The Nixtlaverse: An Open-Source Ecosystem for Forecasting

Olivier Sprangers; Max Mergenthaler Canseco; Marco Peixeiro; Saul Caballero Ramirez; Mariana Menchero García; Jing-Qiang Goh; Han Wang; Nikhil Gupta

cs.LGSep 30, 2026

The Nixtlaverse: An Open-Source Ecosystem for Forecasting

Authors: Olivier Sprangers, Max Mergenthaler Canseco, Marco Peixeiro, Saul Caballero Ramirez, Mariana Menchero García, Jing-Qiang Goh, Han Wang, Nikhil Gupta, +3 more

Organizations: Nixtla, San Francisco, CA, United States

Abstract

Large forecasting applications often combine statistical, machine-learning, and neural models. These families solve the same problem but differ in fitted state, training procedures, and how they parallelize work. Forecasting software must therefore either hide these differences behind a single estimator interface, or keep the families in separate packages, forcing users to rewrite data preparation and evaluation for every package. We present the Nixtlaverse, an ecosystem of open-source Python libraries for time series forecasting, as a case study of a third design: all libraries share the same long-format panel data and keyed forecast outputs, while every model family keeps its own specialized implementation. We demonstrate this design through three use cases on the public M5 competition data. First, we evaluate statistical, machine-learning, and neural models, and an external engine from a separate ecosystem, in a single rolling-origin evaluation with per-series and hierarchy-weighted metrics. Second, we profile runtime and peak memory from 100 to 30,490 series and locate each family's bottleneck: statistical fitting scales approximately linearly in the number of series, feature construction dominates machine-learning memory, and neural training time is nearly independent of panel size under a fixed training budget. Third, we reconcile the forecasts of multiple engines, including the external one, over all 42,840 series of the M5 hierarchy, with sparse reconciliation where dense implementations exhausted memory. These use cases establish the costs, boundaries, and utility of shared data and output contracts. The Nixtlaverse has seen substantial public distribution, scholarly reuse, and adoption through other forecasting frameworks, and is released under permissive open-source licenses with public datasets, reproducible examples, and verifiable benchmark artifacts.

Figures & tables

Component	Version	Licence	Public project assets
StatsForecast	2.0.3	Apache-2.0	source, documentation, tests/CI, guide
MLForecast	1.0.31	Apache-2.0	source, documentation, tests/CI, guide
NeuralForecast	3.2.0	Apache-2.0	source, documentation, tests/CI, guide
HierarchicalForecast	1.3.1	Apache-2.0	source, documentation, tests/CI, guide
CoreForecast	0.0.18	Apache-2.0	source, documentation, tests/CI
UtilsForecast	0.2.16	Apache-2.0	source, documentation, tests/CI, guide

Table 1: Open-source components used in the paper. Reported package versions and their release licenses were recorded together with the public project assets in July 2026 from the projects’ public repositories (URL on the title page). “Guide” denotes a repository-level contribution guide. The artifact records the version-pinned dependency closure used by the benchmark.

Panel of Figure 1 — Figure 1: Overview of the open-source Nixtlaverse. Packages are drawn as sharp-cornered boxes and the shared data contracts as rounded boxes; the legend gives the meaning of the fills. The forecasting engines consume the same long panel representation, optionally pre-processed by UtilsForecast (e.g. filling date gaps), and return keyed forecast dataframes, which downstream evaluation and reconciliation consume regardless of the producing engine. The dashed line indicates that the three engines build on the compiled grouped operations of CoreForecast .

Library	GitHub stars	PyPI downloads	Google Scholar citations
StatsForecast	4,876	1,942,775	135
MLForecast	1,269	505,473	—
NeuralForecast	4,248	285,120	146
HierarchicalForecast	756	289,221	29

Table 2: Public indicators of reach for the principal Nixtlaverse libraries, retrieved on 21 August 2026. Downloads are distribution events recorded by PyPI during the preceding 30 days. Citation counts are those displayed for the canonical software records in Google Scholar; MLForecast did not have a separately indexed software record. These indicators are not necessarily counts of distinct users, deployments, or citing works across the ecosystem.

