Chronos-2: A foundation model for multivariate time series forecasting

AI summary

Overview

Research area: Machine learning — time series foundation models (zero-shot probabilistic forecasting with multivariate structure and covariates), with applications across epidemiology, macroeconomics, earth science, and hydrology.

Technical level: Intermediate. The core ideas (attention, patching, quantile regression, in-context learning) are explained conceptually, but familiarity with forecasting metrics (MASE, WQL, WIS, NSE) helps.

Scope: The paper introduces Chronos-2, a 120M-parameter pretrained Transformer that forecasts collections of related time series and their covariates in a single zero-shot forward pass, and evaluates it on general benchmarks plus four domain-specific scientific case studies.

What This Paper Is About

Most forecasting models are built and tuned separately for each dataset, and most existing time series foundation models treat every series in isolation, so they cannot exploit relationships between co-evolving variables or external drivers. Chronos-2 is a single pretrained model that infers those relationships directly from the historical context at inference time — no task-specific training — and produces probabilistic (quantile) forecasts for univariate and multivariate problems alike. The goal is to move forecasting from per-dataset model fitting toward a generalist model that transfers across scientific and industrial domains.

Key Contributions

1. A unified multivariate architecture with group attention. Chronos-2 alternates time attention (aggregating across patches within one series) with group attention (aggregating across all series in a group at a shared patch index), letting one model handle univariate series, multivariate variates, and targets-with-covariates by simply changing group assignments at inference time.

2. A scalable synthetic multivariate pretraining pipeline. Univariate signals from KernelSynth, Trend-Seasonality-Irregularity (TSI), autoregressive (AR), and exponential smoothing (ETS) generators are coupled via instantaneous transformations (random dense/sparse/sign/aggregation matrices, random MLPs) and sequential transformations (temporal causal models simulating VAR, random RNNs) to expose the model to linear and non-linear, lagged and contemporaneous dependencies.

3. State-of-the-art zero-shot accuracy with large efficiency gains. Across benchmarks, the model improves skill scores by 5.2 percentage points and win rates by 10 percentage points relative to the strongest baselines, while forecasting 100 time series in 0.8 seconds on a single mid-range NVIDIA A10G GPU — 50× and 1000× faster end-to-end than the best task-specific point (CatBoost) and probabilistic (TFT) forecasters.

4. Domain case studies showing parity with specialists. Zero-shot Chronos-2 is evaluated under each field's own protocols against operational ensembles and supervised baselines in infectious disease forecasting, German CPI forecasting, FLUXNET carbon/energy flux prediction, and CAMELS-DE rainfall-runoff simulation.

Main Findings

• Best average accuracy on fev-bench (100 tasks, 68 with multivariate dependencies). Chronos-2 gains an absolute MASE (SQL) win-rate increment of 41.6 (46.9) percentage points over the best statistical model (Statistical Ensemble), 33.9 (55.1) over the best tree-based method (CatBoost), 39.9 (37.6) over the leading deep learning model (TFT), and 9 (7.2) over the strongest foundation-model baseline (TiRex). Wilcoxon signed-rank tests give p < 10⁻⁶ against all baselines.

• Skill-score improvements. On MASE (SQL) skill scores, Chronos-2 improves by 11.8 (15.1) percentage points over the best task-specific model (CatBoost/TFT) and 5.2 (4.7) over the strongest foundation models (TimesFM-2.5/TiRex).

• Group attention (in-context learning) is what drives the gain. Enabling group attention on fev-bench yields absolute improvements of 10.7 percentage points in win rate and 4.1 in skill score versus univariate inference. Gains also appear on the primarily univariate Chronos-Bench-II (n=27) and GIFT-Eval (n=97), and are especially pronounced on Chronos-Bench-II, which focuses on short-history tasks — suggesting the model "cross-learns" from related series when history is scarce.

• Most reliable single model, not just best average. Chronos-2 has the lowest MASE (SQL) improvability at 4.5% (3.6%), versus 10.5% (10.4%) for the next best model (TimesFM-2.5/TiRex) — even where it is not the top performer, the gap to the best available model is small.

