Chronos-2: Universal Time Series Forecasting

• Chronos-2 is a pretrained time-series forecasting model that enables universal zero-shot forecasting across univariate, multivariate, and covariate-informed tasks using group attention.
• It employs an encoder-only Transformer with alternating time and group attention layers to efficiently separate temporal and cross-series aggregation.
• Benchmark results demonstrate state-of-the-art performance on diverse datasets, though performance may vary under regime shifts and high-dimensional noise.

Chronos-2 is a pretrained time-series foundation model for univariate, multivariate, and covariate-informed forecasting in a zero-shot manner. Its defining mechanism is group attention, which enables in-context learning across related series presented together at inference time, so that a group can represent multiple related univariate series, the variates of a multivariate series, or targets together with past-only and known covariates. The model is designed as an inference-only forecasting system that can be used “as is,” without task-specific fine-tuning, and its generality is obtained by training on synthetic datasets that impose diverse multivariate structures on univariate series (Ansari et al., 17 Oct 2025).

1. Concept and scope

Chronos-2 extends the Chronos line from “one series at a time” forecasting to what its authors call universal forecasting. In the formulation reported for the model, this universality consists of support for multiple related targets, past-only covariates, known future covariates, categorical covariates, and in-batch cross learning across groups of series. The grouping is a runtime choice rather than a fixed training-time schema, which means the same model can be used in univariate zero-shot mode, in multivariate mode, or in covariate-informed mode simply by changing the group composition and masks supplied at inference time (Ansari et al., 17 Oct 2025).

This design situates Chronos-2 within the recent family of pretrained time-series models while differentiating it from predecessors and contemporaries. Previous Chronos models such as Chronos-Bolt were univariate-only and lacked general covariate support. PatchTST and TimesNet are described as strong univariate transformers; Moirai-1 supports multivariate inputs but flattens them internally; Toto introduces cross-variate attention but lacks known and categorical covariate support; COSMIC advances covariate usage via synthetic augmentations but remains univariate for targets; and TabPFN-TS can incorporate known covariates but not past-only covariates or multivariate targets. Chronos-2 is presented as broader in scope because it combines universal in-context learning, multivariate forecasting, and covariate-informed forecasting within a single architecture (Ansari et al., 17 Oct 2025).

A practical implication is that Chronos-2 is not merely a multivariate forecaster. It is a group-conditioned forecasting system whose operational unit is the group, not the individual series. This suggests that the model’s main abstraction is relational structure supplied at inference time rather than a fixed notion of “channel” learned for one dataset.

2. Architectural design

Chronos-2 uses an encoder-only Transformer following the T5 encoder design, with rotary position embeddings (RoPE) in the time-attention layers. Its layers alternate between time attention, which operates along the temporal axis within each series, and group attention, which operates across items in the same group at the same patch index. The architectural purpose of this alternation is to separate temporal aggregation from cross-series aggregation, rather than flattening time and variates into one long token sequence (Ansari et al., 17 Oct 2025).

The input pipeline begins with per-dimension robust scaling and variance stabilization:

$$\tilde{x}_{t,d} = \sinh^{-1}\!\left(\frac{x_{t,d} - \mu_d}{\sigma_d}\right),$$

with $\mu_d,\sigma_d$ computed over historical observed values. Missing values are excluded when computing these statistics; after masking, missing entries are replaced by $0$ in the input stream. Chronos-2 also appends a relative time index and a binary observation mask indicating observed versus missing values. Sequences are then split into non-overlapping patches of length $P$, and for each patch the concatenated triple of input values, time indices, and masks is embedded by a residual network

$$f^{\mathrm{in}}_\phi:\mathbb{R}^{3P}\rightarrow\mathbb{R}^{D_h}.$$

A special REG token is inserted between context and future patch blocks, acting as a separator and attention sink. Categorical covariates are embedded to real values prior to scaling; univariate targets use target encoding, and multivariate targets use ordinal encoding (Ansari et al., 17 Oct 2025).

The group-attention mechanism is defined over patch-aligned item embeddings. If $X^{\mathrm{group}}_p\in\mathbb{R}^{S_g\times D_h}$ collects the embeddings of the $S_g$ items in group $g$ at patch index $p$, Chronos-2 forms

$$Q_p = X^{\mathrm{group}}_p W_Q,\quad K_p = X^{\mathrm{group}}_p W_K,\quad V_p = X^{\mathrm{group}}_p W_V,$$

and computes masked attention strictly within the group. By construction, items have no natural order across the group dimension, so no positional encoding is applied there (Ansari et al., 17 Oct 2025).

