ASTEP gives language models a vocabulary for steering time-series forecasts

Time-series foundation models are running into a familiar maturity problem: benchmark scores keep moving, but it is increasingly difficult to tell whether the architecture, the training data or the recipe deserves the credit. A joint project from Rensselaer Polytechnic Institute and IBM Research, described in Red Hat Research Quarterly, proposes a more useful target: make forecasts steerable by language.

What changed

The project developed two related models. PatchTST-FM uses a conventional encoder-only Transformer with patch tokenization, missing-value normalization and a quantile output head. With public data, contiguous masking and a long context window, the team says this deliberately ordinary backbone matched or exceeded specialized time-series models on GIFT-Eval and TIME. The result is less a victory lap than a warning: leaderboard gains attributed to novel architectures may instead come from data and training differences.

ASTEP, short for Abstract Semantic Tokenizer-Predictor, tackles a different limitation. Conventional time-series models expose continuous numerical patches, which give a person or another model no clean way to express a high-level expectation. ASTEP converts a series into discrete tokens describing recurring temporal shapes at several scales. Coarse tokens can represent a gradual rise or plateau; finer tokens fill in shorter fluctuations.

That hierarchy matters. The team reports that multi-scale tokenization outperformed a single-scale version and that ASTEP's larger variant led its zero-shot GIFT-Eval comparison. More importantly, a user or language model can specify the coarse pattern and leave ASTEP to generate a probabilistic forecast consistent with it. The article says this intervention helps most at long horizons, where unconstrained forecasts diverge.

Who should care

The immediate audience is teams building forecasting agents for planning, monitoring and control. The broader project, called the Time-Series Data Agent, couples a language model to a forecasting model so a request such as “expect a slow rise, then a plateau” can become part of the forecast rather than an instruction applied after the fact.

The work also matters to model evaluators. PatchTST-FM is an argument for controlling data and training variables before crediting architectural novelty. ASTEP is an argument for measuring capabilities beyond zero-shot accuracy, including whether a model can accept structured human or machine guidance.

What to try

Practitioners should treat ASTEP as a research architecture, not a product announcement. The useful test is architectural: compare a standard forecasting baseline under a documented training recipe, then ask whether intervention improves decisions at the horizons that matter. If a forecasting agent cannot translate contextual reasoning into a constraint on the forecast, a higher leaderboard score may not solve the operational problem.