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

Aymane Harkati; Moncef Garouani; Olivier Teste; Julien Aligon; Mohamed Hamlich

cs.LGMar 25, 2026

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

Authors: Aymane Harkati, Moncef Garouani, Olivier Teste, Julien Aligon, Mohamed Hamlich

Organizations: 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

Abstract

Accurate forecasting of multivariate time series remains challenging due to the need to capture both short-term fluctuations and long-range temporal dependencies. Transformer-based models have emerged as a powerful approach, but their performance depends critically on the representation of temporal data. Traditional point-wise representations preserve individual time-step information, enabling fine-grained modeling, yet they tend to be computationally expensive and less effective at modeling broader contextual dependencies, limiting their scalability to long sequences. Patch-wise representations aggregate consecutive steps into compact tokens to improve efficiency and model local temporal dynamics, but they often discard fine-grained temporal details that are critical for accurate predictions in volatile or complex time series. We propose IPatch, a multi-resolution Transformer architecture that integrates both point-wise and patch-wise tokens, modeling temporal information at multiple resolutions. Experiments on 7 benchmark datasets demonstrate that IPatch consistently improves forecasting accuracy, robustness to noise, and generalization across various prediction horizons compared to single-representation baselines.

Figures & tables

Datasets	Horizon	PatchTST (with overlapping)		PatchTST (w/o overlapping)
Datasets	Horizon	MSE	MAE	MSE	MAE
ILI	24	1.963( $\blacktriangle+6.72\%$ )	0.878( $\blacktriangle+6.50\%$ )	2.095	0.939
	36	1.981( $\blacktriangledown-1.00\%$ )	0.898( $\blacktriangle+0.33\%$ )	1.961	0.901
	48	2.022( $\blacktriangle+0.19\%$ )	0.935( $\blacktriangle+1.54\%$ )	2.026	0.949
	60	1.991( $\blacktriangledown-3.61\%$ )	0.924( $\blacktriangledown-0.86\%$ )	1.919	0.916
Electricity	96	0.200( $\blacktriangle+2\%$ )	0.288( $\blacktriangle+1.73\%$ )	0.204	0.293

Table 1: Performance of PatchTST with/without overlapping patches across various datasets and horizons. Best scores are in bold blue, underlined values denote equal performance. Blue indicates gains, and red indicates losses relative to baseline.

Figure 1: (a) The proposed dual-stream architecture. The input sequence $X=(x_{i,j})_{1\leq i\leq L,\;1\leq j\leq M}$ is first divided into $N$ patches and combined with positional encoding. It is then passed into a Transformer Encoder block for patch attention and an Autocorrelation block for inner-patch autocorrelation. The outputs of both blocks are concatenated and flattened, then passed through a Linear layer to generate the final prediction $\hat{X}=(\hat{x}_{i,j})_{L+1\leq i\leq L+H,\;1\leq j\leq M}$ . (b) Detailed view of the Transformer Encoder attention and the Autocorrelation block.

Datasets	Weather	Electricity	ILI	ETTh1	ETTh2	ETTm1	ETTm2
Features (M)	21	321	7	7	7	7	7
Timesteps	52696	26304	966	17420	17420	69680	69680

Table 2: Description of the datasets used in the experiments.

Model		IPatch (Ours)		TimeMixer		PatchTST		TimesNet		Crossformer		MICN		FiLM		DLinear		FEDformer		Stationary		Autoformer		Informer
Metric		MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ILI	24	1.604( $\blacktriangle+9,17\%$ )	0.751( $\blacktriangle+12,38\%$ )	1.766	0.844	1.963	0.878	2.616	1.079	3.197	1.149	2.341	1.038	2.223	0.986	2.203	1.036	2.565	1.080	2.301	1.035	2.651	1.114	3.179	1.278
ILI	36	1.689( $\blacktriangle+5,50\%$ )	0.799( $\blacktriangle+9,38\%$ )	1.782	0.874	1.981	0.898	2.030	0.945	3.653	1.264	2.339	1.052	2.155	0.997	2.323	1.076	3.099	1.227	2.109	1.004	2.651	1.114	2.788	1.125
	48	1.562( $\blacktriangle+19,91\%$ )	0.783( $\blacktriangle+15,58\%$ )	1.873	0.905	2.022	0.935	2.032	0.917	3.903	1.317	2.437	1.068	2.144	1.011	2.436	1.146	3.183	1.239	2.477	1.094	2.846	1.146	2.942	1.160
	60	1.445( $\blacktriangle+24,84\%$ )	0.783( $\blacktriangle+12,26\%$ )	1.804	0.910	1.991	0.924	1.836	0.879	4.695	1.451	2.260	0.996	2.003	0.936	2.305	1.088	2.706	1.109	2.429	1.086	2.716	1.128	3.024	1.166
Electricity	96	0.153( $\blacktriangleright+0\%$ )	0.239( $\blacktriangle+3,34\%$ )	0.153	0.247	0.200	0.288	0.168	0.272	0.219	0.314	0.180	0.293	0.198	0.274	0.210	0.302	0.193	0.308	0.169	0.273	0.201	0.317	0.274	0.368

