- The paper introduces a unified Darts API that standardizes integration across multiple time series foundation models for zero-shot forecasting.
- It demonstrates high-fidelity performance with benchmarks showing deviations under 1.3% and significant gains over seasonal baselines.
- Practical insights include support for probabilistic forecasting, model-agnostic explainability, and flexible fine-tuning for enhanced deployment.
Unified Zero-Shot Time Series Forecasting with Darts Foundation Model
Introduction
This paper introduces a unified foundation for zero-shot time series forecasting within the Darts Python library, encompassing integration, supervision, and interface standardization for contemporary Time Series Foundation Models (TSFMs), namely Chronos-2, TimesFM 2.5, TiRex, and PatchTST-FM. The work addresses the fragmentation of TSFM libraries and interfaces, which historically posed substantial barriers to interoperability, fair evaluation, pipeline integration, and real-world adoption. By embedding TSFMs in a consistent framework, Darts aligns with general-purpose forecasting objectives and advances the practical deployment of large pre-trained models for time series analysis.
Context and Motivation
Recent advances in TSFMs—analogous to LLMs in NLP—have demonstrated strong zero-shot performance across a broad array of time series forecasting tasks, underpinned by large-scale pretraining on real and synthetic datasets (Ansari et al., 17 Oct 2025, Das et al., 2023, Auer et al., 29 May 2025, Wen et al., 6 Feb 2026). Despite empirical successes in domains such as epidemiology, retail, and manufacturing [(Ansari et al., 17 Oct 2025), bruned2025showdown, zhang2025defect], the lack of standardized interfaces and pipeline compatibility limits their comparative evaluation and integration with established data science practices. Darts, a mature time series library, is extended here to provide a unified API, robust support for key TSFM features, and seamless compatibility with established data processing and evaluation toolchains.
Unified FoundationModel API and Architecture
The FoundationModel class provides a thin, standardized extension atop Darts' TorchForecastingModel, streamlining integration for PyTorch-native TSFMs. The core fit and predict methods, consistent with prior Darts models, operate over immutable, type-safe TimeSeries containers. This design ensures compatibility with batch forecasting, covariate handling, probabilistic outputs, and evaluation flows, leveraging PyTorch Lightning for backend orchestration and cross-device execution.
Provider integration is straightforward: developers subclass PLForecastingModule to specify model forward passes and FoundationModel to instantiate these modules. Hugging Face–style safetensors support and optional connector classes are supplied, enabling rapid deployment and distribution of large TSFM weights.
Core Capabilities
Covariate and Multi-Series Support
Native support for exogenous covariates (past and future) is realized through TimeSeries alignment and containerized API patterns, enhancing out-of-the-box utility for TSFMs designed for covariate-aware forecasting (e.g., Chronos-2). Simultaneous forecasting of multiple series is enabled via batched execution, optimizing throughput without loss of semantic mapping to input series.
Historical Forecasting and Backtesting
Efficient historical forecasting and robust error backtesting are cornerstones of production-grade forecasting. The historical_forecast method allows both retraining and non-retraining regimes, supports multi-series input, and leverages batching for speed. Full interoperability with Darts' evaluation and residual analysis methods is preserved.
Probabilistic Forecasting
FoundationModel generalizes probabilistic forecasting across TSFMs by supporting standard Darts likelihoods (e.g., QuantileRegression) and Monte Carlo or direct parameter prediction modes. This permits uncertainty quantification and risk-aware inference, with sampling tailored to the likelihoods used during TSFM pretraining or custom user-defined distributions.
Fine-Tuning Flexibility
Recognizing the value of TSFM adaptation, the enable_finetuning flag provides granular control over trainable parameters, supporting zero-shot (frozen), full, or partial fine-tuning regimes via shell-style matching. Though parameter-efficient adapters like LoRA are not currently supported, this mechanism covers the dominant paradigms and is extended to non-TSFMs in the Darts ecosystem.
Model-Agnostic Explainability
SHAP-based explainability for all PyTorch-based models is available through Permutation SHAP, permitting forecast attribution analysis without architecture-specific implementation. Deep and Gradient SHAP are omitted due to architectural limitations, but the default approach ensures applicability across current and future TSFMs.
Additional Features
Further capabilities inherited from Darts include auto-regressive support, mixed-precision computation, ensemble modeling, conformal prediction for calibrated intervals, and built-in hyperparameter search, establishing a comprehensive and production-ready environment.
Empirical Validation
Benchmarking of implemented TSFMs (Chronos-2, TimesFM 2.5, TiRex) on the fev-bench-mini suite (Shchur et al., 30 Sep 2025), using skill scores for both probabilistic (SQL) and deterministic (MASE) regimes, indicates high-fidelity reproduction versus original implementations. Deviations do not exceed 1.3%, and all models show substantial gains over naive seasonal baselines (e.g., Chronos-2 achieves 50.4% SQL and 39.3% MASE). These results confirm both the correctness of the unified API and the capability for seamless evaluation within the Darts framework.
Practical and Theoretical Implications
The unified Darts FoundationModel API closes the deployment gap for TSFMs by facilitating fair, extensible, and reproducible evaluation, lowering barriers for both end-users and model developers. For practitioners, rapid model swapping, ensembling, uncertainty quantification, and explainability now require minimal plumbing, promoting practical adoption in heterogeneous forecasting pipelines.
On the research front, the architecture standardizes evaluation protocols and simplifies ablation, fostering improved rigor in TSFM research and comparative studies. However, challenges remain—current limitations include support for only four TSFMs and the absence of parameter-efficient fine-tuning. Ongoing work and community contributions are expected to enrich model diversity and introduce further adaptation mechanisms.
Conclusion
This work delivers a unified, extensible framework for zero-shot and fine-tuned TSFM deployment, grounded in the Darts library's mature time series infrastructure. Through standardized APIs, integrated probabilistic and explainability features, and robust benchmarking, the platform bridges the interoperability divide and accelerates both applied and methodological advances in time series forecasting with foundation models. Expansion to additional TSFMs and further fine-tuning paradigms will shape the trajectory of foundation model application and evaluation in temporal analytics.
References:
- "Unified Zero-Shot Time Series Forecasting: A Darts Foundation" (2606.27438)
- "Chronos-2: From Univariate to Universal Forecasting" (Ansari et al., 17 Oct 2025)
- "A decoder-only foundation model for time-series forecasting" (Das et al., 2023)
- "TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning" (Auer et al., 29 May 2025)
- "Revisiting the Generic Transformer: Deconstructing a Strong Baseline for Time Series Foundation Models" (Wen et al., 6 Feb 2026)
- "fev-bench: A Realistic Benchmark for Time Series Forecasting" (Shchur et al., 30 Sep 2025)
