Complete Time Series Forecasting Bootcamp in Python (2025)

Understand the random walk as a time series with white noise, autocorrelation, and differencing to assess stationarity, then forecast using a simple simulation and Amazon price data.

Simulate a random walk in Python by generating 1000 standard normal steps, computing their cumulative sum, and visualizing with matplotlib while fixing a seed for reproducibility.

Identify whether a time series is a random walk by testing stationarity and applying differencing, then use the augmented Dickey-Fuller test to detect a unit root in Python.

Explore stationarity by simulating stationary and non-stationary time series, compare their mean and variance, and validate with the augmented Dickey-Fuller test; apply differencing to achieve stationarity.

Explore autocorrelation and the autocorrelation function (ACF) to assess the correlation of a time series with its past values, identify stationary versus random walk, and implement differencing in Python.

Plot and interpret the autocorrelation function of a stationary differenced random walk, confirm non-significant lags, and apply the workflow to Amazon's closing price for forecasting practice.

Learn how to forecast a random walk using naive baseline models, understand how forecast horizon affects performance, and implement methods in your notebook.

Forecast random walks with baselines like mean, last value, and drift, and evaluate predictions using mean absolute error on simulated data and stocks. See how one-step forecasts reveal naive-model limits.

Explore the moving average model, determine its order using stationarity checks and the ACF, and forecast with a simulated data set in Python.

Implement moving average forecasting in Python using the MA model, simulate MA processes, test stationarity with the ADF test, and perform rolling predictions to compare horizons with baselines.

Define the autoregressive model and identify its order using the ACF and pacf, forecasting with AR(p) on simulated data while linking the concepts to the moving average model.

Explore autoregressive forecasting with AR(p) in Python, using PACF to choose p, test stationarity with ADF, and forecast AR(2) series via rolling predictions, comparing to a baseline.

Learn how the arma model combines autoregressive and moving average components, determine p and q with stationary data, acf and pacf, and forecast using a reusable modeling procedure.

Design a general modeling procedure for arma time series: stationarize, fit multiple p, q combinations, pick the lowest aic, and validate with residuals via q-q plots and ljung-box tests.

Forecast with an ARMA(p,q) model via grid search guided by AIC and Ljung-Box tests; assess residuals with diagnostics and box tests, then forecast and compare to baselines.

Learn how to use the ARIMA model to forecast non-stationary data by differencing and integration, selecting p, d, q with AIC, and applying to real Python time series data.

Apply ARIMA modeling to real quarterly electricity production in Australia; determine d=1 and select ARIMA(4,1,4) via grid search, then forecast in original scale with rolling predictions.

Explore modeling seasonality with the Sarima model, using monthly milk production data to forecast with seasonal differencing, AIC-driven model selection, and Python implementation.

Forecast seasonal data with SARIMA in Python, using STL to identify seasonality, conduct stationarity tests, and perform a grid search with a last-season baseline.

Learn to extend arimax models with exogenous variables, demonstrated by forecasting Saskatchewan cattle prices using barley as an exogenous input, and discuss horizon limitations.

Apply a SARIMAX/ARIMAX approach in a Saskatchewan farm-price study, using barley as an exogenous variable, test for stationarity, tune with grid search, and evaluate with rolling predictions and MAPE.

Explore multivariate forecasting with varma and varmax to predict multiple series, using Granger causality to validate bidirectional influences, and apply this approach to cow and calf price forecasting in Python.

Apply the vector autoregression (VAR) model to multi-variable time series, perform grid search and rolling forecasts, and validate with Granger causality tests on cows and calves price data.

Apply Varma forecasting in Python via a grid search over p and q, compare with the var model, and assess residuals and forecasts for cows and calves.

Apply a varmax model with barley as exogenous input, grid-searching p and q. Evaluate residuals with diagnostic tests, compare to baseline and var/varma, and note simpler var models perform best.

Explore simple exponential smoothing, a fast forecast method for series with no trend or seasonality, using a weighted, exponentially decaying average of past data controlled by alpha.

Implement simple exponential smoothing in python using statsmodels, including training and test split, forecasting, and evaluating against a naive seasonal baseline. Demonstrate its limitations with trend and seasonality.

