TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting

Sravan Kumar Ankireddy; Nikita Seleznev; Nam H. Nguyen; Yulun Wu; Senthil Kumar; Furong Huang; C. Bayan Bruss

cs.AIMar 11, 2026

TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting

Authors: Sravan Kumar Ankireddy, Nikita Seleznev, Nam H. Nguyen, Yulun Wu, Senthil Kumar, Furong Huang, C. Bayan Bruss

Organizations: University of Texas at Austin, Austin, TX, USA · Capital One, USA · University of Maryland, College Park, MD, USA

Abstract

Transformer-based time series foundation models face a fundamental trade-off in choice of tokenization: point-wise embeddings preserve temporal fidelity but scale poorly with sequence length, whereas fixed-length patching improves efficiency by imposing uniform boundaries that may disrupt natural transitions and blur informative local dynamics. In order to address these limitations, we introduce TimeSqueeze, a dynamic patching mechanism that adaptively selects patch boundaries within each sequence based on local signal complexity. TimeSqueeze first applies a lightweight state-space encoder to extract full-resolution point-wise features, then performs content-aware segmentation by allocating short patches to information-dense regions and long patches to smooth or redundant segments. This variable-resolution compression preserves critical temporal structure while substantially reducing the token sequence presented to the Transformer backbone. Specifically for large-scale pretraining, TimeSqueeze attains up to 20x faster convergence and 8x higher data efficiency compared to equivalent point-token baselines. Experiments across long-horizon forecasting benchmarks show that TimeSqueeze consistently outperforms comparable architectures that use either point-wise tokenization or fixed-size patching.

Figures & tables

Figure 1: Architectural overview of TimeSqueeze forecasting model. An SSM encoder first processes the raw series at full resolution to extract fine-grained features. Dynamic patching then adaptively compresses the sequence, selecting the salient subset of features. A Transformer backbone performs contextual modeling on the downsampled features, and an unpatching module upsamples the signal to the original resolution while preserving causality. Finally, an SSM decoder combines the compressed and fine-grained features, passing the hybrid features to multi-horizon heads, thereby improving efficiency without sacrificing temporal fidelity.

Models	Metrics	TimeSqueeze ${}_{\text{base}}$		TimeSqueeze ${}_{\text{large}}$		Time-MoE ${}_{\text{base}}$		Time-MoE ${}_{\text{large}}$		Moirai ${}_{\text{base}}$		TimesFM		Moment		Chronos ${}_{\text{large}}$
Models	Metrics	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTh1	96	$0.359$	$0.385$	$0.360$	0.379	0.357	$0.381$	0.350	$0.382$	$0.376$	$0.392$	$0.414$	$0.404$	$0.688$	$0.557$	$0.441$	$0.390$
	192	$0.400$	0.410	$0.402$	$0.407$	0.388	$0.412$	0.384	$0.404$	$0.417$	$0.413$	$0.465$	$0.434$	$0.688$	$0.560$	$0.502$	$0.424$
	336	0.420	$0.423$	$0.423$	0.412	0.411	$0.430$	0.411	$0.434$	$0.433$	0.428	$0.503$	$0.456$	$0.675$	$0.563$	$0.576$	$0.467$
	720	0.428	0.446	$0.441$	$0.448$	0.427	$0.455$	$0.449$	$0.477$	$0.447$	0.444	$0.511$	$0.481$	$0.683$	$0.585$	$0.835$	$0.583$
	Avg.	0.402	$0.416$	$0.407$	0.414	0.394	$0.419$	0.400	$0.420$	$0.417$	0.419	$0.473$	$0.443$	$0.683$	$0.566$	$0.588$	$0.466$

Table 1: Performance comparison of zero-shot forecasting. Best results are highlighted in red and second-best in blue .

