WorldTS: World Modeling for Multimodal Covariate-aware Time Series Forecasting

Yuhan Zhu; Xiangfei Qiu; Hanyin Cheng; Wangmeng Shen; Chenjuan Guo; Bin Yang; Jilin Hu; Christian S. Jensen

cs.LGSep 25, 2026

WorldTS: World Modeling for Multimodal Covariate-aware Time Series Forecasting

Authors: Yuhan Zhu, Xiangfei Qiu, Hanyin Cheng, Wangmeng Shen, Chenjuan Guo, Bin Yang, Jilin Hu, Christian S. Jensen

Organizations: East China Normal University · Aalborg University

Abstract

Time series forecasting is typically framed as learning a direct mapping from historical to future observations in the observation space. However, sequences of observations generally provide only a partial view of the dynamics of the underlying system, with future observations being shaped by latent dynamics. Recent latent-space forecasting methods thus achieve improved performance by predicting future observations from latent-space representations of historical observations rather than directly forecasting future observations in the observation space. Next, while future observations are also shaped by external factors, how to incorporate external, often multimodal, information into forecasting, so that it can shape latent-state formation and evolution directly, remains underexplored. We propose WorldTS, a world-modeling based forecasting framework that integrates multimodal covariates directly into the forecasting to further improve forecasting performance. Specifically, WorldTS employs a two-stage training strategy. First, it learns forecasting-relevant latent state dynamics conditioned on multimodal covariates, yielding encoded future states. Next, the learned state dynamics are frozen, and an observation decoder is trained to map the predicted future states back to future observations. Extensive experiments on 21 real-world datasets offer insight into WorldTS and its effectiveness.

Figures & tables

Figure 1: Comparison of observation-space and latent-space forecasting. (a) Observation-space forecasting supervised by ground-truth future observations. (b) Latent-space forecasting supervised by states encoded from the ground-truth future observations. (c) Forecasts on the Multimodal Solar Power (MMSP) dataset by WorldTS and its observation-space variant (Obs-space). (d) Latent distributions on MMSP, comparing predicted future states with states encoded from the ground-truth future observations.

Figure 2: Overview of WorldTS. TS Encoder/Decoder denote the time-series encoder/decoder; Predictor denotes the covariate-conditioned state predictor. Cov Encoder and Modal Adapter are the numerical-covariate encoder and modality-specific MLP projector. Stage 1 (Latent Dynamics Learning) learns future-state prediction conditioned on multimodal covariates. Stage 2 (Observation Decoding) freezes the encoders and predictor and trains the decoder to recover future observations.

Figure 3: Covariate-conditioned state predictor. Historical covariates are integrated into the historical states through direct fusion or CFA, while future covariates further refine the predicted states.

Models	WorldTS (ours)		KITE (2026)		DAG (2026)		GCGNet (2026)		DUET (2025)		CrossLinear (2025)		Amplifier (2025)		TimeKAN (2025)		TimeXer (2024)		TiDE (2023)		TFT (2021)
Metrics	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae
NP	0.297	0.319	0.325	0.323	0.362	0.344	0.370	0.348	0.411	0.407	0.371	0.387	0.420	0.418	0.405	0.419	0.418	0.371	0.443	0.400	0.379	0.375
PJM	0.087	0.178	0.096	0.178	0.093	0.180	0.095	0.187	0.102	0.197	0.112	0.223	0.137	0.246	0.139	0.262	0.108	0.198	0.142	0.246	0.114	0.207
BE	0.393	0.260	0.427	0.286	0.423	0.279	0.431	0.294	0.514	0.354	0.479	0.337	0.559	0.413	0.548	0.407	0.452	0.290	0.498	0.325	0.454	0.291
FR	0.384	0.210	0.387	0.225	0.414	0.219	0.415	0.234	0.495	0.327	0.482	0.298	0.554	0.408	0.547	0.374	0.427	0.241	0.484	0.281	0.504	0.257
DE	0.340	0.353	0.350	0.370	0.370	0.370	0.401	0.389	0.482	0.430	0.485	0.452	0.473	0.441	0.473	0.445	0.475	0.418	0.499	0.447	0.489	0.446

Table 1: Average results over short and long forecasting horizons on the 12 numerical-covariate datasets. Baseline results are from DAG ( Qiu et al., 2026b ) and KITE ( Cheng et al., 2026e ) under the same evaluation protocol. Full results: Table 5 (Appendix B ). Red : best; blue : second best.

