TimeSAF: Towards LLM-Guided Semantic Asynchronous Fusion for Time Series Forecasting

Fan Zhang; Shiming Fan; Hua Wang

cs.LGApr 14, 2026

TimeSAF: Towards LLM-Guided Semantic Asynchronous Fusion for Time Series Forecasting

Authors: Fan Zhang, Shiming Fan, Hua Wang

Organizations: Shandong Technology and Business University · Ludong University

Abstract

Despite the recent success of large language models (LLMs) in time-series forecasting, most existing methods still adopt a Deep Synchronous Fusion strategy, where dense interactions between textual and temporal features are enforced at every layer of the network. This design overlooks the inherent granularity mismatch between modalities and leads to what we term semantic perceptual dissonance: high-level abstract semantics provided by the LLM become inappropriately entangled with the low-level, fine-grained numerical dynamics of time series, making it difficult for semantic priors to effectively guide forecasting. To address this issue, we propose TimeSAF, a new framework based on hierarchical asynchronous fusion. Unlike synchronous approaches, TimeSAF explicitly decouples unimodal feature learning from cross-modal interaction. It introduces an independent cross-modal semantic fusion trunk, which uses learnable queries to aggregate global semantics from the temporal and prompt backbones in a bottom-up manner, and a stage-wise semantic refinement decoder that asynchronously injects these high-level signals back into the temporal backbone. This mechanism provides stable and efficient semantic guidance while avoiding interference with low-level temporal dynamics. Extensive experiments on standard long-term forecasting benchmarks show that TimeSAF significantly outperforms state-of-the-art baselines, and further exhibits strong generalization in both few-shot and zero-shot transfer settings.

Figures & tables

Figure 1: Comparison of strategy and performance between TimeSAF and other methods.

Figure 2: Overall architecture of TimeSAF.

Models	TimeSAF		CALF		TimeCMA		Time-FFM		UniTime		Time-LLM		GPT4TS		iTransformer		PatchTST		Crossformer		FEDformer
Metrics	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTm1	0.377	0.390	0.395	0.390	0.392	0.405	0.399	0.402	0.385	0.399	0.410	0.409	0.396	0.400	0.407	0.410	0.387	0.400	0.502	0.502	0.448	0.452
ETTm2	0.272	0.320	0.282	0.321	0.289	0.332	0.286	0.332	0.293	0.334	0.296	0.340	0.293	0.338	0.288	0.332	0.281	0.326	1.216	0.707	0.304	0.349
ETTh1	0.426	0.428	0.443	0.435	0.451	0.449	0.442	0.434	0.442	0.447	0.460	0.449	0.456	0.450	0.454	0.447	0.469	0.454	0.620	0.572	0.440	0.459
ETTh2	0.367	0.394	0.371	0.394	0.407	0.419	0.382	0.406	0.377	0.402	0.389	0.408	0.456	0.413	0.383	0.407	0.387	0.407	0.942	0.683	0.436	0.449
Weather	0.240	0.268	0.249	0.273	0.253	0.283	0.270	0.288	0.253	0.276	0.274	0.290	0.278	0.297	0.257	0.278	0.259	0.281	0.258	0.315	0.308	0.360

Table 1: Average MSE and MAE over four prediction lengths. All experiments fix the lookback length T = 96. The prediction length set is H

\in

{96, 192, 336, 720}. The best result is red , the second best result is underlined . Our full results are in Appendix A .

Models	TimeSAF		CALF		TimeCMA		Time-LLM		GPT4TS		PatchTST		Crossformer		FEDformer		TimesNet		DLinear		MICN
Metrics	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTm1	0.483	0.449	0.504	0.462	0.594	0.504	0.636	0.512	0.608	0.500	0.557	0.483	1.340	0.848	0.696	0.572	0.673	0.534	0.567	0.499	0.970	0.674
ETTm2	0.294	0.330	0.302	0.330	0.309	0.344	0.308	0.343	0.303	0.336	0.295	0.334	1.985	1.048	0.356	0.392	0.321	0.354	0.329	0.382	1.073	0.716
ETTh1	0.622	0.527	0.644	0.541	0.721	0.575	0.765	0.584	0.689	0.555	0.683	0.645	1.744	0.914	0.750	0.607	0.865	0.625	0.647	0.552	1.405	0.814
ETTh2	0.422	0.423	0.419	0.427	0.451	0.448	0.589	0.498	0.579	0.497	0.550	0.487	3.139	1.378	0.553	0.525	0.476	0.463	0.441	0.458	2.533	1.158

Table 2: Few-shot forecasting performance on ETT datasets using only 10% of the training data. All experiments fix the lookback length T = 96. The prediction horizon is set to H

\in

{96, 192, 336, 720}. Our full results are in Appendix A .

