Time Series Forecasting in R: A Down-to-Earth Approach

A short definition of forecasting: what it is and where it can be used.

A good forecasting process has eight steps. Learn details about them in this lecture.

The mathematical notions we use in this course. (Don't worry, we'll keep mathematical stuff at the minimum necessary.)

In these lecture you'll learn what to look for when you examine your time series data.

Zooming into time series is a great way to identify patterns. Let's see how to do that.

How to find out which trend type is more appropriate for your series (linear, polynomial...). It's easy and very useful.

Sometimes we have to remove seasonal patterns to spot the trend better. In this lecture I will show you a simple method.

Time series forecasting models are usually built and tested in separate data sets. So let's learn how to create a training set and a test set.

Time to learn about the most important forecasting accuracy indicators. (It might surprise you, but R squared is not among them!)

Prediction accuracy is important, but we don't have to overlook error analysis. So let's learn about underpredictions and overpredictions.

The are two big classes of forecasting methods, that is...

An overview of the forecasting methods that learn patterns from data and try to use these patterns for prediction.

The most popular time series forecasting models are reviewed in this lecture.

Everything starts with a definition, so let's give the definition of moving average, an important class of smoothing techniques,

The centered moving averages - what they can and what they can NOT  do.

Only these moving averages can be used for forecasting, so pay close attention to them.

How to compute the weighted trailing moving averages and how to use them for forecasting purposes.

Visualising the centered moving averages for a series with no trend and seasonality.

Visualising the centered moving averages for a series with a dominant trend.

Using the trailing moving averages to forecast a series without trend and seasonality.

Computing the forecasting accuracy for our forecast using custom-created functions.

Forecasting a time series without trend and seasonality using the weighted moving averages.

What is simple exponential smoothing, how it's different from the moving average and how it's computed.

In this lecture we introduce a very important function than can model a great number of forecasting models. We are going to use it in many future lectures.

Forecasting a series without trend and seasonality using the simple exponential smoothing.

Let's try to improve our forecast by modifying the value of our smoothing parameter.

Until now we did only point forecasts. Now we learn how to find a confidence interval for our forecast.

Let's get to more complex stuff: the advanced exponential smoothing techniques, used for series that present trend and/or seasonality.

How the Holt model with an additive trend is built.

Now let's find out something about the Holt models with a multiplicative trend.

Using the Holt model to forecast a time series that presents a trend only. We consider an additive trend this first time.

Let's try to improve our forecasts changing the smoothing parameter in our model.

After building our model, let's see how to use them to predict future sales values.

Adding a confidence interval to our Holt model forecasts.

Let's try to forecast the same time series, with the same Holt model, but this time using a multiplicative trend. Can we improve the forecast accuracy?

Predicting the future sales values with a multiplicative trend.

Using the Holt model to forecast a time series with a dominant trend and high random variations.

Predicting the future MS stock prices using the Holt model.

Time to forecast series with both trend and seasonality: the Holt-Winters model with three smoothing parameters.

Forecasting a seasonal time series with the Holt-Winters model.

Predicting future sales values based on our model.

Building models manually, and comparing them, is always preferable. However, if you want to let the program choose the best model for you automatically, please watch this lecture.

Let's start to learn about the extended exponential smoothing, very useful when we deal with series with double seasonality.

Building a TBATS model to forecast a series with double seasonality.

Predicting the bike rental levels for the next year using our model.

Building a STLM model to forecast a series with double seasonality.

Forecasting future bike rentals (for one year in advanced) based on our STLM model.

What are regression models, in a nutshell.

How to build regression models for time series that present a trend only.

Forecasting shampoo sales using a linear trend model.

Forecasting shampoo sales using a quadratic trend model.

Forecasting shampoo sales using a cubic trend model.

Forecasting shampoo sales using an exponential trend model.

Using a regression model to forecast a time series with high random variations.

How to introduce the seasonal component in a regression model - it's very simple.

Forecasting a seasonal time series with a trend-only regression model (we ignore seasonality for the moment, to see what we get.)

Now let's introduce the seasonal component in our model. Hopefully, we'll improve our forecast accuracy a lot!

Let's try a regression model on another seasonal time series - electricity production in the US.