Models	WorldTS (ours)		VoT (2026)		TaTS (2026)		GPT4TS (2025)		Time-VLM (2025)		CALF (2025)		GPT4MTS (2024)		RaFT (2025)		iTransformer (2024)		PatchTST (2023)
Metrics	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae
Agriculture	0.285	0.338	0.288	0.355	0.290	0.350	0.291	0.338	0.322	0.359	0.314	0.355	0.301	0.342	0.301	0.383	0.300	0.352	0.309	0.351
Climate	1.072	0.838	1.088	0.845	1.179	0.885	1.171	0.883	1.203	0.896	1.177	0.883	1.152	0.876	1.295	0.930	1.131	0.864	1.168	0.877
Economy	0.214	0.363	0.223	0.374	0.229	0.381	0.233	0.386	0.245	0.396	0.216	0.363	0.226	0.379	0.280	0.425	0.230	0.385	0.229	0.382
Energy	0.090	0.201	0.091	0.218	0.105	0.232	0.111	0.243	0.114	0.253	0.102	0.224	0.111	0.244	0.123	0.255	0.121	0.258	0.107	0.235
Health	0.778	0.542	0.898	0.596	0.899	0.612	0.854	0.629	1.198	0.727	0.964	0.609	0.985	0.658	1.297	0.769	1.261	0.746	1.006	0.650

Table 2: Forecasting results on eight Time-MMD datasets with a prediction horizon of 12. Baseline results are taken from VoT ( Wang et al., 2026 ) under the same evaluation setting.

Category	Ours	Multimodal		Specialized Forecasters					Pretrained Models			UniCA-adapted
Models	WorldTS	MM-TSF	FusionSF	TFT	D.AR	NB.S	TiDE	P.TST	TTM-R2	Moirai	PFN-TS	Chronos-Bolt	TimesFM	Time-MoE
MSE	0.061	0.090	0.478	0.067	0.097	0.107	0.297	0.662	0.095	0.206	0.633	0.090	0.098	0.653
MAE	0.171	0.200	0.566	0.177	0.216	0.219	0.438	0.682	0.263	0.378	0.711	0.193	0.229	0.765

Table 3: Forecasting results on the MMSP with a prediction horizon of 24. Baseline results are taken from UniCA ( Han et al., 2026 ) under the same evaluation setting.

Figure 4: Parameter sensitivity of WorldTS. We vary the weights of three losses (

\lambda_{\mathrm{latent}}

\lambda_{\mathrm{sim}}

, and

\lambda_{\mathrm{VIC}}

), the latent-state dimensionality, and the symmetric causal patch length one at a time on Colbun, Climate, and MMSP, and report MAE. Other settings remain fixed.

Dataset [-1pt](MSE / MAE)	Full	w/o $\mathcal{L}_{\mathrm{latent}}$	w/o $\mathcal{L}_{\mathrm{sim}}$	w/o $\mathcal{L}_{\mathrm{VICReg}}$	Pointwise MLP	One-stage MAE	w/o Covariates
Colbun	0.052 / 0.080	0.056 / 0.088	0.055 / 0.081	0.053 / 0.083	0.066 / 0.091	0.059 / 0.096	0.073 / 0.088
Climate	1.072 / 0.838	1.145 / 0.871	1.089 / 0.852	1.111 / 0.852	1.086 / 0.844	1.149 / 0.875	1.134 / 0.869
MMSP	0.061 / 0.171	0.078 / 0.199	0.063 / 0.176	0.068 / 0.188	0.065 / 0.187	0.085 / 0.217	0.101 / 0.279

Table 4: Ablation results on representative datasets with numerical, textual, and visual covariates.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: TS Encoder and Decoder. Overlapping causal observation patches yield time-aligned states, while local state patches reconstruct target observations.

