Organizations: Institute of Advanced Technology, University of Science and Technology of China, Hefei, China · School of Computer Science and Technology, University of Science and Technology of China, Hefei, China · Department of Automation, University of Science and Technology of China, Hefei, China
Irregular multivariate time series forecasting supports healthcare monitoring, environmental analysis, and activity recognition. Channel-independent (CI) approaches capture each variable's temporal patterns but do not explicitly use observations from other variables. Channel-dependent (CD) modeling exploits relationships among variables, yet asynchronous measurements make these relationships difficult to learn. Incorporating cross-variable information while preserving each variable's irregular history therefore remains a key challenge. We propose Temporal-Reference Complementary Forecasting (TRCF), a framework that augments a compact CI forecaster with a complementary CD branch. ANCHOR aggregates observed values and timestamps at shared temporal locations, retaining observation support to form compact CI features. BRIDGE constructs complementary features through history completion. It first estimates each variable's historical values from its own observations, then uses differences between other variables' observed and estimated values at nearby times to correct missing entries. The completed history preserves observed values and is encoded to supply context alongside the direct CI representation. Finally, the two representations are fused, combined with the query-time embedding, and passed to a shallow multilayer perceptron for forecasting. Experiments on four real-world datasets show that TRCF achieves the lowest mean squared error and mean absolute error in the main benchmark comparison, while reducing measured peak inference memory by 17.28% to 87.98% relative to APN.
Figures & tables
Fig. 1: Two core challenges in IMTS forecasting: (a) uneven observation support and (b) asynchronous cross-variable context.
Fig. 2: The overall framework of TRCF. Given an irregular multivariate history, (a) ANCHOR aggregates observations into shared temporal slots with explicit support features to form direct representations, while (b) BRIDGE constructs an aligned history via bidirectional temporal references, nearest-residual relay, and target-excluded cross-variable correction. The two representations are summed and concatenated with the query-time embedding in the forecasting decoder. (c) BRIDGE relays residuals along time, excludes the target residual from cross-variable correction, and restores observed values.
Dataset
Variables
Samples
History window
Scored targets
HA
12
1,360
3 s
5,910
USHCN
5
1,114
∼ 3 years
368
P12
36
11,981
36 h
27,008
MIMIC
96
21,250
36 h
23,015
TABLE I: Dataset statistics.
Fig. 3: Accuracy and inference cost on P12 and MIMIC. Panel-specific memory keys define bubble sizes.
Fig. 4: Peak inference memory of APN and TRCF across four datasets. Percentages show the reduction relative to APN.
Dataset
Adaptive
Adaptive
ANCHOR
Original P
P=16
K=16
HA
0.04212
0.04216
0.04088
USHCN
0.15397
0.16027
0.13654
P12
0.29315
0.29241
0.28842
MIMIC
0.42319
0.42548
0.42222
TABLE III: Comparison of direct encoders on four datasets (MSE).
Model
HA
USHCN
P12
MIMIC
TRCF
0.04070
0.14305
0.28721
0.40428
w/o ANCHOR
0.04960
0.14196
0.29726
0.41416
w/o BRIDGE
0.04088
0.13654
0.28842
0.42222
TABLE IV: Ablation study of the dual complementary paths (MSE).
Model
P12
MIMIC
MSE
MAE
MSE
MAE
Widened ANCHOR
0.2891
0.3530
0.4240
0.3908
Interpolated context
0.2881
0.3524
0.4195
0.3907
Temporal context
0.2876
0.3523
0.4154
0.3894
TRCF
0.2863
0.3521
0.4015
0.3820
TABLE V: Comparison with capacity-matched controls on P12 and MIMIC.
Model
P12
MIMIC
MSE
MAE
MSE
MAE
TRCF
0.28721
0.35160
0.40428
0.38160
w/o cross correction
0.28808
0.35244
0.41849
0.38984
w/o residual relay
0.28752
0.35185
0.41192
0.38469
TABLE VI: Ablation study of core components in BRIDGE.
Fig. 5: Sensitivity analysis of main hyperparameters (slot count K , grid resolution G , and fusion scale β ).
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
HA
USHCN
P12
MIMIC
Variables N
12
5
36
96
Direct width D
56
6
24
16
Time width dt
8
32
8
16
Token width dv
32
16
16
32
Batch size
16
32
32
32
Learning rate
0.01
0.01
0.03
0.02
Appendix
TABLE VII: Dataset-specific TRCF settings.
Dataset
Model
Parameters
MiB
HA
APN
28,789
120.10
TRCF
36,292
14.44
USHCN
APN
1,970
74.13
TRCF
6,843
21.09
P12
APN
6,701
31.20
KAFNet
65,775
16.21
Appendix
TABLE VIII: Inference memory and parameter counts.
Model
Metric
HA
USHCN
P12
MIMIC
Direct-only reference
ANCHOR
Params
21,649
987
5,265
4,113
A. Matched capacity, λh=0
Widened ANCHOR
Params
36,339
6,850
12,174
19,359
MSE
0.0408
0.1395
0.2891
0.4240
MAE
0.1126
0.2476
0.3530
0.3908
Appendix
TABLE IX: Capacity-matched context and supervision controls.
