Guide For Time Series Forecast Projects

A time series forecast process is a set of steps or a recipe that leads you from defining your problem through to the outcome of having a time series forecast model or set of predictions.

In this lesson, you will discover time series forecast processes that you can use to guide you through your forecast project. After reading this lesson, you will know:

• The 5-Step forecasting task by Hyndman and Athanasopoulos to guide you from problem definition to using and evaluating your forecast model.
• The iterative forecast development process by Shmueli and Lichtendahl to guide you from defining your goal to implementing forecasts.
• Suggestions and tips for working through your own time series forecasting project.

A. 5-Step Forecasting Task

The 5 basic steps in a forecasting task are:

1. Problem Definition
2. Gathering Information
3. Preliminary Exploratory Analysis
4. Choosing and Fitting Models
5. Using and Evaluating a Forecasting Model

5-Step Forecasting Task

B. Iterative Forecast Development Process

Their process can be summarized as follows:

• Define Goal.
• Get Data.
• Explore and Visualize Series.
• Pre-Process Data.
• Partition Series.
• Apply Forecasting Method/s.
• Evaluate and Compare Performance.
• Implement Forecasts/Systems

Iterative Forecast Development Process

Explore and Visualize Series => Get Data. Data exploration can lead to questions that require access to new data.

Evaluate and Compare Performance => Apply Forecasting Method/s. The evaluation of models may raise questions or ideas for new methods or new method configurations to try.

C. Suggestions and Tips

This section lists 10 suggestions and tips to consider when working through your time series forecasting project. 

1. Select or devise a time series forecast process that is tailored to your project, tools, team, and level of expertise.
2. Write down all assumptions and questions you have during analysis and forecasting work, then revisit them later and seek to answer them with small experiments on historical data.
3. Review a large number of plots of your data at different time scales, zooms, and transforms of observations in an effort to help make exploitable structures present in the data obvious to you.
4. Develop a robust test harness for evaluating models using a meaningful performance measure and a reliable test strategy, such as walk-forward validation (rolling forecast).
5. Start with simple naive forecast models to provide a baseline of performance for more sophisticated methods to improve upon.
6. Create a large number of perspectives or views on your time series data, including a suite of automated transforms, and evaluate each with one or a suite of models in order to help automatically discover non-intuitive representations and model combinations that result in good predictions for your problem.
7. Try a suite of models of differing types on your problem, from simple to more advanced approaches.
8. Try a suite of configurations for a given problem, including configurations that have worked well on other problems.
9. Try automated hyperparameter optimization methods for models to flush out a suite of well-performing models as well as non-intuitive model configurations that you would not have tried manually.
10. Devise automated tests of performance and skill for ongoing predictions to help to automatically determine if and when a model has become stale and requires review or retraining.