Models		WorldTS (ours)		KITE (2026)		DAG (2026)		GCGNet (2026)		DUET (2025)		CrossLinear (2025)		Amplifier (2025)		TimeKAN (2025)		TimeXer (2024)		TiDE (2023)		TFT (2021)
Metrics		mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae	mse	mae
NP	24	0.171	0.217	0.179	0.223	0.202	0.237	0.208	0.237	0.246	0.287	0.210	0.266	0.252	0.303	0.273	0.310	0.236	0.266	0.284	0.301	0.219	0.249
	360	0.423	0.420	0.471	0.422	0.521	0.451	0.531	0.459	0.576	0.528	0.531	0.508	0.587	0.534	0.538	0.529	0.600	0.475	0.601	0.498	0.539	0.501
	Avg	0.297	0.319	0.325	0.323	0.362	0.344	0.370	0.348	0.411	0.407	0.371	0.387	0.420	0.418	0.405	0.419	0.418	0.371	0.443	0.400	0.379	0.375
PJM	24	0.056	0.141	0.056	0.143	0.057	0.143	0.060	0.150	0.072	0.166	0.088	0.191	0.096	0.208	0.115	0.244	0.075	0.166	0.106	0.214	0.095	0.195
	360	0.118	0.216	0.135	0.214	0.130	0.218	0.129	0.223	0.131	0.228	0.135	0.254	0.177	0.285	0.162	0.281	0.140	0.231	0.177	0.279	0.133	0.219

Appendix

Table 5: Full results on the 12 numerical-covariate datasets. Avg denotes the average result over the two forecasting horizons. Baseline results follow the same evaluation protocol as Table 1 . The best results are in red and bold , and the second-best results are blue and underlined .

Explore similar work

DiTS: Multimodal Diffusion Transformers Are Time Series Forecasters
Feb 6, 2026Haoran Zhang, Haixuan Liu, Yong Liu +4Time Series Forecasting Probabilistic Time Series Forecasting
School of Software, BNRist, Tsinghua University
Pythia: Toward Foundation World Models for Multimodal Time Series
Oct 4, 2026Xilin Dai, Hongzhou Chen, Yifan Hu +3Multivariate Time Series Forecasting World Model Learning
Ant International
Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting
May 28, 2026Haoxin Liu, Yichen Zhou, Rajat Sen +2Multivariate Time Series Forecasting LLM Post-Training
Georgia Institute of Technology · Google Research

Feb 6, 2026, cs.LGJ/K move · Enter open · S save

DiTS: Multimodal Diffusion Transformers Are Time Series Forecasters

Haoran Zhang, Haixuan Liu, Yong Liu, Yunzhong Qiu+3

School of Software, BNRist, Tsinghua University

Time Series Forecasting Probabilistic Time Series Forecasting Diffusion Transformer Multimodal Diffusion Models

Open paper AskPDF arXiv

While generative modeling facilitates probabilistic time series forecasting, incorporating heterogeneous exogenous information remains challenging. Diffusion Transformers (DiT) provide a scalable framework for conditional generation, yet their adaptation to forecasting calls for conditioning mechanisms tailored to time series. Endogenous targets and exogenous covariates differ in sources, semantics, and statistical characteristics, while sharing temporal coordinates that support fine-grained conditional guidance. Covariates can describe future variability and temporal dependence beyond the conditional mean targeted by direct regression. Motivated by these considerations, we propose Diffusion Transformers for Time Series (DiTS), a Multimodal Diffusion Transformer for covariate-aware forecasting. DiTS models endogenous targets and exogenous covariates as distinct modalities, jointly conditioning future generation on target history and available covariates. Flow matching makes covariate-dependent distributional information relevant to velocity prediction conditioned on noisy future states. We introduce Time-aligned Modulation, extending AdaLN with the temporal-alignment prior to generate patch-wise modulation parameters from aligned covariates and diffusion time. Across diverse covariate-aware forecasting tasks, DiTS achieves strong performance in both deterministic and probabilistic forecasting, demonstrating the effectiveness of conditional generation for both point and distributional forecasting.

Abstract

Explore similar work

DiTS: Multimodal Diffusion Transformers Are Time Series Forecasters

Pythia: Toward Foundation World Models for Multimodal Time Series

Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting