PRISM : A Hierarchical Multiscale Approach for Time Series Forecasting
Zihao Chen, Alexandre Andre, Wenrui Ma, Ian Knight, Sergey Shuvaev, Eva Dyer · 2025
Read PRISM : A Hierarchical Multiscale Approach for Time Series Forecasting by Zihao Chen, Alexandre Andre, Wenrui Ma, Ian Knight, Sergey Shuvaev, Eva Dyer (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-24898.txt.
Introduction
Zihao Chen, Alexandre Andre, and Wenrui Ma present PRISM, a forecasting method that splits a time series into a learnable tree of scales. Real series mix global trends, fine local structure, and patterns in between, which single-scale models miss. At the tree root, a global representation captures coarse trends, while recursive splits reveal finer detail level by level.
Iterative sequence modeling over that partitioned representation is the core of the method, whose full name spells out partitioned representation for iterative sequence modeling. Finance, biology, and healthcare forecasting are named targets. Readers get a multiscale design that lets a model attend to slow drift and quick wiggles together.
Catalog record
| CREATOR | Zihao Chen, Alexandre Andre, Wenrui Ma, Ian Knight, Sergey Shuvaev, Eva Dyer |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 70 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-24898
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
PRISM : A Hierarchical Multiscale Approach for Time Series Forecasting Zihao Chen, Alexandre Andre, Wenrui Ma, Ian Knight, Sergey Shuvaev, Eva Dyer Abstract Forecasting is critical in areas such as finance, biology, and healthcare. Despite the progress in the field, making accurate forecasts remains challenging because real-world time series contain both global trends, local fine-grained structure, and features on multiple scales in between. Here, we present a new forecasting method, PRISM (Partitioned Representation for Iterative Sequence Modeling), that addresses this challenge through a learnable tree-based partitioning of the signal. At the root of the tree, a global representation captures coarse trends in the signal, while recursive splits reveal increasingly localized views of the signal. At each level of the tree, data are projected onto a time-frequency basis (e.g., wavelets or exponential moving averages) to extract scale-specific features, which are then aggregated across the hierarchy. This design allows the model to jointly capture global structure and local dynamics of the signal, enabling accurate forecasting. Experiments across benchmark datasets show that our method outperforms state-of-the-art methods for forecasting. Overall, these results demonstrate that our hierarchical approach provides a lightweight and flexible framework for forecasting multivariate time series. The code is available at https://github.com/nerdslab/prism . 1 Introduction Time-series forecasting is crucial for decision-making across domains, enabling planning, adaptation, and timely intervention 1 ; 3 ; 33 . Accurate time-series forecasting is challenging and remains an area of active research. One of the factors that makes forecasting difficult is that signals within time series evolve across multiple scales 1 that need to be accounted for. Global trends interact with seasonal rhythms and local fluctuations, creating a hierarchical structure in which information at one scale sets context for the next. We argue that effective forecast models must capture this hierarchy explicitly, aligning long-term dynamics with fine-scale variability in a coherent representation. Recent work has moved toward this goal but current approaches fall short of learning unified hierarchical time-frequency representations. Some models incorporate frequency cues to capture periodicity and compact spectral structure 25 ; 32 , others emphasize temporal hierarchy through coarse-to-fine decompositions 2 ; 26 , and still others operate mainly in the frequency domain 27 ; 28 without modeling the temporal structure. While these directions highlight the value of hierarchical design, they construct it in only one domain at a time or flatten frequency into shallow features without linking it to temporal context. What remains missing is a coherent and reconstructible hierarchy that organizes both time and frequency and captures long-term structure, short-term variability, and the interactions between them in a unified representation. In this work, we introduce PRISM ( Partitioned Representations for Iterative Sequence Modeling ), a new model that fills this gap by jointly learning a hierarchical decomposition in both time and frequency. PRISM builds a binary tree over time, recursively partitioning the input into overlapping segments that preserve local context. At each node, a time-frequency decomposition (via Haar wavelets 15 or alternative transforms like exponential moving average filters) produces multiple bands, which are then adaptively