The computational consequence is central to the model’s positioning. Time attention costs $\mu_d,\sigma_d$0 per item, while group attention costs $\mu_d,\sigma_d$1 per patch. Because cross-series aggregation is confined to same-patch indices, the model’s memory scales linearly in the number of variates and covariates for fixed per-patch grouping, which the paper summarizes as Memory Scaling $\mu_d,\sigma_d$2. This is the technical rationale offered for using group attention instead of a globally flattened attention pattern (Ansari et al., 17 Oct 2025).

3. Pretraining, objective, and in-context learning

Chronos-2 is trained with a multi-quantile pinball loss over 21 quantiles,

$\mu_d,\sigma_d$3

and its output head produces direct multi-step forecasts

$\mu_d,\sigma_d$4

The loss is computed only on target dimensions. During inference, the normalized forecast is mapped back by

$\mu_d,\sigma_d$5

Training is staged: Stage 1 uses maximum context $\mu_d,\sigma_d$6 and a small number of output patches, and Stage 2 post-trains the model to extend context to $\mu_d,\sigma_d$7 and increase the maximum number of output patches (Ansari et al., 17 Oct 2025).

Because large corpora of multivariate and covariate-rich time series are scarce, Chronos-2 develops its “universal” behavior by training on synthetic datasets that impose multivariate structure on univariate generators. The base generators are AR, ETS, TSI, and KernelSynth. Two families of “multivariatizers” are used. Cotemporaneous constructions impose instantaneous relationships such as linear or nonlinear mixing, while sequential constructions induce lagged dependencies, shared trends, and cointegration. Covariates are produced by designating some variates as known covariates, with future masks equal to $\mu_d,\sigma_d$8, or as past-only covariates, with future masks equal to $\mu_d,\sigma_d$9 (Ansari et al., 17 Oct 2025).

In-context learning is implemented through prompt construction via groups. For univariate zero-shot forecasting, each series receives a distinct group ID. For multivariate forecasting, all variates of the multivariate target share a group ID. For covariate-informed forecasting, the targets and all covariates share a group ID, and only known covariates are populated in the future input matrix. The model then infers dependencies from context, masks, and group membership rather than from task-specific retraining (Ansari et al., 17 Oct 2025).

The reported ablations characterize the returns from this design unevenly across task classes. Cross learning improves univariate skill scores substantially on Chronos Benchmark II and on the univariate subsets of fev-bench and GIFT-Eval. Gains on pure multivariate tasks are described as modest. The largest gains arise on covariate-informed tasks, where Chronos-2 is reported to consistently outperform TabPFN-TS and COSMIC (Ansari et al., 17 Oct 2025). A plausible implication is that group attention is most valuable when it can exploit explicitly informative side information rather than merely co-movement among targets.

4. Forecasting modes and operational use

Chronos-2 supports three principal inference modes. In univariate zero-shot mode, each target is processed independently by assigning a unique group ID. In multivariate forecasting, all variates share one group ID so that group attention can exchange information across dimensions. In covariate-informed forecasting, targets, past-only covariates, and known covariates are grouped together; future values are supplied only for known covariates, while targets and past-only covariates remain masked in the future segment (Ansari et al., 17 Oct 2025).

The operational pipeline is explicit. One assembles the items and group IDs, applies the asinh-standardization per dimension, constructs the time index and mask, fills future-known covariates with observed values and mask $0$0, sets future targets and past-only covariates to $0$1 with mask $0$2, patches each item, inserts the REG token, runs the alternating time/group-attention stack, and denormalizes the quantile outputs. Point forecasts use the median $0$3, while predictive intervals are formed from quantile pairs such as $0$4–$0$5 (Ansari et al., 17 Oct 2025).

The reported model sizes are ~120M parameters for the base model and ~28M parameters for the small model. On an NVIDIA A10G, the base model achieves ~300 time series per second for a batch of 1,024 items, context length 2,048, and horizon 64. The guidance given for deployment is to keep groups small or moderate, co-batch items with similar frequencies and horizons, and use the 28M variant for CPU-only or latency-sensitive settings. The implementation is in PyTorch and uses an encoder-only stack with RoPE and a quantile head (Ansari et al., 17 Oct 2025).

These usage conventions make Chronos-2 unusual among forecasting systems in that the modeling decision with the largest practical effect may be group formation rather than model retraining. The papers recommend grouping by metadata such as store, category, region, product family, or other evidence of shared dynamics, and they caution that the quality of grouping strongly influences zero-shot performance (Ansari et al., 17 Oct 2025).