• Best-calibrated probabilistic forecasts. On 90% [0.05, 0.95], 95% [0.025, 0.975], and 98% [0.01, 0.99] intervals, Chronos-2 achieves the best calibration on 23 out of 24 metric-interval pairs across fev-bench, Chronos-Bench-II, and GIFT-Eval, measured by Weighted Interval Score and Mean Absolute Coverage Error.

• Covariates sharpen and disambiguate forecasts (qualitative). In retail data, univariate inference misses promotion- and holiday-driven fluctuations that the covariate-aware model tracks; in a store-closure regime change, the "Open" covariate lets the model keep tight intervals during zero-sales periods and anticipate reopening, where a univariate model inflates uncertainty.

• The model behaves like a simulator under hypothetical futures. Feeding simulated promotion schedules moves forecasts between a baseline lower bound (no promo) and an upper bound (high promo); an always-on promotion schedule — never seen in history — yields high predicted sales with markedly widened uncertainty bands. Simulated customer-count trajectories (stable, 20% increased, 20% decreased) produce sales forecasts that track the simulated footfall.

• Epidemiology. Against the operational RespiCast Hub-Ensemble over 67 rolling evaluation windows (Oct 2024–Jan 2026) across 31 European countries, Chronos-2 is competitive at 1-week horizons and significantly outperforms the ensemble at horizons > 1 week, with in-context learning adding further gains by exploiting cross-country correlations (p < 0.001).

• Macroeconomics. On 110 German CPI product groups (Level 4 COICOP), univariate Chronos-2 is comparable to AutoARIMA, AutoETS, and DynOptTheta; joint forecasting within Level 2 COICOP groups outperforms AutoARIMA and AutoETS on the majority of item-window pairs at both 3-month and 12-month horizons (p < 0.001).

• Carbon flux. On FLUXNET daily NEE, LE, and H across 65 sites for June–August 2020, zero-shot Chronos-2 with covariates has the lowest site-level RMSE distributions versus pooled and site-specific supervised XGBoost (p < 0.01), and outperforms its own univariate mode.

• Hydrology. On 1,582 CAMELS-DE catchments, zero-shot Chronos-2 with covariates reaches a median Nash–Sutcliffe Efficiency of 0.68 versus 0.15 for univariate inference; the domain-specific LSTM ensemble reaches 0.90. (The provided text is truncated here, so the fine-tuning result and any hierarchical-forecasting results are not reported in the available content.)

Methodology in Plain English

The model is built on a small encoder-only Transformer (120M parameters, derived from T5) that never generates step by step. Instead, each time series is normalized using its own historical statistics, given extra "meta features" marking the time index and whether each value is missing, and chopped into non-overlapping patches that are embedded into vectors. Because future values of some covariates (holidays, promotions, weather forecasts) may be known, both the history and the forecast horizon are fed in, with unknown future values zero-filled and flagged by a binary mask — a trick that also handles missing data.

The distinctive ingredient is how information flows. In each Transformer block, a time attention layer mixes patches within a series, then a group attention layer mixes across all series in a group at the same patch index. A "group" is a deliberately flexible notion of relatedness: one series, several related series, variates of a multivariate series, or targets plus covariates. Changing group assignments switches the same weights between univariate and multivariate behavior, so dependencies are inferred from context rather than learned per task.

Training data supplies the multivariate experience that real datasets lack. The authors first sample diverse univariate series from Gaussian-process kernels (KernelSynth), stochastic trend/seasonality/irregularity combinations, AR and ETS processes, then couple them — either instantaneously, via random linear operators or random MLPs applied at each time step, or with time lags, via randomly sampled causal graphs resembling VAR systems and random RNNs. The model is trained with a quantile-regression objective through a quantile head that outputs a set of quantiles, including extreme 0.01 and 0.99 quantiles, rather than a single point or a fixed parametric distribution.

Evaluation uses MASE for point forecasts and Scaled/Weighted Quantile Loss for probabilistic ones, reported as win rate (share of pairwise comparisons won), skill score (average percentage improvement over Seasonal Naive), and improvability (gap to the best hindsight model per task), against statistical ensembles, tree-based methods (LightGBM, CatBoost), task-specific deep models (DeepAR, TFT, PatchTST), and foundation models including TiRex, TimesFM-2.5, Toto-1.0, Moirai-1.1, Moirai-2.0, Granite-TTM-R2, TimePFN, and TabPFN-TS. All foundation models are used zero-shot.