weighted by a lightweight router that reweighs bands. The outputs of child nodes are stitched together through linear mixing, yielding multiscale representations that are both interpretable and effective for forecasting. Across benchmark datasets, PRISM achieves accuracy competitive with strong state-of-the-art baselines and surpasses them in many forecasting settings. The model also provides interpretability through its frequency-weighting mechanism, revealing which bands contribute most to predictions, and robustness through its reconstructible design, which stabilizes learning and scales naturally to longer contexts. Together, these results demonstrate that hierarchical multiscale representations, when organized into a unified time–frequency tree, provide a principled pathway toward more accurate and efficient forecasting. • We propose PRISM , a forecasting architecture that organizes time series data through a binary temporal hierarchy combined with multiresolution frequency encoding. This structure captures dependencies across temporal and spectral scales, linking long-range trends with short-term fluctuations. • PRISM learns importance scores over all time–frequency nodes, enabling the model to focus on the most predictive components, and promoting interpretable importance patterns. • Through extensive evaluation across diverse datasets and horizons, PRISM demonstrates strong generalization and efficiency. The model consistently outperforms competitive baselines while maintaining a compact design and transparent interpretability. 2 Related Work We group prior work by how they use time and frequency information. We first review models that process both domains at once. We then cover models that build a hierarchy only in time. Next we discuss methods that work mainly in the frequency domain. We conclude this section by positioning our approach among these prior works. Joint time and frequency modeling methods. Several methods combine information from the time and frequency domains at the same time. TimesNet 25 first discovers dominant periods with an FFT search. It reshapes a 1D series into 2D period–phase tensors and applies 2D CNN blocks to capture patterns inside each period and across periods. This gives a compact time stack guided by frequency peaks. FEDformer 32 adds a seasonal–trend decomposition to a Transformer. It replaces standard attention with frequency-enhanced blocks that operate on a sparse set of Fourier or wavelet modes. This reduces cost and focuses the model on informative bands. ETSformer 24 builds on exponential smoothing and adds frequency attention. It organizes the network into level, growth, and seasonality modules, so each module has a clear role. TFDNet 11 forms a time–frequency matrix with the short-time Fourier transform (STFT), then uses multi-scale encoders and separate time and frequency blocks for trend and seasonal parts. This design lets the model align slow and fast dynamics across domains. These models mix domains and gain efficiency from compact spectral views. However, they do not build a learnable hierarchy in both time and frequency, and they do not define a reconstructable pathway that stitches local pieces back together. Another noteworthy line of work is represented by CoST 23 , which uses time-domain and frequency-domain contrastive losses to learn disentangled trend and seasonal representations for downstream forecasting. Including time and frequency decomposition is complementary to our work. Time domain hierarchical decomposition methods. A second line of work builds a multiscale structure only along time. Early statistical models such as ARIMA 1 and SARIMA 1 decompose a series into trend, seasonality, and noise with fixed forms and linear dynamics. More recent deep learning methods keep the idea of decomposition but learn it from data. For instance: N-HiTS 2 uses hierarchical interpolation and multi-rate sampling. It assembles forecasts at several resolutions, so that coarse blocks capture slow movements and fine blocks refine details. DLinear 29 is another notable work which uses seasonal-trend decomposition integrated with linear layers, maintaining a simple structure with impressive performance. TimeMixer 21 is an MLP forecaster that performs decomposable multi-scale mixing on the past. It also ensembles several simple predictors for the future to improve stability. SCINet 8 applies a recursive split with downsampling, convolutions, and cross-branch interaction. The split-and-interact pattern captures multi-resolution patterns while keeping computation low. N-BEATS 14 is a basis-expansion stack with backcast and forecast heads. The interpretable variant uses polynomial and harmonic bases to separate trend and seasonality. Residual stacking gives a deep hierarchy over time. Autoformer 26 places a progressive trend–seasonal decomposition inside the network. An auto-correlation module models long-range