Chapter 2: An Overview of Amazon Forecast

Time series forecasting is a very popular and widely used technique to deliver insights to a large variety of business stakeholders. Although the values predicted by a forecast can be used as they are (this is, for instance, what you do when you use spreadsheet formulas to feed financial planning practices and predict a business unit's (BU's) results to inform the appropriate business development decisions), they are usually a means to an end: forecasting then becomes an essential cog that feeds other models or business processes (forecasting product demand to optimize the production of manufactured goods).

Time series forecasting was around before machine learning (ML) and artificial intelligence (AI) became so pervasive. In this book, you are going to leverage several services built thanks to ML: these technologies are usually at their best when a massive amount of data is available or when the relationships between this data become impossible to tackle with a mere spreadsheet. When a retailer is building a demand forecast as part of their supply chain planning, the number of stock-keeping units (SKUs) and location combinations is often in the millions, making it valuable to leverage the latest AI technologies.

Amazon Forecast is the first Amazon Web Services (AWS)-managed service that we will explore in detail in this book. By the end of this chapter, you will understand how to frame a forecasting problem that is suitable for Amazon Forecast and will be ready to dive into the next chapters of of this book to develop a sound understanding of the key steps necessary to build your first forecast.

In this chapter, we're going to cover the following main topics:

• What kinds of problems can we solve with forecasting?
• What is Amazon Forecast?
• How does Amazon Forecast work?
• Choosing the right applications

Technical requirements

No hands-on experience of a language such as Python or R is necessary to follow along with the content from this chapter. However, we highly recommend that you follow along in the AWS console, from where you can access the Amazon Forecast service.

If you already have an AWS account, you can connect to the AWS console, click on the search bar at the top, and enter Forecast. In the Services section, click on Amazon Forecast to bring up the landing page of the service console.

If you don't have an AWS account, you will need to create one before you can log in to your console. To create an AWS account, proceed as follows:

1. Open https://portal.aws.amazon.com/billing/signup and follow the online instructions. You will receive a phone call to obtain a verification code for signing up.
2. Create an Identity and Access Management (IAM) administrator user: when you create an AWS account, you get a root user account that gives you complete access to all of the AWS resources in your account. It is strongly recommended to adhere to the best practice of creating individual IAM users. To create such a secure user, see the Creating your first IAM admin user and user group section of the IAM service documentation, at https://docs.aws.amazon.com/IAM/latest/UserGuide/getting-started_create-admin-group.html.

You are now ready to use Amazon Forecast!

What kinds of problems can we solve with forecasting?

Forecasting is the science of predicting the future values of a time series as accurately as possible, given all the available information (including historical values of the very time series of interest, other related time series, and knowledge of any events that can impact your forecast). After reading the previous chapter, you might understand why it can also be considered an art!

Forecasting is a regular task required by many activities: most of the time, a forecast is a means to an end and not an actual objective in itself—except for financial planning, as mentioned in the introduction, although you could argue that financial planning is also a means to ensure good business decision-making. Here are a few examples of situations where you can leverage forecasting to achieve better business outcomes:

• Forecasting future demand for a product to decide how much a plant will have to manufacture and by when. Failing to predict high demand in a given region can cause missed revenue opportunities while manufacturing too much of a product in another region will generate more inventory and increase the capital cost.
• Forecasting the sale of thousands of items in hundreds of stores so that a retailer will know how much inventory to stock in stores or warehouses to serve local clients or reduce delivery delays and costs.
• Predicting the electricity demand for households at an hourly level so that energy providers can make the cheapest electricity trading decision while maintaining the balance of the electricity grid.
• Improve workforce planning by predicting transactional quantities (rendezvous at bank tellers; customer traffic at supermarket checkouts…).
• Predicting replacement parts' demand so that planners can pre-order parts ahead of time, as part of predictive maintenance planning.
• Provide equipment failure probability, with a confidence level, based on sensor data installed on subsystems of the industrial equipment. Linked to replacement parts' prediction, the outcome of this forecast helps plant managers optimize their inventory, limiting the impact of reactive maintenance and reinforcing their predictive maintenance practice.
• Predict a company website's traffic to scale the underlying infrastructure and predict the associated cost of serving content to its audience.
• Predict memory, central processing unit (CPU), input/output (I/O) throughput, storage, and other operational loads so that cloud services such as Amazon Redshift (for hosting data warehouses) or Amazon DynamoDB (NoSQL database) can provision and scale their underlying infrastructures.

