Old-School Convolutional Networks Strike Back in Long-Range Time Series Forecasting

For the better part of a decade, the story of long-term time series forecasting has been written by Transformers. These attention-driven architectures, borrowed from natural language processing, swept through the forecasting literature on the promise that their global receptive fields could capture long-range dependencies that older models supposedly could not. Convolutional neural networks, by contrast, were widely dismissed for long-horizon prediction because their filters only see a small window of data at a time. A new study challenges that assumption head-on. Researchers at Taiyuan University of Technology have revisited convolutional forecasting and built a hybrid architecture, called MRTFNet, that combines multi-resolution temporal modeling with frequency-domain analysis, achieving the best average performance on seven of nine real-world benchmark datasets. The work, published in Cluster Computing, argues that the perceived weakness of CNNs was never an inherent flaw but a design problem waiting to be solved.

The challenge the researchers set out to address is one that anyone who has tried to predict electricity demand, traffic flow, or weather knows intimately: real-world time series are messy, layered objects. They contain fine-grained short-term fluctuations riding on top of daily and weekly cycles, all entangled with cross-variable dependencies, such as how one sensor’s readings influence a neighboring sensor hours later. A model that only looks at one temporal scale tends to miss part of this structure. A model that only looks at local windows misses the global periodicity. Transformer-based models such as Informer, Autoformer, FEDformer, and Crossformer attacked these problems with increasingly sophisticated attention mechanisms, but they carry a well-known computational burden: attention scales quadratically with sequence length, making very long inputs expensive.

The Taiyuan team, led by Zefei Ning and Haodong Zhao with corresponding author Li Wang, took a different route. Their starting point was a temporal multi-resolution embedding module, which constructs representations of the input series at multiple temporal scales simultaneously. Instead of feeding the network a single fixed-resolution view of the past, the module produces a hierarchy of coarser and finer representations, allowing convolutional filters to capture both rapid, fine-grained dynamics and slow-moving long-term trends. The idea echoes techniques from computer vision, where multi-scale feature pyramids have long been standard, and from recent forecasting models such as Pyraformer and Pathformer, which introduced pyramidal and adaptive multi-scale pathways within Transformer frameworks. MRTFNet brings that same multi-resolution philosophy to a convolutional backbone.

But multi-resolution alone does not solve the fundamental complaint about CNNs: their limited global receptive field. A stack of convolutional layers, however deep, still integrates information only across a finite span unless the architecture grows impractically large. The researchers’ answer is a dual-branch time-frequency architecture. The first branch, a spatio-temporal multi-resolution representation module, operates in the time domain and captures temporal dependencies and cross-variable correlations at varying resolutions. The second branch, a frequency-aware global representation module, works in the frequency domain, where the Fourier transform converts the series into a sum of sinusoidal components. In that spectral representation, a global periodic pattern that spans thousands of time steps in the raw data becomes a single localized spike at the corresponding frequency, which the network can modulate directly.

This is the conceptual heart of the paper. Periodicity is one of the strongest and most exploitable structures in long-term forecasting, and the frequency domain exposes it in a way the time domain cannot. A weekly cycle in hourly data repeats every 168 steps; in the Fourier domain it appears as one coefficient. By learning to emphasize or suppress key spectral components, the frequency-aware module gives the convolutional network access to genuinely global structure without requiring attention over the full sequence. The approach aligns with a growing body of frequency-domain forecasting research, including FEDformer’s frequency-enhanced decomposition, FreDF’s frequency-domain learning, and work showing that frequency-domain multilayer perceptrons can be remarkably effective learners for time series.

The final piece of the architecture is a multi-head attention mechanism that fuses the outputs of the two branches. Rather than relying on attention to model long-range dependencies in the raw series, MRTFNet uses it at a higher level of abstraction: to integrate the complementary representations produced by the time-domain and frequency-domain modules before the final prediction. This is a much lighter use of attention, applied to compact feature representations rather than full-length sequences, which helps keep the overall computational profile manageable. The design reflects a broader trend in the field, visible in models such as TF P-Mixer and WPMixer, of treating time-domain and frequency-domain processing as complementary branches rather than rivals.

The empirical evaluation is where the argument becomes persuasive. The researchers tested MRTFNet on nine real-world benchmark datasets, a standard suite in the long-term forecasting literature that typically includes electricity consumption, traffic, weather, and other multivariate series with long horizons. Across these benchmarks, MRTFNet achieved the best average performance on seven of the nine datasets, outperforming strong Transformer-based baselines and recent multi-scale and frequency-aware competitors. The consistency of the result across diverse domains matters more than any single number, because it suggests the architectural principles, multi-resolution embedding plus time-frequency fusion, generalize rather than merely fitting the quirks of one dataset. The datasets used are publicly available through the widely used Time-Series-Library repository, which should make the results straightforward for other groups to verify and extend.

Why does this matter beyond the leaderboards? Long-term forecasting underpins decision-making in energy grid management, transportation planning, cloud resource allocation, and financial risk assessment, and the field has been locked in an expensive arms race of ever-larger attention mechanisms. If convolutional architectures, which are cheaper, faster, and better understood from an engineering standpoint, can match or beat Transformers when equipped with the right inductive biases, the practical implications are significant. Convolutional models deploy easily on edge devices and in streaming systems, and their computational cost scales linearly with input length. The Taiyuan study suggests that the local receptive field, long treated as a disqualifying limitation, can be compensated for by a modest frequency-domain branch that captures global periodicity at negligible cost.

There are also scientific lessons in the paper’s framing. The authors explicitly describe their work as a revisiting, a deliberate re-examination of assumptions that hardened into conventional wisdom. The history of machine learning is full of such reversals: techniques abandoned for the wrong reasons later return with better engineering, as happened with convolutional networks in vision and recurrent ideas in sequence modeling. The forecasting community has recently begun questioning other orthodoxies too, with work at major conferences asking whether global dependencies are even necessary and whether inverted Transformer variants might outperform the standard formulation. MRTFNet adds to this corrective momentum, showing that architectural progress often comes not from inventing new primitives but from combining old ones in ways that respect the structure of the data.

Caveats remain, as they always do. The study relies on benchmark datasets rather than live production deployments, and the paper’s abstract does not report inference-time comparisons that would quantify the efficiency advantage in practice. The frequency-domain branch adds design complexity, and the choice of resolutions and spectral filtering strategies introduces hyperparameters that practitioners will need to tune. Still, the core finding stands on its own: when convolutional networks are given multi-resolution views of the past and a window onto global periodic structure through the frequency domain, they are no longer the underdogs of long-term forecasting. For a field that has spent years assuming attention was the only road to long horizons, that is a result worth paying attention to, and it may well reshape how the next generation of forecasting systems is built.

Subject of Research: CNN-based multi-resolution and time-frequency fusion for long-term time series forecasting

Article Title: Revisiting CNN-based multi-resolution modeling for long-term time series forecasting

Article References: Revisiting CNN-based multi-resolution modeling for long-term time series forecasting. (n.d.). https://doi.org/10.1007/s10586-026-06542-7

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06542-7

Keywords: time series forecasting, convolutional neural networks, multi-resolution modeling, frequency domain, Transformers, deep learning, time-frequency analysis, MRTFNet, benchmarks, periodicity, machine learning, Cluster Computing