Organizations: Institute of Advanced Technology, University of Science and Technology of China, Hefei, China · Department of Automation, University of Science and Technology of China, Hefei, China
A fixed number of recent observations may span minutes for one variable and days for another. Earlier measurements can still reveal whether the latest value is unusual for that variable. We propose TRESS (Timestamped Recent Events and Statistical Summaries), a forecasting model that retains the latest observations at their original times and summarises the full history of each variable. For a given set of variables and event budget, this representation stays the same size as the record grows. Its query-conditioned event lifting module weights deviations from the latest value according to the age of each observation and the forecast horizon. A channel-specific decoder combines the resulting features with historical statistics and a summary shared across variables to predict a correction to the latest value. Retaining 8 to 16 observations per variable, TRESS records the lowest mean test MSE and MAE among the methods compared on all four benchmarks. Serving 32 successive query times from one encoded history reduces CPU inference time by a factor of 2.12--2.60 across the four datasets. The code is at https://github.com/SSMa913/NICLab-TRESS.
Figures & tables
Figure 1. Observation timing in irregular histories. In the constructed example, numbers inside circles give the observed values. The brackets measure elapsed time from each channel’s last observation to the first query. The right panel summarizes the training histories. Each channel observed at least once in a record contributes one age, measured from its latest observation to the common history endpoint. Age is expressed relative to the record’s historical span S , as defined in Section 3.1 . Two example channels share the same values at different times. A bar chart reports differences in the age of the latest observation across four training sets.
Figure 2. MSE on each dataset (Table 1 ) and computational costs averaged arithmetically over four datasets. Numbers show actual values; K/M/G denote 103/106/109 . Times are milliseconds per batch ( b=32 ); training includes forward, backward, and Adam. Memory is extra peak inference allocation in MiB. Matrix FLOPs count profiled matrix products and convolutions. Each axis is independently scaled in log space, with lower values farther out. Figure 5 reports the costs for each dataset. Radar chart comparing five models on test MSE for each dataset and on five costs averaged across datasets.
Figure 3. TRESS forecasts variable c at query time q from historical values X , mask M , and timestamps t . QEL multiplies deviations dc from the latest value ac by event weights ωc(q) , both over slots 2,…,K , then combines them with log lags, support indicators uc , and eight statistics mc . History and query projections are added to obtain rc(q) ; query features ηc(q) derive from q , the history endpoint t0 , span S , and event lags. CJD forms hc(q) through channel modulation and a residual block. Its scalar gate αc(q) modulates context z projected from joint descriptors p . The readout combines hc(q) , four direct statistics mc(4) , and gated context zc(q) , with a unit prior on the latest-value weight. Dashed module outlines group computations; dashed event slots denote padding. Two connected modules encode recent observations relative to each forecast time and decode them with statistics and gated cross-variable information.
Dataset
HumanActivity
USHCN
PhysioNet †
MIMIC
Metric
MSE ↓
MAE ↓
MSE ↓
MAE ↓
MSE ↓
MAE ↓
MSE ↓
MAE ↓
PrimeNet
4.2507 ± 0.0041
1.7018 ± 0.0011
0.4930 ± 0.0015
0.4954 ± 0.0018
0.8128 ± 0.00004
0.6977 ± 0.00006
0.9073 ± 0.0001
0.6614 ± 0.0001
NeuralFlows
0.1722 ± 0.0090
0.3150 ± 0.0094
0.2087 ± 0.0258
0.3157 ± 0.0187
0.3910 ± 0.00303
0.4360 ± 0.00201
0.6085 ± 0.0101
0.5306 ± 0.0066
CRU
0.1387 ± 0.0073
0.2607 ± 0.0092
0.2168 ± 0.0162
0.3180 ± 0.0248
0.5917 ± 0.00340
0.5604 ± 0.00199
0.5895 ± 0.0092
0.5151 ± 0.0048
mTAN
0.0993 ± 0.0026
0.2219 ± 0.0047
0.5561 ± 0.2020
0.5015 ± 0.0968
0.3711 ± 0.00459
0.4231 ± 0.00356
0.9408 ± 0.1126
0.6755 ± 0.0459
SeFT
1.3786 ± 0.0024
0.9762 ± 0.0007
0.3345 ± 0.0022
0.4083 ± 0.0084
0.7912 ± 0.00223
0.6860 ± 0.00079
0.9230 ± 0.0015
0.6628 ± 0.0008
Table 1. Forecasting performance on four IMTS datasets (mean ± standard deviation). Bold and underline mark the lowest and second-lowest reported means. † denotes PhysioNet with a repaired test set, including reproduced baselines. Other baseline entries follow the reference comparison. TRESS results are averaged over five training seeds.