In a nutshell, the outputs of any forecasting activities are critical to ensure any operations continue to run smoothly or to take the appropriate actions. Some activities are easy to forecast, while others are just plain random walks (new values taking a random value with seemingly no link to previous ones, such as stock prices' evolution). When deciding to forecast a time series, you should assess whether the following apply:

• You have enough data available: The amount of data necessary to achieve a good forecast will depend both on the type of algorithm you leverage and on the underlying dynamic of your time series.
• You understand the contributing factors: This is how much you understand the underlying factors that impact how a time series evolves.

When using predictions in a real-life system, you should also assess how likely it is that the results of your forecasts will impact the behavior of its environment. Let's take a few examples, as follows:

• If an energy provider predicts the energy consumption for a city, this will likely have no impact on the actual electricity consumption of the said city.
• On the other hand, if you provide individual electricity consumption predictions to all households of the same city for the days to come, they might adjust their behavior to reduce their consumption (for example, turning down the heating to waste less energy and save money). This will make your energy prediction forecast less good, but you will have achieved your business objective to reduce the energy consumption of the local population.
• As a major financial institution, if you provide advice on stock market investment to your customers, you will influence this market and have an effect on the stock price quantitative analysis another department may be trying to forecast.

In other words, understanding how much your forecasts can become self-fulfilling prophecies is key to ensuring the overall dynamic stability of your predictions and the environment they are part of. In addition, understanding this helps you explain what degraded performance looks like for a forecasting model. At the early stages, providing forecasts to end users could be performed using an A/B testing approach (for example, sending an energy forecast to part of the city over a period of time): this will allow you to measure the impact your forecasts can have on the overall dynamic system and build better performance metrics.

Now that you understand the types of problems forecasting can solve, let's take some time to properly frame a given forecasting problem.

Framing your forecasting problem

As hinted in Chapter 1, An Overview of Time Series Analysis, there are many ways to frame a problem dealing with time series data, and forecasting is only one type of these problems. As its name suggests, Amazon Forecast is suitable for solving forecasting problems: problems where you want to know what will happen, when it will happen, and—possibly—why it happened.

Time series can be found in many areas and situations, and it is easy to find forecasting problems everywhere. In ML, forecasting is one of the most challenging problems to solve (after all, we are trying to predict the future!), and deciding to go down that path may prove difficult—if not impossible—depending on the actual forecasting predictive power the available data contains. Let's take two examples where forecasting is not the best path to take, as follows:

• You own a video-on-demand service, and based on past views and collected interests from your users, you would like to predict what they will likely watch next. You could structure your data to try to forecast the next views so that you can expose these shows on a personalized home page for your users. Although this approach can be rich (it does take the time dimension into consideration), it is very hard to solve efficiently as your time series will be very sparse (a few data points for each user every day at most). This kind of problem is better solved with a recommendation engine (check out Amazon Personalize for more details on how you could build a recommendation engine, for instance).
• You are operating a shampoo production plant and you want to predict the quality of any given batch in production. On your manufacturing line, you collect the product recipe, the characteristics of the raw material that are part of the chemical formulation, and the process parameters you apply with the setpoints you configure on each piece of equipment of the product line. You could use all these time series to try to predict the quality of a given batch; however, the underlying assumption is that the predictive power contained in the time series and tabular data collected is high enough to predict the appearance of the defects you are looking for.

These defects might come from the environmental conditions (hygrometry of the factory; external temperature; rainfall; raw material supplier; bad programming sequence entered by an operator). This kind of problem might be better framed as an anomaly detection problem: in this case, a service such as Amazon Lookout for Equipment (which will be our focus in Part 3, Detecting Abnormal Behavior in Multivariate Time Series with Amazon Lookout for Equipment, of this book) might be better suited to provide the insights you are looking for.

After deciding that forecasting is indeed what you want to leverage to solve your problem, there are several questions you need to address in the early stages of your forecasting project, as follows:

• What would you like to forecast?
• What is the frequency of the available data (hourly, daily, monthly…)?
• What is the desired forecast horizon?
• How frequently do you need to forecast?
• What is the expected output of the forecasting exercise?
• How will you evaluate the accuracy of your forecasts?
• How will the forecast outputs be used further down the road?

Moreover, a key task is obviously to collect, assemble, and prepare the historical data needed to build your forecasting model. You will also need to identify whether related data known at the same time can lend some statistical power to your model to improve your forecast accuracy; as time series forecasting practitioners, you might have used approaches such as Granger predictive causality tests to validate whether a given time series can be used to forecast another one.