Double exponential smoothing extends simple smoothing with a trend component and a beta parameter, using a damping factor phi to curb overshoot in trend-only forecasts without seasonality.

Implement double exponential smoothing with Holt in Statsmodels, compare damping and without damping, extract alpha and beta, and evaluate forecasts against simple exponential smoothing.

Explore triple exponential smoothing, also known as the Holt-Winters method, which uses level, trend, seasonality, and the gamma parameter to forecast time series with additive or multiplicative seasonality.

Use triple exponential smoothing in a rolling forecast with statsmodels, specifying additive trend and additive seasonality, 12-month periods, and evaluating with mean absolute error.

this lesson covers bats and tbats for multi-seasonal time series with daily and weekly cycles, using box-cox and arma errors, with tbats employing fourier terms for non-integer periods.

Forecast with bats and tbats using sktime in Python to capture double seasonality, including daily and weekly patterns, and compare against a baseline using mean absolute percentage error.

Explore the theta model for time series forecasting, decomposing into z0 and z2 lines, removing seasonality, extrapolating with linear regression and exponential smoothing, and reintroducing seasonality in Python.

Apply the theta model from Statsmodels to weekly CO2 data in a rolling forecast, interpolate missing values, and compare theta, baseline, and triple exponential smoothing using the MAP error.

Compare theta to SARIMA in python by implementing SARIMA, testing stationarity, applying differencing, and using a grid search to optimize p,d,q and seasonal terms, then evaluate with Ljung-Box and MAP.

Explore when to use deep learning for time series and outline single step, multi step, and multi-output models; cover data readiness and Fourier transform for seasonality.

Preprocess time series data for deep learning by encoding daily and yearly seasonality with sine and cosine, applying Fourier transform, and creating train, validation, and test splits with min-max scaling.

Explore linear models in deep learning, linking input to output with no hidden layers for single-step, multi-step, and multi-output forecasts. Learn to implement these simple yet powerful models in TensorFlow.

Explore building and training linear and deep learning time series models in TensorFlow with Keras, including single-step, multi-step, and multi-output forecasting using custom time series windowing, early stopping, and evaluation.

Explore deep neural networks by adding hidden layers between input and output to model non-linear relationships using activation functions like ReLU, enabling efficient computation and strong gradient propagation.

Implement deep neural networks in a Jupyter notebook to forecast time series, building single-step, multi-step, and multi-output models, and tune with a learning rate scheduler to stabilize training.

Learn how long short term memory networks extend RNNs for time series forecasting by using cell state and gates: forget, input, and output to combat vanishing gradients, with TensorFlow implementations.

Implement and evaluate LSTM models in a Jupyter notebook using Keras, including single-step, multi-step, and multi-output architectures, with training schedulers and performance plotting.

Build and evaluate convolutional neural networks for time series forecasting using 1d CNNs, single-step and multi-step windows, kernel length six, and Keras. Explore CNN-LSTM hybrids and model evaluation.

Explore convolutional neural networks (CNNs) for time series, learning how convolution with a kernel filters noise, builds robust features, and pairs with LSTM for efficient forecasting.

Explore Prophet, an open source general additive model that handles multiple seasonal periods with Fourier terms, combining trend and holiday effects to forecast series and guide when to use it.

Use Prophet in Python out of the box to forecast time series. Prepare ds and y, fit the model, create a future frame, and compare to a baseline.

Explore advanced Prophet features to plot model components, visualize trend changes, and tune seasonality using Fourier terms, confidence intervals, and interval width in time series forecasting with Python.

Plot forecasts and components to visualize training data, predictions, confidence intervals, and trend with weekly and daily seasonality.

Analyze cross-validation and performance metrics in Prophet for hyperparameter tuning. Use rolling windows, cutoff dates, and fine-tune change point, seasonality priors, holiday priors, and seasonality mode in Python.

Explore how to use Prophet for hyperparameter tuning with cross validation and performance metrics, tuning change point prior scale and seasonality prior scale across horizon and rolling window.

Tune Prophet for weekly CO2 data with cross-validated hyperparameters, then compare with theta model and SARIMA to assess performance.