A Family of Open Time-Series Foundation Models for the Radio Access Network

Ioannis Panitsas; Leandros Tassiulas

cs.NIApr 5, 2026

A Family of Open Time-Series Foundation Models for the Radio Access Network

Authors: Ioannis Panitsas, Leandros Tassiulas

Organizations: Department of Electrical and Computer Engineering, Yale University

Abstract

The Radio Access Network (RAN) is evolving into a programmable and disaggregated infrastructure that increasingly relies on AI-native algorithms for optimization and closed-loop control. However, current RAN intelligence is still largely built from task-specific models tailored to individual functions, resulting in model fragmentation, limited knowledge sharing across tasks, poor generalization, and increased system complexity. To address these limitations, we introduce TimeRAN, a unified multi-task learning framework for time-series modeling in the RAN. TimeRAN leverages a lightweight time-series foundation model with few task-specific heads to learn transferable representations that can be efficiently adapted across diverse tasks with limited supervision. To enable large-scale pretraining, we further curate and open-source TimeRAN DataPile, the largest time-series corpus for RAN analytics to date, comprising over 355K time series and 0.56B measurements across diverse telemetry sources, protocol layers, and deployment scenarios. We evaluate TimeRAN across a comprehensive set of RAN analytics tasks, including anomaly detection, classification, forecasting, and imputation, and show that it achieves state-of-the-art performance with minimal or no task-specific fine-tuning. Finally, we integrate TimeRAN into a proof-of-concept 5G testbed and demonstrate that it operates efficiently with limited resource requirements in real-world scenarios.

Figures & tables

Figure 1. Overview of the TimeRAN DataPile.

Panel of Figure 3 — Figure 3. Proof-of-concept over-the-air 5G testbed.

Tasks	Supervision	Datasets
Anomaly Detection	Zero-shot Linear probing	AERPAW 18/23 ( Asokan and others, 2024 ; Singh, 2024a ) , BLT ( Fechete and others, 2025 ) , Hetnets ( Chatzistefanidis and others, 2023 ) , Jamshield ( Panitsas and others, 2025b )
Classification	Linear probing Full-finetuning	AERPAW 20 ( Reddy and others, 2025 ) , Irish ( Raca and others, 2020 ) , Spotlight ( Sun and others, 2024 ) , TelecomTS ( Feng and others, 2025 ) , Tractor ( Groen and others, 2024 )
Forecasting	Linear probing Full-finetuning	AERPAW 18/24/25 ( Asokan and others, 2024 ; Singh, 2024c ; Singh, 2024b ) , QoE Aware ( Kabeer and others, 2025 ) , Queens ( Elsherbiny and others, 2020 )
Imputation	Zero-shot Linear probing	Irish ( Raca and others, 2020 ) , Open Ireland ( Xavier and others, 2024 ) , Hetnets ( Chatzistefanidis and others, 2023 ) , WINS ( Ghoshal and others, 2025 ) , NOK ( Fondo-Ferreiro and others, 2023 )

Table 1. TimeRAN evaluation benchmark.

Panel of Figure 4 — Figure 4. Training loss and size of TimeRAN variants.

Methods	TimeRAN LP	TimeRAN 0	MOMENT LP	MOMENT 0	Autoformer	Informer	TimesNet
Dataset	Adj. F1	Adj. F1	Adj. F1	Adj. F1	Adj. F1	Adj. F1	Adj. F1
AERPAW 18 ( Asokan and others, 2024 )	0.93	0.86	0.83	0.90	0.91	0.90	0.89
AERPAW 23 ( Singh, 2024a )	0.99	0.98	0.98	0.98	0.93	0.94	0.94
BLT ( Fechete and others, 2025 )	0.98	0.95	0.93	0.91	0.86	0.89	0.83
Hetnets ( Chatzistefanidis and others, 2023 )	0.99	0.99	0.98	0.98	0.96	0.95	0.96
Jamshield ( Panitsas and others, 2025b )	0.97	0.93	0.91	0.90	0.92	0.92	0.93

Table 2. Anomaly detection performance on the curated evaluation benchmark.