Dataset	TimeSqueeze ${}_{\text{base}}$		Time-MoE ${}_{\text{base}}$		iTransformer		TimeMixer		TimesNet		PatchTST		DLinear		CycleNet ${}_{\text{MLP}}$		TQNet
Dataset	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTh1	0.398	0.419	0.379	0.406	$0.454$	$0.447$	$0.448$	$0.442$	$0.454$	$0.450$	$0.468$	$0.454$	$0.529$	$0.522$	$0.457$	$0.441$	$0.441$	$0.434$
ETTh2	0.350	0.393	0.346	0.386	$0.383$	$0.406$	$0.364$	$0.395$	$0.414$	$0.496$	$0.386$	$0.406$	$0.942$	$0.683$	$0.388$	$0.409$	$0.378$	$0.402$
ETTm1	$0.383$	0.386	0.345	0.381	$0.407$	$0.409$	$0.381$	$0.395$	$0.400$	$0.405$	$0.387$	$0.400$	$0.513$	$0.495$	$0.379$	$0.396$	0.377	$0.393$
ETTm2	0.259	0.321	$0.271$	$0.335$	$0.288$	$0.332$	$0.275$	$0.323$	$0.291$	$0.332$	$0.280$	$0.326$	$0.757$	$0.610$	0.266	0.314	$0.277$	$0.323$
Weather	$0.243$	$0.279$	0.236	$0.275$	$0.250$	$0.280$	0.240	0.271	$0.259$	$0.286$	$0.258$	$0.280$	$0.258$	$0.315$	$0.243$	0.271	$0.242$	0.269

Table 2: Performance comparison of full-shot forecasting. Best results are highlighted in red and second-best in blue .

Panel of Figure 2 — Figure 2: Computational efficiency comparison between TimeSqueeze ${}_{\text{base}}$ and Time-MoE: (a) Training memory and time requirements across different batch sizes and context lengths. TimeSqueeze achieves comparable performance while reducing memory usage by $3.4\times$ and training time by $\approx 20\times$ . (b) Inference throughput across prediction horizons. TimeSqueeze delivers up to $10.5\times$ higher throughput for longer prediction horizons.

Panel of Figure 3 — Figure 3: Ablation: (a) Average MSE across five benchmark datasets for prediction horizon 96 with different model components. (b) Pretraining Context Length vs Forecasting Performance: Longer pretraining context translates to improved performance, even when the inference context remains fixed at 512.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Models	ETTh1		ETTh2		ETTm1		ETTm2		Weather		Electricity
Models	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
Autoformer [2021]	0.496	0.487	0.450	0.459	0.588	0.517	0.327	0.371	0.338	0.382	0.227	0.338
FEDformer [2022]	0.498	0.484	0.437	0.449	0.448	0.452	0.305	0.349	0.309	0.360	0.214	0.327
DLinear [2023]	0.461	0.457	0.563	0.519	0.404	0.408	0.354	0.402	0.265	0.315	0.225	0.319
TimesNet [2023]	0.495	0.450	0.414	0.427	0.400	0.406	0.291	0.333	0.251	0.294	0.193	0.304
PatchTST [2023]	0.516	0.484	0.391	0.411	0.406	0.407	0.290	0.334	0.265	0.285	0.216	0.318

Appendix

Table 3: Multivariate time series forecasting results averaged across prediction horizons

T=\{96,192,336,720\}

with fixed input length

L=96

for all datasets. Best results are highlighted in red and second-best in blue .

Table 4: Model configurations.

	Energy	Finance	Healthcare	Nature	Sales	Synthetic	Transport	Web	Other	Total
# Seqs.	2,875,335	1,715	1,752	31,621,183	110,210	11,968,625	622,414	972,158	40,265	48,220,929
# Obs.	15.981 B	413.696 K	471.040 K	279.724 B	26.382 M	9.222 B	2.130 B	1.804 B	20.32 M	309.09 B
Percent %	5.17%	0.0001%	0.0001%	90.50%	0.008%	2.98%	0.69%	0.58%	0.006%	100%

Appendix

Table 5: Key statistics of the pre-training dataset Time-300B from various domains.

Dataset	Domain	Freq.	# Time Series	# Obs.	Source
Weatherbench (Hourly)	Nature	H	3,984,029	74,630,250,518	( Rasp et al., 2020 )
Weatherbench (Daily)	Nature	D	301,229	3,223,513,345	( Rasp et al., 2020 )
Weatherbench (Weekly)	Nature	W	226,533	462,956,049	( Rasp et al., 2020 )
Beijing Air Quality	Nature	H	4,262	2,932,657	( Chen, 2019 )
China Air Quality	Nature	H	17,686	4,217,605	( Zheng et al., 2015 )
CMIP6	Nature	6H	14,327,808	104,592,998,400	( Nguyen et al., 2023 )

Appendix

Table 6: Key properties of Nature dataset from Time-300B..

Figure 4: Performance scaling with training data size: Average MSE for 96-horizon forecasting across five benchmarks shows consistent improvement with increased training tokens.

