Quick start

If you're new to Chronos, welcome! This guide will interactively walk you through all the steps needed to start forecating on the platform. We'll focus on using the Precision forecasting interface but a lot of the same concepts we share here will apply to other areas of the platform.

If you're new to forecasting as a whole, we have a lot of introductory content to explore that teaches the basics of working with time series. We also have sections on how to think about time series forecasting in a business context so that users can make the most of the forecasts they generate.

Precision Forecasting

Follow along the interactive guide to learn how to start forecasting on the platform. We'll walk through all the steps needed, from data sourcing and upload, horizon and algorithm selction, and ending with making a forecasts. It's okay if any of these concepts are new, we'll step through each section of the process with explainations to go with.

Let's get into it.

Before we can make a forecast, we need some data to work with. If you already have a series you want to forecast, great, otherwise FRED is a great resource for sample time series.

You will also need to have an account before you can start forecasting. You can get started for free. Create an account

When uploading your own data, we currently support .csv and .xlsx files where the headers contain a date column named date, and any number of columns named after the respective series. Date values should be in the format yyyy-mm-dd for daily data, yyyy-mm for monthly data, or yyyy for yearly data. Learn more about the ISO 8601 international date format standard we follow if you're interested.

You can also use the sample data file below to get started. Click here to download

You can either drag-and-drop the file into the dotted area or browse your file system to upload. If the data is in the correct format and Chronos can parse the data, you should see a preview of the data. It may only show a subset of the data but you should see your data set and some properties about it like how many observations there are and the date range it spans.

In order to upload the data, you need to select which is your target variable and click the upload button. The target variable is the time series that you want to generate a forecast for. In other apps like Batch forecasting or Scenario-based forecasting, Chronos accepts the upload of additional time series. These use cases work a little differently and require data in different formats, but for now we'll focus on working with a single time series.

When using one of the platform's app like Precision or Batch forecasting, you will see the data you've uploaded take the place of the Upload interface. In this guide, we've broken things down so that we can provide explainations in between each step.

Now that we have data on the platform, we begin the process of defining how we want the forecast to behave. First, this involves choosing a horizon. The horizon is how far from the last point of the uploaded data the forecasting algorithm should predict. When your data end relatively close to the present date, the horizon usually represents a forecast into the future. This isn't a requirement and there are cases where you may want to predict data you already have the true values for, however in general future predictions are the most common use case.

As you drag the horizon slider (or adjust from the number input), you should how the blank space to the right of your uploaded series changes in size. This is the horizon into which the forecasting algorithm will generate its forecasts for.

With the time series visualization element, you can also use that to get a better look at your data. Scrolling with the mouse hovered over the element will let you zoom in and out and a left-click and drag will pan the view.

Choosing how large to make your horizon can depend on a few things. First, you want to make sure that you've uploaded enough observed data so that the algorithm has enough history to learn from properly. If you want to predict one year into the future, you should have multiple years of history uploaded so that the patterns taking place over that time period can be learned from and used to forecast into the future. Second depends on how you plan to use the forecast and the algorithm you've chosen to forecast with. The next section will explore this idea further.

Chronos Precision forecasting currently supports three different forecasting libraries that each have their own strengths and weaknesses. The following section will give a brief breakdown of each of them and how to start using one in your forecasting.

• Prophet: An open source project originally developed at Facebook (now Meta) that generates forecasts using an additive model where non-linear trends are fit with various seasonalities, plus holiday effects.
  When to use: with time series that have strong seasonal effects and several seasons of historical data. It is also effective when there is missing data or shifts in the trend, and typically handles outliers well.
• autoARIMA: An extension of the popular statistical method ARIMA, autoARIMA automates the parameter selection process with something akin to a grid search. It will automatically select the model that minimizes some information criterion.
  When to use: often best used as a baseline when beginning a search for the best algorithm to forecast your data with, autoARIMA does well on time series with just a single seasonality.
• GluonTS: A toolkit for probabilistic time series modeling, focusing on deep learning-based models. Chronos exposes a number of models, including DeepAR.
  When to use: best used where there may be non-linear patterns in the data that a more generic statistical process may not be able to learn.

With these three, users can generate a wide range of forecasts, each preforming best when given the right kind of data. Knowing which algorithm to choose is as much a skill as it is an art. Chronos makes it easy to try them all and them compare metrics between them to find the best suited for the data.

Using the dropdown selection above, you can change which algorithm to use when forecast. Choosing an algorithm will then give users access to a number of parameters that they can adjust to change the way the algorithm behaves.

Each algorithm will have different parameters that users can adjust, however without any adjustments, the default parameters for each should generate reasonable forecasts that can serve as a baseline before continuing experimentation with other algorithms or adjusted parameters.

We now have everything we need to make our first forecast. We've uploaded some data, chosen a horizon and forecasting algorithm, and now we're ready to send it all off to Chronos to handle forecasting.

In other guides, we get more into the details of model selection, using external data, additional parameters available only for specific algorithms, and more. For now, we have everything we need for a quick start. Hit the forecast button below when you're ready.

After a minute or two, you should see your forecast results as a blue line in the section above. This blue line is the forecasting model's output. Some models will just give you a forecast, or values forward of your uploaded data, while others will give you predictions across the entire historical and forecasting period.

Models that predict in the historical period are particularly helpful because you can get a sense of how well the model "fit" your data. The closer the blue prediciton line is to the actual values the more likely it is that the model will be able to make good predicitons.

Despite our best efforts as humans to "eyeball" results, the best way to see how well a model performs on a dataset is to look at the error metrics. Error metrics are calculations done that consider the model's predictions and the true values observed. There are many different kinds of error metrics and each one of them tells the story of model accuracy a little differently.

Ideally, each of these error values is as close to zero as possible. Zero values for an error metric means that the predicitons are exactly the same as the true values. In the real world, this is almost never likely to happen. Nonetheless, that doesn't mean we can't do our best to get as close as possible.

Further documentation will go into greater depth about how these error metrics are calculated and what each of them mean and what they can tell us about a forecast.

If you've followed this guide closely, you should have been able to generate a forecast on some data you uploaded. In the guide, we break down the steps that you will usually need to go through be able to effectively use the platform. When you actually start using one of the apps, like Precision or Scenario-based forecasting, you will see a different user interface. Still, the basic elements shown here will be the same and should be familiar when you switch over to that app.

Thank you for following along and thank you for choosing to forecast with Chronos. Reach out to the Chronos team here or send us a DM on Twitter for any questions or comments you might have.