Dataset
Train
Validation
Test
C
K
H
R
κ
Learning rate
Weight decay
HumanActivity
949
193
218
12
8
12
12
0.50
0.003
0.001
USHCN
901
101
112
5
9
16
15
0.50
0.010
0.075
PhysioNet
9,703
1,079
1,199
36
16
20
12
1.00
0.001
0.001
MIMIC
17,212
1,913
2,125
96
9
16
9
0.25
0.001
0.001
Table 2. Data partitions and TRESS configurations. C counts variables, K is the event budget, H and R are the widths of the local and context representations, and κ controls the initial event gate.
Dataset
HumanActivity
USHCN
PhysioNet †
MIMIC
Metric
MSE ↓
MAE ↓
MSE ↓
MAE ↓
MSE ↓
MAE ↓
MSE ↓
MAE ↓
w/o local representation
0.0427 ± 0.0000
0.1172 ± 0.0003
0.1495 ± 0.0000
0.2595 ± 0.0001
0.2902 ± 0.0003
0.3531 ± 0.0002
0.3985 ± 0.0007
0.3826 ± 0.0013
w/o direct statistics
0.0408 ± 0.0004
0.1166 ± 0.0010
0.1889 ± 0.0018
0.2978 ± 0.0048
0.2836 ± 0.0004
0.3524 ± 0.0020
0.3892 ± 0.0014
0.3778 ± 0.0010
w/o gated context
0.0408 ± 0.0000
0.1132 ± 0.0001
0.1513 ± 0.0001
0.2596 ± 0.0003
0.2873 ± 0.0002
0.3512 ± 0.0003
0.4048 ± 0.0004
0.3806 ± 0.0003
TRESS
0.0405 ± 0.0001
0.1142 ± 0.0003
0.1513 ± 0.0002
0.2597 ± 0.0003
0.2844 ± 0.0002
0.3502 ± 0.0002
0.3862 ± 0.0008
0.3740 ± 0.0004
Table 3. Component ablations on four IMTS datasets (test mean ± sample standard deviation over five seeds). Each variant removes one of the three readout feature groups and uses TRESS’s selected hyperparameters. Removing direct statistics also removes the latest-value coordinate and its unit coefficient prior. Values are rounded to four decimal places. † marks the repaired PhysioNet test set.
Persistence
TRESS
Dataset
Targets
MSE
MAE
MSE
MAE
HumanActivity
5,910
0.0605
0.1225
0.0405
0.1142
USHCN
368
0.3640
0.3291
0.1513
0.2597
PhysioNet
27,008
0.3985
0.3919
0.2844
0.3502
MIMIC
23,015
0.5702
0.4254
0.3862
0.3740
Table 4. Five-seed test errors against the persistence forecast. The target count is the number of observed scalar forecast values.
Figure 4. Test MSE of the retrained model without context minus that of TRESS. Positive values favor the full model. A recent variable was last observed less than one quarter of the history span before its endpoint, an older variable was last observed at least one quarter-span before the endpoint, and an unseen variable has no history in that record. Bars are 95% paired record-bootstrap intervals. Panels use separate horizontal scales. Paired changes in MSE by age of the last observation for two clinical cohorts.
Figure 5. Measured computational costs on an RTX 3090 at batch size 32; lower is better. Each panel uses a logarithmic vertical axis. Training-step time includes forward, masked MSE, backward, and Adam; inference starts from raw histories and includes model-specific input conversion. Matrix FLOPs count profiled matrix products and convolutions rather than all operations. Memory is the extra eager peak beyond resident model and input tensors. Section 5.7 gives the measurement protocol. Five panels compare parameters, matrix operations, training time, inference time and extra peak memory by dataset and model.
TRESS, earlier-observation retention p
Dataset
1
1/2
1/4
0
Persistence
HumanActivity
0.0405
0.0430
0.0454
0.0576
0.0605
USHCN
0.1513
0.1541
0.1680
0.2812
0.3640
PhysioNet
0.2844
0.2913
0.3010
0.3450
0.3985
MIMIC
0.3862
0.3955
0.4064
0.4131
0.5702
Table 5. Test MSE after independently thinning earlier observations. Each variable retains its latest measurement. The persistence column applies at every retention probability p .
Figure 6. Time to return forecasts for eight records as the number of requested times per record increases. Joint execution batches the times. Reuse history and separate execution issue them one at a time. Points are medians over three batches on a single CPU thread. Both axes use logarithmic scales. Latency for separate calls, cached histories, joint execution, and APN at increasing query counts in four datasets.
Dataset
Full
Recent
Capped
HumanActivity
0.0405 / 0.1142
0.0404 / 0.1134
0.0404 / 0.1134
USHCN
0.1513 / 0.2597
0.2289 / 0.3096
0.2288 / 0.3096
PhysioNet
0.2844 / 0.3502
0.2845 / 0.3497
0.2845 / 0.3498
MIMIC
0.3862 / 0.3740
0.3860 / 0.3735
0.3860 / 0.3737
Table 6. Test MSE / MAE when the window used for statistics changes. Full uses the complete history. Recent uses the latest K values with the original count. Capped also limits the count to K . Entries average the same five training seeds.