Figure 5: Performance scaling with training data size: Average MSE for 96-horizon forecasting across five benchmarks shows consistent improvement with increased training tokens.

Figure 6: Performance comparison between TimeSqueeze and Time-MoE for a given context length for prediction horizon $336$ . TimeSqueeze noticeably outperforms the point-embedding baseline when the available context is limited.

Model / Variation	ETTh1		ETTh2		ETTm1		ETTm2		Weather		Average
Model / Variation	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
TimeSqueeze ${}_{\text{base}}$	$0.357$	$0.384$	$0.281$	$0.336$	$0.311$	$0.343$	$0.181$	$0.270$	$0.166$	$0.216$	0.259	0.310
Time-MoE ${}_{\text{base}}$	$0.357$	$0.381$	$0.305$	$0.359$	$0.338$	$0.368$	$0.201$	$0.291$	$0.160$	$0.214$	0.272 (+5.0%)	0.323 (+4.2%)
TimeSqueeze w/ fixed patching (size 4)	$0.373$	$0.396$	$0.455$	$0.448$	$0.359$	$0.382$	$0.335$	$0.380$	$0.178$	$0.232$	0.340 (+31.3%)	0.368 (+18.7%)
TimeSqueeze w/ fixed patching (size 2)	$0.376$	$0.402$	$0.475$	$0.462$	$0.361$	$0.384$	$0.357$	$0.396$	$0.179$	$0.233$	0.340 (+31.3%)	0.368 (+18.7%)
TimeSqueeze w/ linear patching (no SSM)	$0.379$	$0.401$	$0.481$	$0.463$	$0.370$	$0.375$	$0.375$	$0.402$	$0.158$	$0.174$	0.353 (+36.3%)	0.363 (+17.1%)

Appendix

Table 7: Ablation study on zero-shot forecasting performance for prediction horizon 96.

Figure 7: Forecasting performance improves with context length up to 512-800 tokens, then plateaus or slightly degrades.

Method	ETTh1 (MSE/MAE)	ETTh2 (MSE/MAE)	ETTm1 (MSE/MAE)	ETTm2 (MSE/MAE)	Weather (MSE/MAE)
Dynamic patching (avg patch size 4)	0.342 / 0.364	0.280 / 0.347	0.351 / 0.370	0.201 / 0.292	0.175 / 0.229
Fixed-size patching (size 4)	0.366 / 0.394	0.406 / 0.421	0.370 / 0.388	0.362 / 0.406	0.185 / 0.250

Appendix

Table 8: Evaluation of TimeSqueeze Dynamic patching with a generic transformer backbone.

Method	ETTh1		ETTh2		ETTm1		ETTm2		Weather		Avg.
Method	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
Dynamic patching	0.369	0.396	0.364	0.400	0.411	0.415	0.287	0.363	0.204	0.265	0.327	0.368
Fixed-size patching	$0.382$	$0.403$	$0.405$	$0.428$	$0.439$	$0.424$	$0.334$	$0.394$	$0.209$	$0.265$	$0.354$	$0.383$
Point embedding	$0.392$	$0.415$	$0.567$	$0.506$	$0.426$	$0.441$	$0.660$	$0.548$	$0.235$	$0.302$	$0.456$	$0.442$

Appendix

Table 9: Performance of TimeSqueeze-10M variant with pretraining of GiftEvalPretrain dataset, showing consistent gains of dynamic patching over fixed patching.

Models	Metrics	TimeSqueeze ${}_{\text{base}}$		Time-MoE ${}_{\text{base}}$		iTransformer		TimeMixer		TimesNet		PatchTST		DLinear		CycleNet ${}_{\text{MLP}}$		TQNet
Models	Metrics	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTh1	96	0.354	0.384	0.345	0.375	$0.386$	$0.405$	$0.375$	$0.400$	$0.384$	$0.402$	$0.414$	$0.419$	$0.423$	$0.448$	$0.375$	$0.395$	$0.371$	$0.393$
	192	0.397	0.412	0.372	0.396	$0.441$	$0.436$	$0.436$	$0.429$	$0.421$	$0.429$	$0.460$	$0.445$	$0.471$	$0.474$	$0.436$	$0.428$	$0.428$	$0.426$
	336	0.418	0.427	0.389	0.412	$0.487$	$0.458$	$0.484$	$0.458$	$0.491$	$0.469$	$0.501$	$0.466$	$0.570$	$0.546$	$0.496$	$0.455$	$0.476$	$0.446$
	720	0.423	0.454	0.410	0.443	$0.503$	$0.491$	$0.498$	$0.482$	$0.521$	$0.500$	$0.500$	$0.488$	$0.653$	$0.621$	$0.520$	$0.484$	$0.487$	$0.470$
	Avg.	0.398	0.419	0.379	0.406	$0.454$	$0.447$	$0.448$	$0.442$	$0.454$	$0.450$	$0.468$	$0.454$	$0.529$	$0.522$	$0.457$	$0.441$	$0.441$	$0.434$