Dataset
Encoder, slots
Params
MSE
MAE
HA
Adaptive, original
28,789
0.04212
0.11564
Adaptive, 16
21,973
0.04216
0.11561
ANCHOR, 16
21,649
0.04088
0.11269
USHCN
Adaptive, original
1,970
0.15397
0.26083
Adaptive, 16
1,130
0.16027
0.26602
ANCHOR, 16
987
0.13654
0.24976
Appendix
TABLE X: Direct-encoder accuracy and parameter counts.
Model
HA
USHCN
P12
MIMIC
MSE
TRCF
0.04070
0.14305
0.28721
0.40428
w/o ANCHOR
0.04960
0.14196
0.29726
0.41416
w/o BRIDGE
0.04088
0.13654
0.28842
0.42222
MAE
TRCF
0.11235
0.25237
0.35160
0.38160
Appendix
TABLE XI: Ablation study of the complementary paths on four datasets.
Fig. 6: Ablation of temporal aggregation and observation support.
Fig. 7: Forecasting with reduced historical observation support.
Fig. 8: Forecasting accuracy and model size with different ANCHOR widths.
A. Variable attention and historical supervision
Configuration
USHCN
MIMIC
MSE
MAE
MSE
MAE
TRCF
0.1431
0.2524
0.4043
0.3816
Diagonal attention
0.1433
0.2524
0.4049
0.3812
λh=0
0.1431
0.2527
0.4015
0.3820
λh=0.03
0.1432
0.2524
0.4038
0.3817
Appendix
TABLE XII: Variable attention and training objectives.
Content / sources
HA
USHCN
P12
MIMIC
MSE
Raw / observed
0.04080
0.14258
0.28667
0.40935
Raw / nearest
0.04073
0.14271
0.28574
0.40283
Residual / observed
0.04071
0.14298
0.28752
0.41192
Residual / nearest
0.04070
0.14305
0.28721
0.40428
MAE
Appendix
TABLE XIII: Temporal reference and source propagation.
Model
HA
USHCN
P12
MIMIC
MSE
TRCF
0.04070
0.14305
0.28721
0.40428
w/o correction
0.04073
0.14289
0.28808
0.41849
w/o residual relay
0.04071
0.14298
0.28752
0.41192
Model-filled raw
0.04073
0.14272
0.28627
0.40846
MAE
Appendix
TABLE XIV: BRIDGE component removals and model-filled raw completion.
Irregularly sampled multivariate time series arise in many real-world domains such as healthcare, climate science, and large-scale monitoring systems. Characterized by asynchronous observations and heterogeneity across variables, such time series pose significant challenges for forecasting. In this work, we introduce Cross-Variable Temporal Attention (CVTA), a short-term, continuous-time forecasting framework that learns prediction functions from its compact, interpretable state summarizing recent temporal dynamics across variables. CVTA uses Variable-Conditioned Embeddings to project heterogeneous variables into variable-specific latent spaces, and applies temporal attention to capture cross-variable dependencies from this state. Furthermore, we demonstrate that CVTA can be integrated as an auxiliary module into existing state-of-the-art prediction models through a dynamic meta-decision model. Extensive experiments on multiple real-world benchmarks show that CVTA achieves strong short-term predictive accuracy, supports efficient online inference, improves existing prediction models when used as an auxiliary module, and learns attention patterns that align well with domain knowledge.
Ankitkumar Joshi, Milos Hauskrecht
Department of Computer Science, University of Pittsburgh, Pittsburgh, PA 15260
Multivariate time series forecasting often struggles to capture long-range dependencies due to fixed lookback windows. Retrieval-augmented forecasting addresses this by retrieving historical segments from memory, but existing approaches rely on a channel-agnostic strategy that applies the same references to all variables. This neglects inter-variable heterogeneity, where different channels exhibit distinct periodicities and spectral profiles. We propose CRAFT (Channel-wise retrieval-augmented forecasting), a novel framework that performs retrieval independently for each channel. To ensure efficiency, CRAFT adopts a two-stage pipeline: a sparse relation graph constructed in the time domain prunes irrelevant candidates, and spectral similarity in the frequency domain ranks references, emphasizing dominant periodic components while suppressing noise. Experiments on seven public benchmarks demonstrate that CRAFT outperforms state-of-the-art forecasting baselines, achieving superior accuracy with practical inference efficiency.
Irregular multivariate time series forecasting is a challenging yet important problem in real-world applications, where observations are often irregularly sampled and asynchronously recorded across variables. Existing time series foundation models are mostly built on regularly sampled sequences, making them difficult to generalize to irregular time intervals and asynchronous cross-variable dependencies. To address these challenges, we propose QiYao-I, a manifold based foundation model for irregular multivariate time series forecasting. Specifically, we introduce a novel sampling-conditioned temporal manifold attention mechanism that maps real timestamps into a learnable temporal manifold feature space and injects temporal manifold biases into attention layers, enabling the model to capture both irregular time intervals and local sampling structures. Further, we propose a dynamic variable interaction mechanism with frequency awareness. It selectively performs cross-variable message passing under asynchronous observations. Extensive experiments on real-world irregular multivariate forecasting benchmarks demonstrate that QiYao-I achieves superior performance compared with both time series foundation models and end-to-end irregular forecasting models, showing strong generalization ability in zero-shot and few-shot settings.
Linfeng Wang, Ruitong Zhang, Kai Zhao +6
East China Normal University · Huawei Technologies Ltd.