Time series analytics has been widely applied across various domains. While existing time series foundation models (TSFMs) have demonstrated promising results, they still face two major challenges: 1) Most TSFMs mainly rely on stacking Transformer layers to increase parameter size, lacking the ability to adaptively model complex time series patterns from multi-domain data. 2) Current TSFMs are primarily pre-trained for specific tasks, e.g., forecasting or anomaly detection, and a few models support multiple tasks, but they fail to fully leverage task specificity and commonality, resulting in suboptimal performance. To address these challenges, we propose TimeSeek, a novel generalized time series analytics model that adaptively models time series patterns from multi-domain data and unifies multiple tasks in pre-training. Conceptually, TimeSeek is guided by selective information sharing under domain and task heterogeneity. From the perspective of multi-domain data, we propose a Multi-Domain Feature Extractor that adaptively captures temporal dependencies and channel correlations based on domain-specific characteristics, including progressive multi-scale modeling and adaptive slot channel modeling. From the perspective of the diversity of tasks, we propose multi-task collaborative training, which leverages task specificity and task commonality through task customization and cross-task training, effectively unifying multiple tasks in pre-training and better adapting to diverse downstream scenarios. After pre-training on large-scale time series data, TimeSeek achieves state-of-the-art performance in forecasting, anomaly detection, and classification tasks across both few-shot learning and zero-shot situations, together with favorable inference efficiency.
Figures & tables
Fig. 1: TimeSeek leverages multi-domain data and unifies multiple tasks in pre-training.
Models
Multi-Task Pre-training
Multi-Scale Modeling
Channel Modeling
Zero-Shot
Timer [ 10 ]
✗
✗
✗
✓
MOIRAI [ 14 ]
✗
✗
✓
✓
TimesFM [ 13 ]
✗
✗
✗
✓
Chronos [ 16 ]
✗
✗
✗
✓
Moment [ 11 ]
✗
✗
✗
✗
UniTS [ 15 ]
✓
✗
✓
✓
TABLE I: Compare TimeSeek with existing TSFMs.
Fig. 2: The framework of TimeSeek.
Fig. 3: The framework of the Multi-Domain Feature Extractor (MFE).
Fig. 4: Multi-Tasks pre-training framework of TimeSeek, where the Fine-Scale Extractor is put on top to ease description.
Method
TimeSeek
Timer
TimesFM
Chronos
Time-MoE
UniTS
MOIRAI
TEMPO
Metric
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
ETTm(Avg)
0.311
0.350
0.636
0.515
0.391
0.389
0.422
0.392
0.322
0.368
0.548
0.457
0.333
0.354
0.391
0.393
ETTh(Avg)
0.390
0.406
0.466
0.441
0.439
0.422
0.462
0.424
0.391
0.413
0.394
0.416
0.432
0.423
0.394
0.413
Weather
0.217
0.250
0.292
0.313
–
–
0.288
0.309
0.256
0.289
0.275
0.298
0.260
0.275
0.282
0.316
Traffic
0.555
0.367
0.613
0.407
–
–
0.615
0.421
–
–
–
–
–
–
0.503
0.358
Solar
0.192
0.257
0.771
0.604
0.500
0.397
0.393
0.319
0.411
0.428
0.845
0.669
0.714
0.704
0.416
0.374
TABLE II: The zero-shot forecasting results of time series foundation models. The average results of all predicted lengths {96,192,336,720} are listed here. A dash (’-’) denotes datasets included in the model’s pretraining and therefore excluded from testing. ETTm represents the average results across ETTm1 and ETTm2, while ETTh represents the average across ETTh1 and ETTh2. Bold: the best, and underline: the second best.
Method
TimeSeek (10%)
iTransformer
FITS
TimeMixer
PatchTST
Crossformer
DLinear
Metric
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
ETTm(avg)
0.285
0.337
0.303
0.348
0.306
0.345
0.307
0.349
0.303
0.348
0.481
0.479
0.307
0.350
ETTh(avg)
0.365
0.394
0.400
0.421
0.372
0.407
0.387
0.418
0.385
0.416
0.666
0.57
0.456
0.455
Weather
0.210
0.246
0.232
0.270
0.244
0.281
0.225
0.263
0.224
0.261
0.233
0.292
0.239
0.290
Traffic
0.413
0.273
0.397
0.281
0.429
0.302
0.410
0.279
0.397
0.275
0.521
0.282
0.434
0.295
Solar
0.188
0.253
0.202
0.260
0.232
0.268
0.203
0.261
0.200
0.284
0.204
0.232
0.247
0.309
TABLE III: The forecasting results of TimeSeek and time series-specific models are presented, with average performance reported across all prediction lengths {96,192,336,720} . TimeSeek utilizes only 10% of the training data for few-shot learning, whereas time series-specific models are trained using the full training data. Bold: the best and underline: the second best. Full results are in the Supplemental Material.