Appendix

Table 10: Performance comparison of full-shot forecasting. Bold for best and underscore for 2nd best.

Panel of Figure 8 — Figure 8: Example dynamic patch boundaries for ETTh1 and ETTh2 datasets.

Panel of Figure 9 — Figure 9: Example dynamic patch boundaries for ETTm1 and ETTm2 datasets.

Panel of Figure 10 — Figure 10: Example dynamic patch boundaries for Weather dataset.

Figure 11: Distribution of patch sizes across eval datasets.

Panel of Figure 12 — Figure 12: Forecasting results on ETTh1 and ETTh2 datasets.

Panel of Figure 13 — Figure 13: Forecasting results on ETTm1 and ETTm2 datasets.

Panel of Figure 14 — Figure 14: Forecasting results on the Weather dataset.

Explore similar work

Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer
Mar 27, 2026Yulun Wu, Sravan Kumar Ankireddy, Samuel Sharpe +4Multiscale Time-Series Modeling Time Series Forecasting
Capital One · University of Texas, Austin
IPatch: A Multi-Resolution Transformer Architecture for Robust Time-Series Forecasting
Mar 25, 2026Aymane Harkati, Moncef Garouani, Olivier Teste +2Time Series Forecasting Multivariate Time Series Forecasting
CCPS Laboratory, ENSAM, Hassan II University, Casablanca, 20670, Morocco · IRIT, UMR 5505 CNRS, Université Toulouse Capitole, Toulouse, 31000, France · IRIT, UMR 5505 CNRS, Université Toulouse, UT2J, Toulouse, 31000, France
Scale-Aware Pretraining of Time Series Foundation Models via Multi-Patch Token Alignment and Hybrid Masking
Aug 20, 2026Taihua Chen, Xiang Ma, Yixin Zhang +3Time-Series Foundation Models Self-Supervised Pre-Training
School of Software, Shandong University, Jinan 250101, China · Joint SDU–NTU Centre for Artificial Intelligence Research (C-FAIR), Shandong University, Jinan 250101, China · Nanyang Technological University, Singapore

Mar 27, 2026J/K move · Enter open · S save

Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer

Yulun Wu, Sravan Kumar Ankireddy, Samuel Sharpe, Nikita Seleznev+3

Capital One · University of Texas, Austin

Multiscale Time-Series Modeling Time Series Forecasting

Open paper AskPDF arXiv

Efficiently aggregating spatial or temporal horizons to acquire compact representations has become a unifying principle in modern deep learning models, yet learning data-adaptive representations for long-horizon sequence data, especially continuous sequences like time series, remains an open challenge. While fixed-size patching has improved scalability and performance, discovering variable-sized, data-driven patches end-to-end often forces models to rely on soft discretization, specific backbones, or heuristic rules. In this work, we propose Reinforcement Patching (ReinPatch), the first framework to jointly optimize a sequence patching policy and its downstream sequence backbone model using reinforcement learning. By formulating patch boundary placement as a discrete decision process optimized via Group Relative Policy Gradient (GRPG), ReinPatch bypasses the need for continuous relaxations and performs dynamic patching policy optimization in a natural manner. Moreover, our method allows strict enforcement of a desired compression rate, freeing the downstream backbone to scale efficiently, and naturally supports multi-level hierarchical modeling. We evaluate ReinPatch on time-series forecasting datasets, where it demonstrates compelling performance compared to state-of-the-art data-driven patching strategies. Furthermore, our detached design allows the patching module to be extracted as a standalone foundation patcher, providing the community with visual and empirical insights into the segmentation behaviors preferred by a purely performance-driven neural patching strategy.

Abstract

Explore similar work

Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer

IPatch: A Multi-Resolution Transformer Architecture for Robust Time-Series Forecasting

Scale-Aware Pretraining of Time Series Foundation Models via Multi-Patch Token Alignment and Hybrid Masking