How to Get Started with Time-Series Forecasting

Time-series forecasting is a powerful method that allows you to use data about the past to help predict the future. To make decisions about the future, we almost always rely on information from the past. In essence, we do this every day, even when looking at yesterday’s weather to predict what to wear today. Using time-series forecasting tools becomes particularly powerful when we have a challenge or issue that’s important to us and we’ve gathered or have access to historical data about that issue. We typically use that data to help us develop intuition about the challenge so that we can make decisions. And that intuition is powerful! However, there’s an even better way: We can use the statistical approaches in time-series forecasting and our past data to make quantitative and testable predictions about the future. 

Some real-world examples of time-series forecasting in action include:

• A small retail company using time-series forecasting to adjust its frontline retail staff cost planning to be driven by machine learning predictions and data analysis in place of aspirational operational targets. 
• A nonprofit organization working on a syndromic surveillance system to predict disease outbreaks using electronic health records to protect people from surges of new emergent diseases. They’re using time-series forecasting to enhance their ability to predict and respond to potential public health threats in South America.
• A World Bank country economist leveraging time-series forecasting to predict fishing revenues for small Pacific island states. This is a major source of government revenue in these states and is very volatile. Better insight into revenues makes the fiscal budgeting process much easier and, in turn, helps manage public investments that reduce poverty.

What Can Be Forecasted?

According to the e-book “Forecasting: Principles and Practice, the Pythonic Way,” there are four conditions that need to be met to forecast something:

1. How well we understand the factors that contribute to it
2. How much data is available
3. How similar the future is to the past
4. Whether the forecasts can affect the thing we are trying to forecast

If these four conditions are met, you can forecast. Take for example an energy company that wants to forecast demand. The factors contributing to energy demand, including weather, are well understood. There is ample data on past energy consumption, and consumption levels tend to follow past results. Therefore, energy demand would be relatively easy to forecast.

It’s also important, like with any data science analysis, that your data is well structured. That means it needs to be in a format that computers can understand. Broman and Woo’s “Data Organization in Spreadsheets” is an excellent guide on how to structure your data for any type of analysis.

How Often to Forecast

Forecasts should match your organization’s planning cycles – e.g., monthly, quarterly or annually. Teams must balance the difficulty and expense of forecasting with the benefit of having the most up-to-date forecasting data. Most companies aren’t forecasting daily, except for very limited circumstances, such as supply chain use cases like predicting just-in-time inventory levels. However, automation and new transformer models – similar to large language models (LLMs), but for data – are making forecasting much easier and less expensive and, therefore, enabling organizations to forecast more frequently.

How to Get Started

Getting started with time-series forecasting requires some preparation, understanding what you want to predict, and running and testing the models. 

The basic steps for a forecasting task are:

1. Problem definition, including determining how the forecasts will be used.
2. Gathering information, including what data needs to be considered in the model.
3. Preliminary analysis, including looking for patterns and relationships in and among the data.
4. Choosing and fitting models, from basic regression models to advanced neural networks.
5. Using and evaluating a forecasting model – only after running a forecast can you ascertain whether a model was a good fit for the use case.

Steps 1-3 are the typical things you’re already doing if you’re using your past data to try to understand the future. So, what do Steps 4 and 5 look like?

Choosing and fitting a model can be challenging – there’s a barrier to getting acquainted with the tools, statistical frameworks and model deployment needed to start time-series forecasting. Decisions on model options can vary because choosing the model is highly dependent on the data you have and the types of questions you’re asking. That’s why Steps 1-3 are so important to articulate. 

However, over the last few years, new resources and tools have been developed to help you get started. There are widely used and well-documented open-source packages in R and Python and a canonical guide for getting started with the “Forecasting: Principles and Practice” book, now with a Python version. 

If you’re just getting started, a great way to practice is using benchmark data – data for which we know the answers – so you can try different approaches and easily see how well that approach is working. Find a dataset that matches your general type of data, for the most relevant type of practice, for both the analysis and interpretation of the results. 

A faster way to get to your first prediction is using transformer-based foundation models. These are models that are pretrained on many types of data. Instead of building the model yourself, you can run your data through these models and see the predictions. Some of these foundation models include Chronos, Prophet and TimeGPT, which have various levels of ease of use. Chronos and Prophet, for instance, are open-source models, but you need to host them yourself, which still requires significant infrastructure and setup. TimeGPT is available as a service, and you can access it via Nixtla’s API. There are also R and Python software development kits (SDKs), in which you can generate a prediction in just three lines of code. There’s benchmarking data for many of these models, so you can compare them for accuracy, speed and often for your type of data. These prebuilt foundation models have been proven to work in a wide range of use cases and can save significant time, generally allowing organizations to run their first time-series forecasts within 30 minutes.

With any of these foundation models, you can also continue to try to evaluate and improve accuracy, using fine-tuning or including exogenous variables. A first approach to enhance forecast accuracy is to increase the number of fine-tuning steps. The fine-tuning process adjusts the weights within the model, allowing it to better fit your customized data. You can also include exogenous variables – external factors or predictors that are not part of the target time series but can influence its behavior. Incorporating these variables can provide the model with additional context, improving its ability to understand complex relationships and patterns in the data. Generally, to use exogenous variables, pair each point in your input time series with the corresponding external data.

Following these steps will get you on your way to predicting the future. Forecasting is a continually evolving field with new models and approaches to evaluating predictions, including model fairness, and ways of working with the models and data. The questions you have and the context you work in will help guide you to the next set of tasks you need to learn in forecasting. 

I believe, by the end of 2025, time-series forecasting will be a common practice in businesses that helps them improve data-driven decision-making to save time and money and will lead to better business outcomes. Adding this skill to your toolkit will empower you to better leverage the potential of your data and answer the questions important to your organization.