainty forecasting should look elsewhere.
About This Course
Forecast time series data using linear regression, trend analysis, seasonality, and hybrid forecasting models.
What You'll Learn
Curriculum
Use the time dummy (time-step feature) and lag features to fit linear regression models to serial data, and interpret the coefficients to read off trend and serial dependence.
Model long-term, non-periodic change with moving averages and a polynomial time dummy built using statsmodels' DeterministicProcess.
Create seasonal indicators and Fourier features (CalendarFourier) guided by a periodogram to capture periodic change, deseasonalize a series, and add holiday features.
Use lag features and serial dependence (e.g., cycles) to forecast, including building lagged datasets and selecting useful lags.
Combine the strengths of two learners by fitting a simple model and then a boosting model on its residuals; implement a BoostedHybrid class pairing LinearRegression with XGBRegressor.
Define the forecasting task (origin, horizon, lead time), build multistep datasets for multiple series, and apply the MultiOutput, Direct, Recursive, and DirRec strategies (DirRec with XGBoost on Store Sales).
Prerequisites
- Working knowledge of Python and pandas (DataFrames, indexing, datetime handling)
- Basic familiarity with scikit-learn-style models is strongly recommended (the course uses LinearRegression and XGBRegressor from lesson 1 onward without teaching ML fundamentals); completing Kaggle's Intro to Machine Learning and Feature Engineering first is the natural lead-in
- Comfort reading plots (matplotlib/seaborn); no prior time-series background required
Instructor
Ryan Holbrook
Instructor · Kaggle
Pros & Cons
Pros
- Completely free and frictionless: runs in-browser on Kaggle Notebooks with the data and learntools exercises pre-wired, so there is zero setup and you finish in about five hours
- Genuinely practical and competition-tested: lessons are drawn from winning Kaggle forecasting solutions and feed directly into the Store Sales - Time Series Forecasting competition for immediate application
- Strong, modern feature-engineering toolkit you rarely get in classical courses: Fourier seasonality, hybrid linear+gradient-boosting models, and the four ML forecasting strategies (MultiOutput/Direct/Recursive/DirRec)
- Clear, well-scoped, and beginner-friendly within its niche, taught by Kaggle's own Ryan Holbrook with an emphasis on interpretable models; includes a certificate of completion
- Honest academic grounding: references Hyndman & Athanasopoulos' Forecasting: Principles and Practice and other standard texts rather than overselling
Cons
- Deliberately omits classical statistical forecasting (ARIMA/SARIMA, exponential smoothing/ETS) and Prophet, as well as deep-learning models (RNN/LSTM/Transformers) and probabilistic/uncertainty forecasting, so it is incomplete as a standalone time-series education
- Assumes prior Python/pandas and ML comfort: it uses scikit-learn and XGBoost from the start without teaching the fundamentals, which can leave true beginners stuck
- Short and shallow by design (about five hours, six lessons); light on theory such as stationarity, ACF/PACF diagnostics, and proper time-series cross-validation
- No reliable public rating exists to corroborate quality at scale (Class Central shows only 5.0 from a single review), so the catalog's headline rating should not be treated as verified
Alternatives To Consider
Frequently Asked Questions
Is Time Series free?
Yes — Time Series is free to access. Free with no paywall or audit limitation; all six lessons, exercises, and the certificate of completion are included at no cost, like all Kaggle Learn micro-courses. A free Kaggle account is required.
Who is Time Series for?
Data scientists and analysts who already write Python and pandas comfortably and want a fast, applied path to ML-based forecasting and Kaggle tabular competitions. It is ideal for people who learn by doing, want to feed into the Store Sales competition, and are interested in feature engineering (lags, Fourier seasonality) and hybrid linear+gradient-boosting models rather than statistical theory.
What will you learn in Time Series?
Model time series as a regression problem using time-step (time dummy) and lag features, and interpret the resulting linear coefficients; Capture long-term trend with moving averages and a polynomial time dummy via statsmodels' DeterministicProcess; Engineer seasonal features: seasonal indicators plus Fourier features (e.g., CalendarFourier, order-4) read off a periodogram, and add holiday indicators; Use lags and serial dependence to turn other time-dependent behavior into features for forecasting.
What are the prerequisites for Time Series?
Working knowledge of Python and pandas (DataFrames, indexing, datetime handling); Basic familiarity with scikit-learn-style models is strongly recommended (the course uses LinearRegression and XGBRegressor from lesson 1 onward without teaching ML fundamentals); completing Kaggle's Intro to Machine Learning and Feature Engineering first is the natural lead-in; Comfort reading plots (matplotlib/seaborn); no prior time-series background required.
Is Time Series worth it?
It is an excellent, free, and unusually practical introduction IF you already know Python/pandas and want the ML-feature-engineering approach to forecasting; it is a poor fit if you need classical statistical methods (ARIMA/ETS/Prophet) or deep learning, since the course intentionally omits them.
How we reviewed this course
This is an independent editorial assessment by Cursarium, based on Kaggle's published course materials and aggregated public learner feedback (last reviewed 2026-06). We have not independently completed the course. Links to providers are standard references, not paid placements.
Sources
- Official course - Learn Time Series Tutorials (Kaggle)
- Free Course: Time Series from Kaggle (Class Central - syllabus + 5.0/1-review rating)
- Course notebooks mirror confirming the 6 lessons and content (drakearch/kaggle-courses, GitHub)
- Lesson 6 notebook - MultiOutput/Direct/DirRec strategies, XGBoost, references (raw GitHub)
- Store Sales - Time Series Forecasting companion competition (Kaggle)