Collecting and preparing this related data will be an additional step before you can build your model. You might also need to collect and prepare non-time-related relationships between your time series (such as hierarchies, where some locations are in the same region, or some items are in the same product groups, or some items share the same color).

The following screenshot summarizes the forecasting problem scope:

Figure 2.1 – Time series forecasting components

Let's dive into several of the key areas of this screenshot, as follows:

• Historical data: The historical data is the part of your time series that will be used to train a forecasting model. In Figure 2.1, this is the regular line plot located on the left of the graph.
• Historical data frequency: Your historical time series data can be available at a regular or irregular sampling rate. If your raw data is not regularly sampled or if its frequency is different from the forecast granularity, you may have to aggregate your data at the desired unit. In retail, each product sale is recorded when it happens, but not every product is sold every day, leading to a sparse time series with spikes that correspond to days when a sale has happened. Aggregated sales for each product at a weekly or monthly level will smooth out the original signal.
• Forecast horizon: This is the prediction length you can expect for a forecast. This is usually measured in the number of time steps using the desired unit, which might be different from the original time series. If you expect daily data as a result of your forecast and need predictions for the coming week, your sampling unit will be at the daily level and your forecast horizon will be 7 days.
• Expected output: You can request a point-in-time forecast. In the plot shown in Figure 2.1, this is the dotted line located in the forecast horizon and it helps answer this type of question: What will the value of the time series be by tomorrow? A forecast can also emit a value for several time steps at once. This helps answer questions such as: Which value will the time series take for every day of the coming week? Your forecast may also be consumed by organizations or systems that expect a range of possible values; in this case, you will replace the point values with probabilistic forecasts that emit a range of possible values at each predicted time step (in the previous plot, this is represented by the area plot around the median forecast denoted by a dotted line, delimited by the lower bound and upper bound of the requested quantiles—p10 and p90 in this case).

You now know how to frame a good forecasting problem, so let's apply this fresh knowledge to Amazon Forecast!

What is Amazon Forecast?

Amazon Forecast is one of the AI-/ML-managed services available on the AWS cloud platform. Flexible and accurate forecasts are key in many business areas to predict demand and ensure inventory or raw material are stocked accordingly or predict wait times and ensure your counters are staffed accordingly. While all organizations use some form of forecasting process today, traditional methods cannot leverage the increasing quantity and complexity of time series signals available to us.

Managed services are services where the end users only bring their data and configure some parameters to suit their needs. All other tasks, considered as undifferentiated heavy lifting, are performed on the users' behalf by the service. This includes the automation of all the infrastructure management: as an Amazon Forecast user, you don't have to provision and manage virtual machines (VMs), configure user accounts, manage security, plan for scalability if your request volume increases, decommission unused resources, and so on.

In the case of AI-/ML-managed services, some data preparation, ingestion tasks, and model management activities are also performed under the hood, allowing you to focus primarily on the problem to solve. Amazon Forecast is a scalable managed service that automates the whole end-to-end (E2E) forecasting pipeline, from data ingestion to model deployment and serving forecasts. The service also deals with data preparation under the hood. Amazon Forecast can do the following:

• Fill missing values in your datasets while letting you customize the parameters to use to replace them.
• Prepare the time series so that it can be ingested by multiple algorithms—Amazon Forecast can prepare time series sequences to feed deep learning (DL) models or generate multiple random sequences to output probabilistic forecasts for proprietary algorithms such as non-parametric time series (NPTS).
• Pick the best algorithms depending on your datasets—Amazon Forecast can use AutoML to automatically choose between statistical forecasting algorithms (such as AutoRegressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS), NPTS, or Prophet) and more advanced data-hungry neural network-based architectures (such as DeepAR+ or Convolutional Neural Network-Quantile Regression (CNN-QR) deep neural networks (DNNs)). You will have an overview of the different algorithms used by Amazon Forecast, along with the advantages and disadvantages of using each of them, in Chapter 5, Customizing Your Predictor Training.
• Segment your data and automatically apply the best ensemble of algorithms for each segment.
• Provide confidence intervals thanks to probabilistic forecast generation—traditional algorithms only output point forecasts (usually the mean value) with a confidence interval, relying on strong assumptions about the data distribution.
• Provide explainability scores to help you understand which factors influence the forecast of specific items and for which time duration.
• Manage multiple local statistical models for each of your time series—using ARIMA, ETS, NPTS, or Prophet to train on 100 time series means that you would have to manage 100 models on your own. Amazon Forecast transparently manages all these models for you and lets you easily export all forecasts at once as a single comma-separated values (CSV) file after inference as if it were just one model.
• Perform model evaluation on multiple time windows and provide evaluation results for all your time series data at once in a single CSV file digestible by the usual spreadsheet or business intelligence (BI) visualization software. Amazon Forecast uses backtesting techniques on your behalf (you will deep dive into this feature and how to customize it to suit your needs in Chapter 5, Customizing Your Predictor Training).
• Automatically compute multiple accuracy metrics when evaluating and selecting the best models, including traditional point forecast metrics such as Weighted Absolute Percentage Error (WAPE) and unbiased error metrics for probabilistic forecasts (namely, Weighted Quantile Loss (wQL)). You will discover the details behind these accuracy metrics in Chapter 4, Training a Predictor with AutoML.
• Perform hyperparameter optimization to find the best parameters to train your models.
• Add built-in features such as public holidays or weather data to improve your forecast accuracy. When relevant, Amazon Forecast also feeds features such as lagged values to capture seasonality in your datasets. For instance, if your samples are taken at an hourly frequency, Amazon Forecast will feed values from the last 1, 2, and 3 days to algorithms such as DeepAR+ that can benefit from these additional insights to build more accurate models.

Based on the same technology used at Amazon.com, Amazon Forecast drills down the extensive experience Amazon has on forecasting techniques to make it available to every developer and data scientist. It's a fully managed service that uses statistical and DL-based algorithms to deliver accurate time series forecasts. Although an ML practitioner or a data scientist will be able to leverage their skills to improve the service results and make the best of it, Amazon Forecast does not require any ML experience to deliver future time series predictions based on your historical data.

Amazon Forecast also includes the latest advances in DL to provide DL algorithms (namely DeepAR+ and CNN-QR) that are able to learn more complex models and serve situations where you might need to build forecasts for millions of time series, address cold start situations, or easily build what-if scenarios. Amazon Forecast is also able to automatically segment your data and build an ensemble of models, increasing forecast accuracy with no additional heavy analysis at your end.

Up to now, you have seen how to frame a forecasting problem and how the Amazon Forecast service can help you solve this. You will now see how Amazon Forecast works, which will give you the keys to successfully apply this service to your business problems.

How does Amazon Forecast work?

To build forecasting models, Amazon Forecast deals with the following concepts and resources:

• Datasets: A dataset is a container to host your data. Amazon Forecast algorithms use these datasets to train its models. Each dataset is defined by a schema whereby you can define the columns and their types. Amazon Forecast includes several domains with predefined schemas to help you get started.
• Dataset groups: Amazon Forecast algorithms can leverage several datasets to train a model. A dataset group is a container that packages a group of datasets used together to train a forecasting model.
• Featurization: The featurization configuration lets you specify parameters to transform the data. This is where you specify a null-value filling strategy to apply to the different variables of your dataset.
• Algorithms: Amazon Forecast has access to multiple algorithms including statistical algorithms (ARIMA, ETS, Prophet, and NPTS) and DL algorithms (DeepAR+ and CNN-QR). Algorithm choice can be abstracted away from the user when choosing AutoML at training time while giving the more advanced practitioner the ability to select and configure the algorithm of their choice.
• Training parameters: Some algorithms can be further customized by providing specific hyperparameter values. Another common training parameter to all algorithms is the forecast horizon.
• Recipes: A recipe is a combination of several ingredients: an algorithm, a featurization configuration, and evaluation parameters
• Predictors: This is a trained model. To create a predictor, you provide a dataset group and a forecasting recipe (which includes an algorithm, although you can let Amazon Forecast choose the algorithm on your behalf and work in AutoML mode). The AutoPredictor feature lets Amazon Forecast build an ensemble of algorithms to better adapt to the different segments your data may enclose.
• Metrics: Amazon Forecast uses several accuracy metrics to evaluate the predictors built into the service and help you choose which one to use to generate forecasts that will serve your business purpose. These metrics are evaluated through a backtesting technique that you can configure to split your data between training and testing parts.
• Forecasts: You can run inference (that is, generate new predictions) to generate forecasts based on an existing predictor. Amazon Forecast outputs probabilistic forecasts instead of point forecasts of mean values—for instance, Amazon Forecast can tell you if there is a 10% chance that the actual values of your time series are below the values predicted by the model. By default, forecasts will be generated at 10%, 50%, and 90% probabilities, giving you more insights for sound decision-making.
• Explainability insights: In ML, explainability allows you to better understand how the different attributes in your dataset impact your predictions. With Amazon Forecast, you can request forecast explainability insights—this feature will be exposed both in a dashboard and as a downloadable report, and will help you understand how the different features of your datasets impact forecast results for individual time series and at which points in time.