Model	Engine	mean	95% CI	RMSE	WRMSSE	Total
		MASE
SeasonalNaive	StatsForecast	1.196	[1.175, 1.229]	1.781	0.847	0.99
AutoETS	StatsForecast	1.050	[1.028, 1.084]	1.382	0.673	0.99
LightGBM	MLForecast	1.568	[1.544, 1.598]	1.534	1.045	1.06
NHITS	NeuralForecast	0.874	[0.856, 0.906]	1.421	1.888	0.70
Croston	sktime (external)	1.073	[1.056, 1.101]	1.412	0.957	1.03

Table 3: Rolling-origin results on the complete M5 bottom-level panel (30,490 series, three origins, re-estimation at every origin), for the fixed configurations of A.2 . Metrics use raw, unclipped engine outputs. MASE and RMSE are unweighted means over series and origins; the bracketed 95 percent bootstrap intervals resample whole series, and the corresponding RMSE intervals, which are of comparable width, are in the artifact. WRMSSE is the competition’s weighted metric over all twelve hierarchy levels at the final origin. The Total column reports the aggregate forecast of the final origin relative to the observed total; cross-validation times and the resampling details are in A.3 . The external Croston engine runs in an isolated environment with no Nixtla package installed ( A.6 ).

Figure 3: Resource profile per engine (columns) as the number of series increases, for subsets drawn proportionally in product category and store. The first four rows show the runtime of each computational phase separately (log–log; the statistical and neural engines have no separate preparation phase), the fifth row the total runtime over all recorded phases including data loading, and the bottom row peak memory as aggregate proportional set size (PSS) during preparation, fitting, and prediction. Medians over five cold-start repetitions; axes are shared across engines within each row.

	100 series		1,000 series
Workers	Fit (s)	Speedup	Fit (s)	Speedup	Exponent
1	23.8	—	233.8	—	0.99
4	7.0	$3.4\times$	62.3	$3.75\times$	0.95
20	3.4	$6.9\times$	23.8	$9.8\times$	0.84

Table 4: Effect of parallelization for the statistical engine: StatsForecast fitting AutoETS alone, medians of three repetitions in fresh processes. Twenty workers is every logical CPU of the machine ( A.7 ). The exponent is the slope of fitting time against the number of series on log–log axes between the two panel sizes; speedups are relative to one worker at the same panel size.

Backend	Forecast	Constr.	Total	(GiB)
	Runtime (s)			Peak memory
Multiprocessing	738	—	738	3.1
Ray	865	32	897	19.2

Table 5: Distributed backend against local multiprocessing for the statistical engine only: StatsForecast forecasting the SeasonalNaive–AutoETS portfolio over the full 30,490-series panel through the streaming .forecast interface, one measured run per backend. Construction is the one-time conversion of the in-memory panel into Ray’s distributed object store. Peak memory is aggregate proportional set size over parent and worker processes.

		MASE, all levels			MASE, total level
Forecasts	Runtime (s)	AutoETS	LightGBM	Croston	AutoETS	LightGBM	Croston
Base	—	1.031	1.476	1.047	0.653	2.163	1.457
BottomUp	60 / 57 / —	1.029	1.653	1.047	0.677	1.723	1.439
MinT (wls_var)	179 / 197 / —	1.033	1.321	—	0.672	0.605	—

Table 6: Base and reconciled forecasts on the complete M5 hierarchy (42,840 series, 28-day holdout at the final origin), using sparse reconciliation implementations. MASE is the mean over series; the total level is the single top series. Runtime covers the reconciliation step only, per base engine, from one measured run each; it excludes the time of hierarchy construction. A second complete pass of the same configurations is retained in the artifact as superseded; Section 6 reports how far it differs. The dense configurations attempted, BottomUp and MinT-shrink, did not complete; dense MinT with the wls_var estimator was not run (Section 6 ). Variance-weighted reconciliation is not run for the external engine, which does not supply in-sample fitted values ( A.6 ).