Methods		TimeRAN ${}_{\textbf{FF}}$			TimeRAN ${}_{\textbf{LP}}$			MOMENT ${}_{\textbf{FF}}$			Informer			TimesNet			1D-CNN			LSTM
Dataset	Task	P.	R.	F1	P.	R.	F1	P.	R.	F1	P.	R.	F1	P.	R.	F1	P.	R.	F1	P.	R.	F1
AERPAW 20 ( Reddy and others, 2025 )	Location	1.00	0.99	0.99	0.62	0.80	0.70	1.00	0.98	0.99	0.65	0.76	0.71	0.61	0.69	0.68	0.95	0.90	0.93	0.68	0.72	0.71
Irish ( Raca and others, 2020 )	Mobility	1.00	0.99	0.99	0.96	0.89	0.92	0.99	0.99	0.99	0.95	0.91	0.92	0.96	0.92	0.94	0.98	0.99	0.98	0.96	0.93	0.95
Spotlight ( Sun and others, 2024 )	Anomaly	0.93	0.94	0.93	0.84	0.83	0.84	0.93	0.95	0.94	0.83	0.89	0.86	0.81	0.89	0.84	0.78	0.92	0.81	0.90	0.83	0.87
Spotlight ( Sun and others, 2024 )	Root-Cause	0.99	0.99	0.99	0.55	0.45	0.40	0.96	0.95	0.95	0.57	0.58	0.58	0.61	0.59	0.57	0.63	0.66	0.61	0.64	0.65	0.64
TelecomTS ( Feng and others, 2025 )	Congestion	0.98	1.00	0.99	0.58	0.95	0.72	0.97	0.98	0.98	0.74	0.99	0.85	0.72	0.99	0.84	0.87	0.93	0.90	0.75	0.99	0.86

Table 3. Classification performance on the curated evaluation benchmark.

Methods		TimeRAN ${}_{\textbf{FF}}$		TimeRAN ${}_{\textbf{LP}}$		MOMENT ${}_{\textbf{FF}}$		Autoformer		Informer		TimesNet		1D-CNN		LSTM		ETS		ARIMA
Dataset	Horizon (H)	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
AERPAW 18 ( Asokan and others, 2024 )	32	25.55	3.86	23.55	3.78	31.92	4.33	21.70	3.55	31.30	4.16	22.78	3.67	28.63	4.08	20.43	3.40	34.32	4.61	31.78	4.33
	64	26.46	3.99	26.29	3.97	31.61	4.32	22.79	3.66	23.47	3.77	24.46	3.82	27.41	4.05	22.92	3.70	38.81	4.78	34.12	4.60
	128	28.17	4.07	25.15	3.86	27.45	4.05	23.36	3.76	24.63	3.84	24.11	3.81	25.66	3.93	23.26	3.73	39.21	4.81	37.21	4.71
	208	28.69	4.10	25.71	3.93	27.32	4.04	25.25	3.87	25.81	3.96	24.40	3.81	24.69	3.85	24.84	3.87	41.44	5.11	40.11	5.01
AERPAW 24 ( Singh, 2024c )	32	17.70	3.53	14.75	3.14	19.00	3.56	16.86	3.50	20.26	3.79	14.27	3.10	27.33	4.39	15.45	3.18	29.80	4.65	26.48	4.32

Table 4. Forecasting performance on the curated evaluation benchmark.

Methods		TimeRAN LP		TimeRAN 0		MOMENT FF		Forward Fill		Mean		N. Neighbors		Linear		Rolling Mean
Dataset	Mask Ratio	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
Irish ( Raca and others, 2020 )	10%	4.95	1.44	5.16	1.48	23.04	3.02	13.02	2.34	20.82	2.92	10.80	1.94	8.45	1.86	12.18	2.26
	30%	5.96	1.56	6.21	1.60	22.55	2.98	16.72	2.53	24.59	3.07	14.50	2.40	11.86	2.00	14.68	2.42
	50%	7.38	1.75	8.47	1.87	24.40	3.08	16.77	2.57	23.67	3.05	15.56	2.49	12.40	2.28	14.90	2.45
Open Ireland ( Xavier and others, 2024 )	10%	21.02	2.54	20.80	2.53	23.64	2.78	62.46	4.38	34.52	3.78	62.33	4.37	52.35	4.17	34.46	3.61
	30%	22.75	2.69	22.67	2.68	24.60	2.87	62.80	4.40	34.90	3.80	62.55	4.38	52.16	4.16	35.03	3.80

Table 5. Imputation performance on the curated evaluation benchmark.