Models	TimeSAF		CALF		TimeCMA		Time-LLM		GPT4TS		PatchTST		Crossformer		FEDformer		TimesNet		DLinear		MICN
Metrics	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
h1 → m1	0.749	0.563	0.755	0.574	0.820	0.590	0.847	0.600	0.798	0.574	0.894	0.610	0.999	0.736	0.765	0.588	0.794	0.575	1.439	0.870	0.760	0.577
h1 → m2	0.313	0.355	0.316	0.355	0.329	0.365	0.315	0.357	0.317	0.359	0.318	0.362	1.120	0.789	0.357	0.403	0.339	0.370	2.428	1.236	0.399	0.439
h2 → m1	0.873	0.604	0.836	0.586	1.087	0.673	0.868	0.595	0.920	0.610	0.871	0.596	1.195	0.711	0.741	0.588	1.286	0.705	0.764	0.601	0.778	0.594
h2 → m2	0.317	0.360	0.319	0.360	0.349	0.386	0.322	0.363	0.331	0.371	0.420	0.433	2.043	1.124	0.365	0.405	0.361	0.390	0.527	0.519	0.496	0.496

Table 3: Zero-shot forecasting performance on the ETT datasets, where prediction lengths

H\in\{96,192,336,720\}

. “h1→m1” indicates that models trained on ETTh1 are evaluated on ETTm1, and similarly for the other transfer settings. Our full results are in Appendix A .

Figure 3: Ablation studies of different variants of TimeSAF on multiple datasets.

Panel of Figure 4 — Figure 4: Visualization of the proposed asynchronous fusion mechanism on the Exchange dataset. (a) Cross-attention maps from fusion queries to temporal patches in the fusion stage. (b) Cross-attention maps from temporal patches to fusion queries in the refinement stage. (c) t-SNE projection of temporal features and fusion features before refinement. (d) t-SNE projection of temporal features and fusion features after refinement.

Figure 5: Sensitivity of TimeSAF to fusion configuration.

Models		TimeSAF		CALF		TimeCMA		Time-FFM		UniTime		Time-LLM		GPT4TS		iTransformer		PatchTST		Crossformer		FEDformer
Metrics		MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTm1	96	0.311	0.351	0.319	0.348	0.330	0.371	0.336	0.369	0.322	0.363	0.359	0.381	0.335	0.369	0.334	0.368	0.329	0.367	0.360	0.401	0.379	0.419
	192	0.357	0.372	0.373	0.375	0.368	0.389	0.378	0.389	0.366	0.387	0.383	0.393	0.374	0.385	0.377	0.391	0.367	0.385	0.402	0.440	0.426	0.441
	336	0.390	0.399	0.411	0.400	0.402	0.411	0.411	0.410	0.398	0.407	0.416	0.414	0.407	0.406	0.426	0.420	0.399	0.410	0.543	0.528	0.445	0.459
	720	0.450	0.438	0.478	0.438	0.471	0.450	0.469	0.441	0.454	0.440	0.483	0.449	0.469	0.442	0.491	0.459	0.454	0.439	0.704	0.642	0.543	0.490
	avg	0.377	0.390	0.395	0.390	0.392	0.405	0.399	0.402	0.385	0.399	0.410	0.409	0.396	0.400	0.407	0.410	0.387	0.400	0.502	0.502	0.448	0.452

Table 4: All experiments fix the lookback length T = 96. The prediction length set is H

\in

{96, 192, 336, 720}. The best result is red , the second best result is underlined .