Table 3: Long-term Forcasting : Unified Look-back of 96, with horizon

\in\{96,192,336,720\}

for Electricity, Weather & ETT, ILI horizon is set

\in\{24,36,48,60\}

Bold results are the best outcomes, underlined are second best.

Figure 2: Visual comparison of 96-step forecasts on the Electricity dataset. IPatch (red) tracks the ground truth (black) with markedly higher fidelity than PatchTST, TimeMixer, Crossformer, and DLinear.

Figure 3: Predictive effectiveness vs efficiency on ILI dataset (forecast horizon 60). IPatch (red circle) offers the best performance-efficiency tradeoff among all evaluated architectures

Datasets	Horizon	IPatch (Ours)		Patch only		Autocorrelation only
Datasets	Horizon	MSE	MAE	MSE	MAE	MSE	MAE
ILI	24	1.604	0.751	1.977	0.873	1.790	0.805
ILI	36	1.689	0.799	1.886	0.882	1.875	0.848
	48	1.562	0.783	1.662	0.809	1.636	0.800
	60	1.445	0.789	1.839	0.952	1.711	0.870
Electricity	96	0.153	0.239	0.162	0.246	0.154	0.241

Table 4: Ablation study of patching and point-wise autocorrelation integration. Blue bold results are the best outcomes, underlined are second best.

Datasets	Horizon	IPatch (Fourrier)		IPatch (softmax)
Datasets	Horizon	MSE	MAE	MSE	MAE
ILI	24	1.604	0.751	1.739	0.806
ILI	36	1.689	0.799	1.644	0.798
	48	1.562	0.783	1.635	0.806
	60	1.445	0.789	1.569	0.788
Electricity	96	0.153	0.239	0.153	0.245

Table 5: Ablation study of the final activation function. Blue bold results are the best outcomes.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4: Qualitative comparison of forecasting models (IPatch, TimeMixer, Crossformer, DLinear, and PatchTST) on different datasets. IPatch shows consistently stronger temporal modeling across diverse time series characteristics.

Explore similar work

TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting
Mar 11, 2026Sravan Kumar Ankireddy, Nikita Seleznev, Nam H. Nguyen +4Time Series Forecasting Time Series Transformers
University of Texas at Austin, Austin, TX, USA · Capital One, USA · University of Maryland, College Park, MD, USA
One Step Closer to Ground Truth: A Multi-Scale Residual-Aware Representation Learning Pipeline for Predicting Time Series Data
Jun 9, 2026Amrijit Biswas, Mustafa Kamal, Robin Krambroeckers +4Multivariate Time Series Forecasting Residual Learning
Artificial Intelligence Department, RobotBulls Labs Geneva, Switzerland · Department of Electrical and Computer Engineering, North South University Dhaka, Bangladesh
Rethinking Patch Based Multivariate Time Series Forecasting with Semantic Structured Partitioning
Aug 20, 2026Zhiquan Huang, Jiazhe Wang, Linjing Xue +3Multivariate Time Series Forecasting Time Series Forecasting
School of Information Engineering, Henan University of Science and Technology, Luoyang, 471023, Henan, China · School of Software, Henan University of Science and Technology, Luoyang, 471000, Henan, China

Mar 11, 2026J/K move · Enter open · S save

TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting

Sravan Kumar Ankireddy, Nikita Seleznev, Nam H. Nguyen, Yulun Wu+3

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

Time Series Forecasting Time Series Transformers Time-Series Foundation Models

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting

One Step Closer to Ground Truth: A Multi-Scale Residual-Aware Representation Learning Pipeline for Predicting Time Series Data

Rethinking Patch Based Multivariate Time Series Forecasting with Semantic Structured Partitioning