Let's now have a look at how these concepts are integrated into an Amazon Forecast typical workflow.

Amazon Forecast workflow overview

Any forecasting process involves two steps, as outlined here:

1. We start by looking backward in the historical data to establish a baseline and uncover trends that may continue in the future.
2. We then use these trends to predict the future values of your time series data.

As illustrated in the following diagram, Amazon Forecast provides an approach to tackle these two steps:

Figure 2.2 – Amazon Forecast overview

Let's dive into this diagram to explain the different steps you will have to go through to use Amazon Forecast, as follows:

• Uploading and ingesting data into the service—this will be thoroughly detailed in the next chapter. Amazon Forecast can use data that is located in Amazon Simple Storage Service (S3), a scalable data storage infrastructure used to store any kind of object and dataset. At ingestion time, Amazon Forecast inspects your data and identifies valuable features that will be used at training time.
• Training a predictor, as described in Chapter 4, Training a Predictor with AutoML. Depending on the performance obtained, you can iterate over the data ingestion and training steps a few times to achieve the desired performance. Under the hood, Amazon Forecast can train multiple models using a variety of algorithms, select the best parameters for each of the models, optimize them, and select the most accurate one.
• Once you have a predictor that you are happy with, you can start generating new forecasts that will also be time series (although they may contain only one data point, depending on the prediction length you ask for). When you ask Amazon Forecast to generate a forecast, it will actually deploy and host your model so that you can export all predictions at once.
• If you enable the AutoPredictor mode, Amazon Forecast will also allow you to generate explainability insights that you will be able to visualize in the console user interface (UI). You will also have the opportunity to download a detailed report in CSV format.

A good understanding of the Amazon Forecast workflow would not be complete without a presentation of the key drivers of the pricing model of this service. In the next section, you will match the different steps of this workflow with the different pricing dimensions of Amazon Forecast.

Pricing

As with many AWS services, you only pay for what you use, with no upfront commitment. Although the cost of the service is minimal to build a proof of concept (PoC)—especially if you can benefit from the Free Tier—there are four dimensions to consider for operationalizing a forecasting pipeline with Amazon Forecast, as outlined here:

• Storage: The data is ingested and prepared to ensure the fastest training time by Amazon Forecast. Storage is priced for each gigabyte (GB) ingested in the service.
• Training hours: Each time you train a new custom model based on your data, you are billed for the number of hours of training.
• Generated forecast data points: You are billed for each unit of 1,000 forecast data points generated by the service. A forecast data point is a combination of the number of unique time series, the number of quantiles, and the time points within the forecast horizon.
• Forecast explanations: You are also billed for any explanation generated by the service to explain the impact of different attributes or related variables of your dataset on your forecasts.

If this is the first time you are using Amazon Forecast with any given account, you have access to a Free Tier, which will allow you to use the service for free for 2 months. During this period, you will not be charged if you use the following:

• Storage: Less than 10 GB per month
• Training hours: Fewer than 10 hours per month
• Generated forecast data points: Fewer than 100,000 forecast data points per month

AWS service developers work relentlessly to reduce the operational costs of services, and price reductions happen regularly. At the time this book was written, these components were priced as follows:

• Storage: $0.088 per GB.
• Training hours: $0.24 per hour.
• Generated forecasts: Starts at $2.00 per 1,000 forecasts for the first 100,000 data points. Pricing then decreases according to a tiered pricing table.

For the most up-to-date pricing, you can check the Amazon Forecast pricing page:

https://aws.amazon.com/forecast/pricing/

Up to now, you have read about how to find and frame a proper forecasting problem. You have also learned how Amazon Forecast works. You are now ready to dive into the choices you will need to make to select applications that will benefit the most from Amazon Forecast's capabilities.

Choosing the right applications

You have successfully framed your ML project as a forecasting problem and you have collected some historical time series datasets. Is Amazon Forecast a good candidate to deliver the desired insights? Let's review some considerations that will help you understand whether Amazon Forecast is suitable for your forecasting scenario—namely, the following:

• Latency requirements
• Dataset requirements
• Use-case requirements

All in all, reading through these requirements will help you define when Amazon Forecast is best suited for a given problem. You will also understand when it is not likely to be a good candidate for this. You will have some pointers on what you need to do to adjust your problem framework so that it matches Amazon Forecast's native capabilities.