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Models		TimeSAF		CALF †		TimeCMA †		Time-LLM		GPT4TS		PatchTST		Crossformer		FEDformer		TimesNet		DLinear		MICN
Metrics		MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTm1	96	0.442	0.430	0.468	0.445	0.568	0.483	0.587	0.491	0.615	0.497	0.558	0.478	1.037	0.705	0.604	0.530	0.583	0.503	0.552	0.488	0.677	0.585
	192	0.457	0.437	0.479	0.446	0.532	0.475	0.606	0.490	0.597	0.492	0.539	0.471	1.170	0.778	0.641	0.546	0.608	0.515	0.546	0.487	0.784	0.627
	336	0.476	0.450	0.499	0.463	0.548	0.489	0.719	0.555	0.597	0.501	0.558	0.488	1.463	0.913	0.768	0.606	0.733	0.572	0.567	0.501	0.972	0.684
	720	0.559	0.479	0.572	0.496	0.729	0.571	0.632	0.514	0.623	0.513	0.574	0.498	1.693	0.997	0.771	0.606	0.768	0.548	0.606	0.522	1.449	0.800
	avg	0.483	0.449	0.504	0.462	0.594	0.504	0.636	0.512	0.608	0.500	0.557	0.483	1.340	0.848	0.696	0.572	0.673	0.534	0.567	0.499	0.970	0.674

Appendix

Table 5: Few-shot forecasting performance on ETT datasets using only 10% of the training data. The prediction horizon is set to H

\in

{96, 192, 336, 720}. We report the average MSE and MAE over all horizons; The best result is red , the second best result is underlined .

Models	Metric	TimeSAF		CALF		TimeCMA		Time-LLM		GPT4TS		PatchTST		Crossformer		FEDformer		TimesNet		MICN		DLinear
h1 → m1	96	0.729	0.544	0.767	0.564	0.809	0.577	0.804	0.565	0.809	0.563	0.908	0.596	0.856	0.649	0.731	0.561	0.764	0.563	0.832	0.621	0.735	0.554
	192	0.750	0.554	0.753	0.570	0.814	0.584	0.827	0.593	0.799	0.567	0.927	0.616	0.906	0.684	0.746	0.573	0.798	0.562	1.288	0.854	0.752	0.570
	336	0.744	0.564	0.745	0.575	0.814	0.591	0.835	0.600	0.803	0.577	0.920	0.621	1.104	0.796	0.775	0.596	0.790	0.584	1.721	0.972	0.749	0.579
	720	0.774	0.588	0.758	0.590	0.841	0.607	0.922	0.644	0.783	0.589	0.822	0.608	1.131	0.816	0.808	0.625	0.827	0.594	1.915	1.036	0.805	0.606
	avg	0.749	0.563	0.755	0.574	0.820	0.590	0.847	0.600	0.798	0.574	0.894	0.610	0.999	0.736	0.765	0.588	0.794	0.575	1.439	0.870	0.760	0.577
h1 → m2	96	0.217	0.301	0.218	0.301	0.229	0.310	0.212	0.298	0.218	0.304	0.219	0.305	0.611	0.588	0.257	0.345	0.245	0.322	0.496	0.556	0.239	0.343

Appendix

Table 6: Full results for zero-shot forecasting on the ETT datasets, where prediction lengths

H\in\{96,192,336,720\}

. “h1→m1” indicates that models trained on ETTh1 are evaluated on ETTm1, and the same applies to other items. The best result is red , the second best result is underlined .

Dataset	Variables	Frequency	Scope	Time Range	Predicted Window
ETTh1, ETTh2	7	1 Hour	Energy	2016/07 – 2018/07	{96,192,336,720}
ETTm1, ETTm2	7	15 Mins	Energy	2016/07 – 2018/07	{96,192,336,720}
Electricity	321	1 Hour	Energy	2012 – 2014	{96,192,336,720}
Weather	21	10 Mins	Weather	2020 Whole Year	{96,192,336,720}
Exchange	8	1 Day	Finance	1990 – 2016	{96,192,336,720}

Appendix

Table 7: Dataset descriptions. Variables denotes the dimension of the multivariate time series. Frequency indicates the sampling interval.

Figure 6: Hint templates for specific datasets are used to transcribe multivariate time series segments into natural language descriptions. Each template jointly encodes the time interval, numerical sequence, sampling frequency, and trend statistics to establish a unified representation between structured time series data and language model input.

Dataset	$H$	Time-SAF	Domain	Timestamp	Instruction
ETTh2	96	0.286	0.289	0.292	0.291
	192	0.370	0.375	0.384	0.376
	336	0.391	0.398	0.390	0.395
	720	0.425	0.428	0.426	0.429
Exchange	96	0.078	0.081	0.085	0.080
	192	0.170	0.175	0.178	0.173

Appendix

Table 8: Prompt variant ablation (MSE; lower is better).