Type
Statistical Methods
Task-Specific Models (Superwised)
Time Series Foundation Models (Zero-Shot)
Model
Naïve
Auto
Auto
Auto
DeepAR
TiDE
N-BEATS
PTST.
DLinear
TimesFM
TabPFN
Chronos
MOIRAI
TimeSeek
ETS
ARIMA
Theta
(2020)
(2023)
(2019)
(2022)
(2023)
(2023)
(2025)
(2024)
(2024)
(Ours)
MASE
1.270
1.212
1.074
1.090
1.343
1.091
0.938
0.849
1.061
1.077
0.771
0.822
0.875
0.732
CRPS
1.591
7.489
0.912
1.244
0.853
0.772
0.816
0.587
0.846
0.680
0.544
0.577
0.599
0.528
TABLE IV: Performance of TimeSeek on the GIFT-Eval leaderboard. A lower MASE or CRPS indicates a better performance. Baseline results are from the GIFT-Eval leaderboard.
Dataset
SMD
MSL
SMAP
SWaT
PSM
Metric
P
R
F1
P
R
F1
P
R
F1
P
R
F1
P
R
F1
OCSVM
66.98
82.03
73.75
50.26
99.86
66.87
41.05
69.37
51.58
56.80
98.72
72.11
57.51
58.11
57.81
IForest
71.94
94.27
81.61
53.87
94.58
68.65
41.12
68.91
51.51
53.03
99.95
69.30
69.66
88.79
78.07
AE
69.22
98.48
81.30
55.75
96.66
70.72
39.42
70.31
50.52
54.92
98.20
70.45
60.67
98.24
75.01
DAGMM
63.57
70.83
67.00
54.07
92.11
68.14
50.75
96.38
66.49
59.42
92.36
72.32
68.22
70.50
69.34
LSTM
60.12
84.77
70.35
58.82
14.68
23.49
55.25
27.70
36.90
49.99
82.11
62.15
57.06
95.92
71.55
TABLE V: The anomaly detection results are presented. TimeSeek performs zero-shot inference without utilizing any training data, whereas baselines are trained on the full training data. Bold: the best and underline: the second best. More metric results are in the Supplemental Material.
Dataset
TimeSeek (10%)
Moment
TimesNet
PatchTST
DLinear
Crossformer
OS-CNN
TapNet
Rocket
XGBoost
DTW
EthanolConcentration
0.718
0.300
0.357
0.328
0.362
0.380
0.240
0.323
0.447
0.437
0.323
FaceDetection
0.651
0.689
0.686
0.683
0.680
0.687
0.575
0.556
0.694
0.633
0.529
Handwriting
0.361
0.351
0.321
0.296
0.270
0.288
0.668
0.357
0.567
0.158
0.286
Heartbeat
0.814
0.737
0.780
0.749
0.751
0.776
0.489
0.751
0.718
0.732
0.717
JapaneseVowels
0.986
0.957
0.984
0.975
0.962
0.991
0.991
0.965
0.965
0.865
0.949
PEMS-SF
1.000
0.855
0.896
0.893
0.751
0.859
0.760
0.751
0.856
0.983
0.711
TABLE VI: The classification results are presented. TimeSeek utilizes only 10% of the training data for few-shot learning, whereas baselines are trained using the full training data. Bold: the best and underline: the second best.
Fig. 5: Ablation about multi-scale modeling on Weather and ETTm1 datasets. Sum and Concat represent scale feature summation and scale feature concatenation, respectively. Single Scale denotes modeling with only the fine scale. The results are averaged across four prediction lengths.
Fig. 6: Ablation about channel modeling for the Weather and ETTm1 datasets. W/O Router refers to removing the channel router. Attention represents using standard attention for channel modeling. Channel-independent denotes ignoring channel modeling entirely. The results are averaged across four prediction lengths.
Fig. 7: Ablation study on the classification module across five datasets. Vanilla class token denotes directly initializing a trainable class token for classification.
Fig. 8: Results of fine-tuning TimeSeek under different data percentages with prediction length 96.
Fig. 9: The effect of the number of slots K .
Fig. 10: A comparison of inference time and the parameter counts of time series foundation models, along with their zero-shot performance on the Weather dataset.
Electrical and Computer Engineering, Queen’s University · Ingenuity Labs Research Institute, Queen’s University · Mechanical and Materials Engineering, Queen’s University
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Xiamen Institute of Data Intelligence