Explore similar work

Apr 17, 2026cs.LG

Chronax: A Jax Library for Univariate Statistical Forecasting and Conformal Inference

Time-series forecasting is central to many scientific and industrial domains, such as energy systems, climate modeling, finance, and retail. While forecasting methods have evolved from classical statistical models to automated, and neural approaches, the surrounding software ecosystem remains anchored to the traditional Python numerical stack. Existing libraries rely on interpreter-driven execution and object-oriented abstractions, limiting composability, large-scale parallelism, and integration with modern differentiable and accelerator-oriented workflows. Meanwhile, today's forecasting increasingly involves large collections of heterogeneous time series data, irregular covariates, and frequent retraining, placing new demands on scalability and execution efficiency. JAX offers an alternative paradigm to traditional stateful numerical computation frameworks based on pure functions and program transformations such as just-in-time compilation and automatic vectorization, enabling end-to-end optimization across CPUs, GPUs, and TPUs. However, this modern paradigm has not yet been fully incorporated into the design of forecasting systems. We introduce Chronax, a JAX-native time-series forecasting library that rethinks forecasting abstractions around functional purity, composable transformations, and accelerator-ready execution. By representing preprocessing, modeling, and multi-horizon prediction as pure JAX functions, Chronax enables scalable multi-series forecasting, model-agnostic conformal uncertainty quantification, and seamless integration with modern machine learning and scientific computing pipelines.

Xan Carey, Yash Deshmukh, Aileen Huang +7

Simulacrum New York City, NY, USA

Time Series Forecasting Chronos Global Climate Models

Ask PDF

May 14, 2026cs.AI

Nexus : An Agentic Framework for Time Series Forecasting

Time series forecasting is not just numerical extrapolation, but often requires reasoning with unstructured contextual data such as news or events. While specialized Time Series Foundation Models (TSFMs) excel at forecasting based on numerical patterns, they remain unaware to real-world textual signals. Conversely, while LLMs are emerging as zero-shot forecasters, their performance remains uneven across domains and contextual grounding. To bridge this gap, we introduce Nexus, a multi-agent forecasting framework that decomposes prediction into specialized stages: isolating macro-level and micro-level temporal fluctuations, and integrating contextual information when available before synthesizing a final forecast. This decomposition enables Nexus to adapt from seasonal signals to volatile, event-driven information without relying on external statistical anchors or monolithic prompting. We show that current-generation LLMs possess substantially stronger intrinsic forecasting ability than previously recognized, depending critically on how numerical and contextual reasoning are organized. Evaluated on data strictly succeeding LLM knowledge cutoffs spanning Zillow real estate metrics and volatile stock market equities, Nexus consistently matches or outperforms state-of-the-art TSFMs and strong LLM baselines. Beyond numerical accuracy, Nexus produces high-quality reasoning traces that explicitly show the fundamental drivers behind each forecast. Our results establish that real-world forecasting is an agentic reasoning problem extending well beyond only sequence modeling.

Sarkar Snigdha Sarathi Das, Palash Goyal, Mihir Parmar +6

1Google · 2Pennsylvania State University

Retrieval-Augmented Forecasting Time Series Foundation Models Agentic Framework Agentic

Ask PDF

Sep 30, 2025cs.LG

fev-bench: A Realistic Benchmark for Time Series Forecasting

Benchmark quality is critical for meaningful evaluation and sustained progress in time series forecasting, particularly with the rise of pretrained models. Existing benchmarks often have limited domain coverage or overlook real-world settings such as tasks with covariates. Their aggregation procedures frequently lack statistical rigor, making it unclear whether observed performance differences reflect true improvements or random variation. Many benchmarks lack consistent evaluation infrastructure or are too rigid for integration into existing pipelines. To address these gaps, we propose fev-bench, a benchmark of 100 forecasting tasks across seven domains, including 46 with covariates. Supporting the benchmark, we introduce fev, a lightweight Python library for forecasting evaluation emphasizing reproducibility and integration with existing workflows. Using fev, fev-bench employs principled aggregation with bootstrapped confidence intervals to report performance along two dimensions: win rates and skill scores. We report results on fev-bench for pretrained, statistical, and baseline models and identify promising future research directions.

Oleksandr Shchur, Abdul Fatir Ansari, Caner Turkmen +5

AWS · Keystone AI · Prior Labs +2

Time Series Forecasting Dense Baseline Reproducibility

Ask PDF