A more fluid way to model time-series data

The past is often a powerful guide to the future, whether you’re trying to predict stock prices, traffic on a network, or quarterly sales. But so much depends on the data.

Time-series data can be sliced in slivers of seconds and minutes, or in chunks spaced out over days and years — and sometimes even at different rates within the same sample. Things quickly get complicated once you move from modeling a time series captured at one scale to modeling multiple timescales at once, as foundation models are designed to do.

“The same data can look vastly different depending on the timescale and domain,” said Lars Graf, an IBM researcher focused on time-series analysis. “Traffic data has a seven-day structure because of the work week, while weather data has important daily and seasonal cycles. Skillful forecasting depends on the ability to pick out all these varying patterns.”

IBM’s new time-series foundation model, FlowState, can do just that, dynamically adjusting to the timescale of the prediction task. Its state-space model (SSM) architecture, combined with an innovative decoding mechanism, allows it to analyze a time series at one scale and output a prediction at another.

The model’s ability to seamlessly shift timescales recently landed it in second place for zero-shot forecasting models on GIFT-Eval, a leaderboard from Salesforce for time-series foundation models. The latest addition to IBM’s Granite family of time-series models, FlowState is now available in open source in Hugging Face and will be presented at a time-series modeling workshop at NeurIPS 2025 in December.

In choosing a name for the model, researchers turned to the concept of flow, that state of deep immersion during a creative endeavor when time can feel almost elastic. “It’s a metaphor for the model’s core innovation: continuous-time modeling,” said Graf. “FlowState allows flexible forecasting across arbitrary timescales. It doesn’t just process time. It flows through it.”

Small but mighty

At 9.1 million parameters, FlowState is the smallest model to make GIFT’s top 10 chart for zero-shot forecasting, outperforming rivals more than 20 times its size. It’s also currently the only state space model.

Used for decades in control systems engineering, SSMs have recently become a popular alternative to transformers for their speed at handling long sequences of data, especially text. SSMs have been rapidly making inroads into language modeling as a result; some of IBM’s forthcoming Granite 4.0 LLMs have adopted a hybrid SSM design.

SSMs, however, have been far less explored for modeling time-series data, despite their native ability to standardize and encode data at disparate scales. In analyzing their potential, IBM researchers realized that a decoder was needed to translate data encoded in a time-scale invariant space into a prediction at a specific scale.

They paired an encoder from S5, an existing SSM, with their decoding solution that applies the statistical concept of basis functions, which are the building blocks for more complex functions. The encoder converts a given time series into a timescale-invariant “hidden state,” and passes it to the decoder, which transforms it to a forecast at an arbitrary resolution.

The decoder does this by interpreting each element in the hidden state as a coefficient to a corresponding basis function, leading to a continuous forecast that can be spliced at any interval depending on the task. On a benchmark like GIFT, the timescale of the prediction usually matches that of the initial input time series. FlowState, however, can generalize to timescales it’s never seen.

It’s a key trait that differentiates the tiny SSM-based FlowState from its much larger transformer-based rivals on GIFT. Most of the transformers were either trained on data at all potential timescales or given enough weights to memorize patterns at those scales.

FlowState, by contrast, requires far fewer training examples — and weights — by converting its training data into an abstract representation. This means it potentially can be deployed at a far lower cost than the other leaders on the GIFT chart.

Toward more challenging time-series problems

FlowState’s ability to make time-scale adjusted predictions is currently geared toward tasks with a single variable — things like predicting traffic flow on city streets or whether people click on a website or buy a product.

In this way. FlowState complements IBM’s TinyTimeMixers, a collection of forecasting models that with some fine-tuning can handle multi-variable problems. For example, people rarely decide to buy a product for just one reason. Other factors like time of year and the product’s price compared to similar products may come into play.

The researchers behind FlowState are working to extend the model to complex multi-variable problems like this that more closely mimic the real world. They predict that we will be seeing more SSM-based models applied to time-series modeling in the future.

“This idea of fusing mathematical operations in continuous, abstract space holds the potential for even more powerful architectures that can take on more complex forecasting problems,” said Angeliki Pantazi, a principal research scientist at IBM who led the FlowState team.