3 forecasts about time-series forecasting

Over the last year, we’ve seen an explosion in innovation in time-series forecasting. Along with new statistical models, transformer-based approaches have allowed for the creation of zero-shot foundation models from well-known organizations such as Google, Amazon, and Microsoft and from companies specializing in time-series forecasting such as Nixtla. The pre-trained models make time-series forecasting more accessible and available, especially to smaller organizations with limited resources. If the promises of foundation models materialize, they can revolutionize how practitioners tackle their forecasting tasks.

How will this emergence of zero-shot models impact the field of time-series forecasting? Here are three forecasts in the forecasting field.

Foundational models will slowly replace other methods for most practical use cases.

As user interfaces improve and we continue to improve both general and specialized models, these models will represent viable solutions to most forecasting tasks, replacing traditional approaches based on training task-specific models. In some tasks, such as physics-based time series, traditional approaches will likely continue to outperform foundation models, and they can and should be used in those cases.

Forecasting expertise may shift towards approaches and algorithms for evaluating accuracy and model fairness. We’ve seen this shift in the large language model (LLM) space, where organizations looking to incorporate them need to validate models and ensure they’re free from systematic bias. We’ll need to do that same work in forecasting as it becomes easier to employ forecasting models in more settings.

As with LLMs, we will see a growing family of forecasting models with different strengths and weaknesses (performance, size, speed, specialization).

At the time of writing, the number of foundation time series models is still limited but growing fast. As the field evolves, we’ll start to see more specialized models, for specific industries and purposes, to handle specific data types or to give users finer control to balance speed and accuracy, depending on what’s important in each task.

It won’t just be the models that will specialize but also the options and UI for interacting with them. The variety of options available for improving accuracy, the readily available exogenous variables, and how users access the models via code or interface will also be differentiators.

These developments will make it more challenging to conduct and share “one size fits all” benchmark analyses but will improve the access and accuracy of models in time-series forecasting. So, as we look to the future, we won’t be looking for “the one fastest and most accurate model” but conducting that analysis based on the type of question we’re asking. That means the developer of each model should clarify the tradeoffs to users and help guide them to the most appropriate models and interfaces for their data and questions.

Foundation models will democratize access to forecasting and anomaly detection, sparking a rise in the number of users.

As foundation models become more widely used, they will open the doors to time-series forecasting to more people and unlock the benefits of forecasting for more organizations. This is amazing! Broadening access to data science approaches allows more people to use data to answer the questions that are important to them.

This also means that we will be responsible for providing resources and educational opportunities that provide the context on forecasting, when it’s appropriate to use forecasting and when it isn’t, and how to think about results and use them effectively and appropriately. Importantly, we’ll need to include and incorporate the diversity of people who can use forecasting and approach our sharing and education in multiple and culturally responsive ways.

To capture this future, we’ll need more roles where communication, and not just development, are skills we prioritize. This requires training for developers to expand their skill sets or people with existing solid communication skills to learn more about forecasting. As for all things related to technology, it’s not “if you build it, they will come,” but how you share about it that broadens access. In this way, we can ensure that the future truly is for everyone.

Cristian Challu is co-founder and chief science officer at Nixtla.

—

New Tech Forum provides a venue for technology leaders—including vendors and other outside contributors—to explore and discuss emerging enterprise technology in unprecedented depth and breadth. The selection is subjective, based on our pick of the technologies we believe to be important and of greatest interest to InfoWorld readers. InfoWorld does not accept marketing collateral for publication and reserves the right to edit all contributed content. Send all inquiries to doug_dineley@foundryco.com.