Panel of Figure 5 — Figure 5. TimeRAN learns separable representations in a zero-shot setting without dataset-specific fine-tuning.

Panel of Figure 6 — Figure 6. TimeRAN generalizes to unseen environments in a zero-shot setting.

Panel of Figure 7 — Figure 7. System-level performance of TimeRAN under varying input characteristics during inference.

Panel of Figure 8 — Figure 8. TimeRAN performance across representative RAN analytics tasks on our over-the-air 5G testbed.

Wireless Foundation Models

Wireless foundation models learn reusable representations of radio channels, signals, network traces, and telemetry for transfer across wireless tasks and settings.

See all 33 papers

RRM-GPT: A Framework and Vision for Radio Resource Management Foundation Models
Oct 3, 2026Ahmed Aboulfotouh, Akram Bin Sediq, Koosha Pourtahmasi Roshandeh +4Wireless Resource Allocation Wireless Foundation Models
Department of Electrical and Software Engineering, University of Calgary, Canada · Ericsson Canada Inc., Canada
TimeNet: An Extensible Unified Data Infrastructure for Next-Generation Temporal Foundation Models
Oct 3, 2026Martin Maritsch, Timo Stoffregen, Thomas Kaar +36Time-Series Foundation Models
Aionic Labs, Switzerland · Agentic Systems Lab, ETH Zurich, Switzerland · Stanford University, USA +17
Toward Autonomous O-RAN: A Multi-Scale Agentic AI Framework for Real-Time Network Control and Management
Feb 15, 2026Hojjat Navidan, Mohammad Cheraghinia, Jaron Fontaine +5Multi-Agent Orchestration Language Model-Based Control
IDLab, Department of Information Technology at Ghent University - imec, Ghent, Belgium · Department of Electrical and Computer Engineering, Princeton University, USA

Oct 3, 2026J/K move · Enter open · S save

RRM-GPT: A Framework and Vision for Radio Resource Management Foundation Models

Ahmed Aboulfotouh, Akram Bin Sediq, Koosha Pourtahmasi Roshandeh, Omar Mashaal+3

Department of Electrical and Software Engineering, University of Calgary, Canada · Ericsson Canada Inc., Canada

Wireless Resource Allocation Wireless Foundation Models

Open paper AskPDF arXiv

Learning-based models for radio resource management (RRM) are typically built for a single function and deployment, so each new setting repeats the development pipeline. RRM decisions, however, share a common structure: each is assembled from interdependent fields, defined by the standard, whose values are selected in view of the network state. We propose RRM-GPT, an autoregressive framework for RRM foundation models that generate these decisions as a language model generates text. An encoder maps heterogeneous network observations into a common token representation, and a decoder emits the decision one field at a time, each conditioned on the network state and the fields already committed. Pretraining on unannotated network logs teaches the model what makes a decision valid and how controllers choose among valid decisions; post-training then adapts it to deployment-specific operator objectives through imitation or reinforcement learning. The framework targets two forms of reuse: a function-specific model reused across deployments, and a model shared across RRM functions that generates their interdependent decisions as one sequence. In a 5G New Radio (NR) case study, we demonstrate that a single model generates complete scheduling grants spanning user selection, timing, link adaptation, resource allocation, and control signaling. The model captures dependencies among grant fields and transfers learned behavior to an unseen scenario without adaptation.

Abstract

Wireless Foundation Models

Explore similar work

RRM-GPT: A Framework and Vision for Radio Resource Management Foundation Models

TimeNet: An Extensible Unified Data Infrastructure for Next-Generation Temporal Foundation Models

Toward Autonomous O-RAN: A Multi-Scale Agentic AI Framework for Real-Time Network Control and Management