Figure 7. Validation MSE as one parameter varies around the selected TRESS configuration. Columns correspond to datasets. Lines show three-seed means, bands span one sample standard deviation, and stars mark the settings used by TRESS. The upper row overlays K , H , and R on a common numerical axis. The lower row overlays κ (bottom axis, linear) and learning rate (top axis, logarithmic, in units of 10−3 ), sharing one MSE axis. Validation MSE curves for five hyperparameters on four datasets. Stars mark the settings used to evaluate TRESS.
TRESS
APN
Dataset
L
Full
Encode
Decode
Full
PhysioNet
36
1.73
0.71
0.98
5.68
72
1.84
0.84
0.97
8.54
144
2.06
1.06
0.97
16.75
288
2.45
1.44
0.98
31.49
MIMIC
72
2.62
1.27
1.29
43.16
Table 7. Median inference time in milliseconds per batch with a fixed set of observations and four query times. L counts all allocated history rows. Encode and decode are separate measurements of TRESS’s components.
Dataset
Group
Targets
Persist.
No context
TRESS
PhysioNet
Recent
24,434
0.3926
0.2799
0.2777
Older
2,289
0.3959
0.2908
0.2832
Unseen
285
0.9258
0.8936
0.8761
MIMIC
Recent
16,252
0.5278
0.3557
0.3511
Older
5,039
0.5018
0.3265
0.3063
Unseen
1,724
1.1699
1.0962
0.9502
Table 8. Clinical test MSE by the target variable’s observation history. The middle predictor is TRESS retrained without gated context. Within each row, all three predictors are scored on the same targets.
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
Irregular multivariate time series forecasting (IMTSF) requires capturing temporal dynamics and cross-variable dependencies from sparse, asynchronous observations. Existing methods often rely on learned temporal encoders, coupling historical representation construction with model optimisation and repeatedly processing the same observations. We propose CTLite, a lightweight forecasting framework that decouples historical summarisation from learned prediction. Specifically, a parameter-free continuous-time moment encoder summarises local levels and trends at multiple temporal scales while suppressing poorly supported slope estimates. A compact readout combines damped local extrapolation with a low-rank cross-variable correction that accounts for observation freshness, target availability, and forecast horizon. Since the historical summaries are independent of trainable parameters and forecast queries, they can be reused across optimisation steps and repeated predictions for an unchanged history without altering predictions or parameter gradients. Experiments on four real-world benchmarks show that CTLite achieves the best or second-best mean MSE and MAE among the evaluated methods with only 88-2,537 trainable parameters. Under matched GPU workloads at batch size 32, CTLite reduces inference latency by 18.6-87.3% compared with the latest models, including the cost of historical encoding. Reusing summaries further reduces total runtime by approximately 52% over 16 forecast calls on the same history compared with recomputing the encoder for every call, including the initial encoding cost. The code is at: https://github.com/SSMa913/NICLab-CTLite.
Shusen Ma, Xudong Zuo, Yangchun Xie +1
Institute of Advanced Technology, University of Science and Technology of China, Hefei, China · Department of Automation, University of Science and Technology of China, Hefei, China
Event-informed forecasting requires translating reports and historical responses into changes to a numerical forecast. We propose TIMBRE (Temporal Integration of Memory-Based Responses and Evidence), which combines source-aware representation, state-conditioned response transfer, and reliability-guided fusion before a frozen forecast head. A separate readout adjusts interval widths while preserving the median. In a single-seed, one-epoch development study of 13 tasks, TIMBRE improves MAE over ordinary fusion on eight tasks but over native Chronos-2 on only two. Disabling response transfer in the trained model reduces BTC and AULF MAE by 47.04% and 6.81%, respectively. These findings identify sensitivity to learned response transfer rather than a general forecasting advantage. Missing development-set scores and the absence of retrained ablations limit attribution to individual evidence mechanisms.
Xinyu Guan, Zhirong Zhang, Hongyuan Liu +4
Alibaba Group · Chang’an University · University of Electronic Science and Technology of China +4
Real-world time series are frequently driven by exogenous events and structural shifts, rendering conventional forecasting based solely on historical numerical observations insufficient. While language models can retrieve external news, standard retrieval-augmented approaches struggle with high noise, missing signals, and an inability to reason causally about event impacts. We propose SEER (Self-Evolving Event Reasoning and Retrieval), a closed-loop framework that dynamically optimizes event conditioning for time series forecasting. SEER translates prediction errors into two decoupled feedback mechanisms: (i) a reflective retrieval memory that refines subsequent search queries and filters spurious noise, and (ii) a persistent causal knowledge base that distills transferable domain dynamics. SEER enforces strict chronological boundaries across both event retrieval and reflection, preventing look-ahead bias and data leakage. Across six volatile time-series benchmarks, SEER consistently outperforms state-of-the-art time series foundation models and language model baselines.