5. Benchmark performance and applied studies

Chronos-2 is reported to achieve state-of-the-art performance on fev-bench, GIFT-Eval, and Chronos Benchmark II. On fev-bench, using scaled quantile loss (SQL), it attains 90.7% win rate, 47.3% skill score, 3.6 s median runtime, 0% leakage, and 0 failures. On additional fev-bench metrics it records 88.5% win rate and 51.5% skill for WQL, 87.9% and 35.5% for MASE, and 85.4% and 39.4% for WAPE. On GIFT-Eval, Chronos-2 reaches 81.9% win rate and 51.4% skill in WQL and 83.8% and 30.2% in MASE. On Chronos Benchmark II, it records 79.8% and 46.6% in WQL and 81.5% and 26.5% in MASE. The paper further states that pairwise comparisons against TiRex and TimesFM-2.5 are statistically decisive in fev-bench: the 95% confidence intervals of pairwise win rates exclude 50%, and pairwise skill score intervals exclude 0% (Ansari et al., 17 Oct 2025).

Benchmark	Metric	Chronos-2 result
fev-bench	SQL	90.7% win rate, 47.3% skill
fev-bench	WQL	88.5% win rate, 51.5% skill
GIFT-Eval	WQL	81.9% win rate, 51.4% skill
Chronos Benchmark II	WQL	79.8% win rate, 46.6% skill

The qualitative case studies emphasize domains where covariates are central. In energy (EPF-DE), using load and renewable-generation covariates improves forecast accuracy markedly relative to univariate mode. In retail (Rossmann weekly sales), promotion and holiday covariates allow the model to track the true dynamics over the horizon, whereas univariate forecasts are described as flat with high uncertainty (Ansari et al., 17 Oct 2025).

An independent financial study evaluates Chronos-2 in rolling monthly forecasts from 2000–2025 on the Magnificent-7 equities, U.S. Treasury yields, and a combined panel. Within-domain multivariate forecasting consistently outperforms univariate forecasting. For U.S. Treasury yields, multivariate input reduces mean MAPE from 0.1233 to 0.0497 and mean RMSE from 0.1865 to 0.0418; for equities, mean MAPE falls from 0.0844 to 0.0706 and mean RMSE from 5.0395 to 3.9619. Series-level gains are reported for every maturity and every ticker. However, combining equities and interest rates into one group reduces forecast accuracy relative to within-asset multivariate grouping, which the study interprets as degradation from noisy context (Das et al., 8 May 2026).

6. Limitations, critiques, and research significance

The main Chronos-2 paper identifies several limitations. Zero-shot performance can degrade under extreme regime shifts, structural breaks, or novel seasonalities absent in pretraining. In very high-dimensional settings, memory remains linear in group size, but compute per patch becomes $0$6 if all variates are grouped together, so the authors recommend grouping only strongly coupled variates or splitting into subgroups. Incorrect masks or timestamp misalignment can weaken covariate use, especially when holidays or promotions are off by one period. Because Chronos-2 uses independent quantile regression, quantile crossing is possible, so post-processing for monotonicity or conformal calibration may be required in risk-sensitive deployments (Ansari et al., 17 Oct 2025).

A subsequent controlled study questions how effectively Chronos-2 uses simple covariates. In experiments based on identity, sum, aggregate, and quadratic target–covariate relationships, TabPFN-TS captures these relationships more effectively than Chronos-2, especially for short horizons. In the identity experiment on Electricity with 1344–336, Chronos-2 improves from 0.2037 to 0.0085 when given the target as a covariate, but TabPFN-TS improves from 0.2367 to 0.0056; on Solar with the same horizon, Chronos-2(+id) reaches 0.0456 whereas TabPFN-TS(+id) reaches 0.0000. In the aggregate experiment at 1344–336, Chronos-2 changes from 0.1880 to 0.1940 with covariates, while TabPFN-TS changes from 0.2375 to 0.0008. The same paper reports that as the autoregressive coefficient $0$7 increases in

$0$8

the advantage flips toward Chronos-2, consistent with stronger temporal dependence favoring its time and group attention mechanisms (Berthelier et al., 12 May 2026).

These results refine the interpretation of Chronos-2’s benchmark leadership. They do not contradict its strong performance on broad forecasting benchmarks, but they indicate that high aggregate accuracy does not automatically imply optimal handling of simple contemporaneous target–covariate mappings. The emerging picture is that Chronos-2 is strongest when predictability is mediated by temporal structure, long context, and structured cross-series information, whereas short-horizon tasks dominated by simple feature-to-target mappings may favor models with a stronger in-context regression prior (Berthelier et al., 12 May 2026).

From a research perspective, Chronos-2 is significant because it makes group structure a first-class object in zero-shot forecasting. Its contribution is not only a performance increment over earlier TSFMs, but a reformulation of forecasting as inference over temporally patched, group-conditioned sequences. This suggests a line of development in which future TSFMs may combine Chronos-2’s temporal and cross-series machinery with the more explicit regression-oriented covariate handling that controlled studies found beneficial in TabPFN-TS (Ansari et al., 17 Oct 2025).