Transfer Setting	$H$	Time-SAF		Time-FFM
		MSE	MAE	MSE	MAE
ETTm2 $\rightarrow$ Electricity	96	0.521	0.581	0.616	0.613
	192	0.345	0.335	0.649	0.631
	336	0.486	0.478	0.687	0.651
	720	0.612	0.597	0.736	0.675
ETTm1 $\rightarrow$ Weather	96	0.235	0.268	0.235	0.270

Appendix

Table 9: Cross-dataset zero-shot transfer results (lower is better).

Dataset	$H$	Fusion Trunk	Trunk-Decoder
ETTm1	96	0.311	0.320
	192	0.357	0.363
	336	0.390	0.401
	720	0.450	0.462
Weather	96	0.157	0.162
	192	0.204	0.213

Appendix

Table 10: Structural comparison between Fusion Trunk and Trunk-Decoder (MSE; lower is better).

LLM-Based Time Series Forecasting

LLM-based time series forecasting uses large language models to predict future values from numerical histories and contextual data.

See all 100 papers

FreSia: Frequency-Semantic Instantiation and Alignment for Multivariate Time Series Analysis
Oct 5, 2026Yubo Wang, Hui He, Hezhe Qiao +2Multivariate Time Series Forecasting Frequency-Domain Time Series Modeling
School of Computer Science and Technology, Beijing Institute of Technology, No.5 Zhongguancun South Street, Beijing, 100081, China · School of Computing and Information Systems, Singapore Management University, 80 Stamford Road, 178902, Singapore
Markovian Linguistic-Temporal Bridge: Unlocking the Potential of LLMs for Time Series Forecasting
May 19, 2025Siming Sun, Kai Zhang, Xuejun Jiang +2Time Series Forecasting LLM-Based Time Series Forecasting
Zhejiang University
FM-LLM: A frequency-enhanced mixture-of-experts framework for adapting LLMs to time series forecasting
Aug 12, 2026Rentao Gu, Yihang Ding, Junjie Li +5Zero-Shot Time Series Forecasting Time Series Forecasting
Beijing University of Posts and Telecommunications, Beijing, 100876, China · China Telecom Research Institute, Beijing, 100032, China

Oct 5, 2026J/K move · Enter open · S save

FreSia: Frequency-Semantic Instantiation and Alignment for Multivariate Time Series Analysis

Yubo Wang, Hui He, Hezhe Qiao, Guoqing Ji+1

School of Computer Science and Technology, Beijing Institute of Technology, No.5 Zhongguancun South Street, Beijing, 100081, China · School of Computing and Information Systems, Singapore Management University, 80 Stamford Road, 178902, Singapore

Multivariate Time Series Forecasting Frequency-Domain Time Series Modeling Time Series Forecasting LLM-Based Time Series Forecasting

Open paper AskPDF arXiv

Large Language Models (LLMs) have shown strong potential in multivariate time series forecasting and anomaly detection. Existing studies predominantly inject temporal information into LLMs via direct numerical tokenization or heuristic textual descriptions. However, LLMs still face difficulty in perceiving the underlying structural patterns of numerical time series, particularly the seasonal and trend components obscured by discrete numerical tokens. To bridge this gap, we propose FreSia, a frequency-aware framework that establishes an effective alignment between the semantic space of LLMs and the frequency space of time series. Specifically, FGPrompt, a Frequency-Guided Prompt mechanism within FreSia, distills the frequency-domain structures of time series and projects them into prompts tailored to the semantic space of LLMs. Furthermore, we introduce a Global-driven Context Learning (GCL) component, which uses a global CLS-driven probe to generate global context to bridge the time-frequency domain gap and fuse the multi-modal information. Experiments on eight forecasting benchmarks show that FreSia achieves average improvements of 13.48% and 8.06% in MSE and MAE, respectively.

Abstract

LLM-Based Time Series Forecasting

Explore similar work

FreSia: Frequency-Semantic Instantiation and Alignment for Multivariate Time Series Analysis

Markovian Linguistic-Temporal Bridge: Unlocking the Potential of LLMs for Time Series Forecasting

FM-LLM: A frequency-enhanced mixture-of-experts framework for adapting LLMs to time series forecasting