Latency requirements

With Amazon Forecast, the training must happen in the cloud—if your data is not available in cloud storage such as Amazon S3, the first step will be to transfer it there.

At prediction time, the inference will also happen in the cloud—you will need to send your freshest data to the cloud, and a trained model (the predictor) will be generated and also stored in the cloud. As inference happens in the cloud, you will depend on network latency between your local systems and the internet. If you need forecasting predictions in near real time to feed optimization algorithms in a factory, Amazon Forecast will likely be the wrong candidate and you should explore building a custom forecasting model that you will deploy at the edge (for instance, leveraging Amazon SageMaker and its Edge Manager feature to compile, deploy, and manage ML models on local machines).

Dataset requirements

Although Amazon Forecast includes some statistical methods that have been successfully used on small datasets, large datasets with dense time series are likely more valuable to deliver a business return on investment (ROI) perspective.

The underlying DL algorithms included in the services start to shine when you have at least several hundred time series (the service can scale to millions of distinct time series) and a long historical period to train on (ideally, more than 1,000 data points per time series).

The more advanced algorithms leveraged by Amazon Forecast are also able to capture relationships between multiple related time series and shared non-time related attributes.

Depending on the available length from your time series data, you will also be limited in the maximum forecast horizon you can actually request. This horizon will be the lesser of 500 time steps or one-third of the dataset length. Here are a couple of examples to illustrate this:

• If your historical time series contains 1,000 days of data at a daily granularity (1,000 time steps), your maximum forecast horizon will be 333 days.
• If your historical time series contains 1,000 days of data but at an hourly granularity (24 x 1,000 = 24,000 time steps), your maximum forecast horizon will be 500 hours (if you decide to keep the same sampling rate of 1 hour per time step).

Amazon Forecast is a univariate forecasting service—the target dataset that contains the metrics of interest can only include a single field to forecast for. If you have a multivariate dataset and want to deliver predictions for multiple fields in parallel, you would have to build several individual models with Amazon Forecast. Depending on your use case, this may not be suitable, as dependency between different variables may contain valuable insights to forecast future values for all these variables.

Last but not least, your data must have a compatible time interval—although Amazon Forecast can deal with datasets that have a time interval ranging from 1 minute to 1 year, the more common intervals are hours, days, and weeks. High-frequency datasets (for example, sensor data collected at a 10-millisecond sampling rate) will require you to aggregate the data at a minimum of 1 minute. If you are looking at predictive fast events, this is not a suitable approach (I would argue that a forecasting formulation might not be the right one). On the other end of the spectrum, long periods of time—such as decades—in environmental science will lead to a sparsity of your dataset, which may generate a low-quality forecast.

Use-case requirements

Throughout this chapter, you have read through multiple forecasting examples. You have also seen how much heavy lifting Amazon Forecast performs on your behalf. There are, however, situations and use cases for which Amazon Forecast is not the best match, such as the following:

• Small data: Use cases similar to financial planning, where you need to forecast monthly revenue for a company for the next 5 years based on the last 3 years (36 data points).
• Long periods of time with sparse data: This generally yields small data.
• Random walk: Random processes such as stock market indices where future values are heavily impacted by external factors with low (to no) predictive power in past historical data alone.
• High-frequency data: Datasets with a sampling rate of less than 1 minute or even less than 1 second are not suitable for Amazon Forecast.
• Multivariate data: Although Amazon Forecast can forecast univariate data while leveraging insights coming from related time series data, the service is not built to provide multivariate forecasts in parallel for several interdependent time series.

Summary

Amazon Forecast is an AI-/ML-managed service running in the cloud and provides a one-stop shop to solve many forecasting problems.

In this chapter, you learned how to frame your problem as a forecasting one and discovered the key concepts manipulated by the Amazon Forecast service (for instance, datasets, predictors, forecasts, and explainability insights).

This chapter was important to help you develop a good understanding of which applications will be good candidates to be solved with Amazon Forecast. You also read about how you can improve your forecasting problem-framing approach, should you want to benefit the most from all the undifferentiated heavy lifting that this service can deliver on your behalf.

In the next chapter, we are going to create our first forecasting project by creating a dataset group and ingesting some time series data.