Why This Matters

Impact on research. The results argue that a single pretrained model can capture structure shared across temporal systems, shifting the field's emphasis from dataset-specific fitting to contextual inference over interacting series. The group-attention mechanism gives foundation models a way to represent multivariate dependence and covariates — a capability the paper notes no prior foundation model had demonstrated zero-shot — and the synthetic pipeline offers a reusable recipe for generating multivariate pretraining data where real multivariate corpora are scarce.

Real-world applications:

• Public health: rapidly deployable probabilistic forecasts of acute respiratory infection and influenza-like illness that can supplement operational ensembles like RespiCast, particularly where specialist modeling teams are scarce or slow to calibrate.
• Macroeconomics and monetary policy: joint forecasting of granular CPI components, not just headline inflation.
• Earth science and agriculture: forward-looking ecosystem carbon and energy flux forecasts from meteorological drivers, useful for irrigation optimization and monitoring vegetation stress under increasing aridity.
• Hydrology and energy: rainfall-runoff simulation from meteorological covariates, and electricity load forecasting where temperature, dew point, and pressure data are available in full, partial, or not at all.

Industry relevance. The 120M-parameter footprint runs on consumer-grade GPUs or standard CPU hardware without specialized accelerators, and 0.8-second forecasts for 100 series mean no per-dataset training or tuning cost. Scenario analysis with hypothetical future covariates turns the same model into a planning and simulation tool for retail promotions, staffing, and demand planning.

Future Directions

• Closing the gap to domain-specific supervised models without sacrificing generality. The hydrology comparison shows a supervised LSTM ensemble at NSE 0.90 against 0.68 for zero-shot Chronos-2, which frames fine-tuning, adaptation, or ensembling as open questions.
• Extending the framework to hierarchical and structured forecasting. The paper motivates hierarchical forecasting, but the truncated content does not report those results, leaving the design of group structures for aggregation constraints as an open problem.
• Better handling of out-of-distribution regimes and epistemic uncertainty. The model widens intervals appropriately under an always-on promotion schedule it never saw, but the paper treats this as reflecting uncertainty rather than as a solved extrapolation capability.
• Validating generalization beyond the current benchmark and case-study set. Whether the synthetic interaction generators cover the coupling structures of other scientific systems — and how sensitive results are to that coverage — is left to future work.

Target Audience

Forecasting researchers and practitioners who need accurate probabilistic predictions across heterogeneous datasets without per-dataset training; machine learning engineers building forecasting products on constrained hardware; and domain scientists in epidemiology, macroeconomics, climate/ecosystem science, and hydrology who want to know whether a generalist model can compete with their specialized tools. Readers new to forecasting will find the architecture explanation accessible, but the benchmark analysis assumes familiarity with scaled error metrics and quantile losses.

Authors’ abstract

Accurate forecasting with uncertainty quantification is critical in a variety of scientific applications ranging from epidemiology and economics to energy systems and earth sciences. Yet most forecasting models still need to be developed and trained separately for each domain-specific dataset. Time series foundation models (TSFMs) aim to remove this limitation and enable the application of the same model across domains through large-scale pretraining. However, most existing TSFMs treat each time series independently and therefore cannot model interactions among time series and external drivers. Here, we introduce Chronos-2, a foundation model for zero-shot forecasting that models the interactions between multivariate time series and heterogeneous covariates, which commonly arise in scientific applications. The model infers relationships among time series directly from historical context and generates probabilistic forecasts without task-specific treatment. Across various benchmarks, Chronos-2 achieves state-of-the-art accuracy on average by improving skill scores by 5.2 percentage points relative to the strongest baselines and increasing win rates by 10 percentage points. Chronos-2 is also computationally efficient, generating forecasts for 100 time series in under one second on a single mid-range GPU, 50x and 1000x faster end-to-end than the best task-specific point and probabilistic forecasters, respectively. In addition, we demonstrate its effectiveness in diverse scientific settings -- including infectious-disease forecasting, macroeconomic modeling, carbon-flux prediction, and runoff simulation -- where the model matches or surpasses specialized methods developed for each domain. These results show that a single pretrained model can capture structure shared across temporal systems.