ModernTCN: Pure Convolutional Forecaster

• ModernTCN is a CNN-based architecture that decouples temporal, channel-wise, and variable-wise dependencies for improved forecasting.
• It utilizes large-kernel convolutions to capture extensive temporal receptive fields, enabling efficient direct multi-step predictions.
• Applied across trajectory prediction, RUL estimation, and financial forecasting, it serves as a robust and adaptable baseline.

Searching arXiv for the original ModernTCN paper and related references. ModernTCN denotes Luo and Wang’s ICLR 2024 pure-convolution architecture for general time-series analysis, and in subsequent literature it is treated as a representative “modern” CNN-based forecaster rather than as a recurrent or attention-driven model. Across the cited works, it is characterized by explicit separation of dependency types, large temporal receptive fields obtained through convolutional design, and deployment as a strong baseline or backbone in forecasting pipelines spanning user trajectory prediction, remaining useful life estimation, and multi-horizon financial forecasting (Hao et al., 4 Aug 2025, Fan et al., 27 Oct 2025, Saidd, 27 Feb 2026).

1. Definition and scope

In the trajectory-prediction literature, ModernTCN is introduced as an advanced CNN-based time-series forecasting model alongside TCN and SCINet. Its defining description is that it “decouples and separately models temporal, channel-wise, and variable-wise relationships” and “uses large convolutional kernels to capture a broad receptive field,” making it a pure convolutional architecture designed to capture long-range temporal dependencies without recurrent or attention components (Hao et al., 4 Aug 2025).

This positioning distinguishes ModernTCN from classical TCN formulations based primarily on stacks of dilated residual convolutions. The financial benchmark further describes it as using large-kernel depthwise $1$D convolutions, pointwise mixing, multi-stage downsampling or hierarchy, structural reparameterisation, and RevIN-based normalization, again emphasizing a modernized convolutional inductive bias rather than sequence recurrence or full self-attention (Saidd, 27 Feb 2026).

The nomenclature is not entirely uniform across later applications. The SARNet study refers to the same backbone as both “ModernTCN” and “Modern Temporal Convolutional Network” or “MTCN,” but preserves the central idea of a pure convolutional forecaster used as the only deep component in a larger prognostics pipeline (Fan et al., 27 Oct 2025). A plausible implication is that ModernTCN is best understood as a specific architecture with a recognizable design philosophy, rather than as a single frozen implementation.

2. Architectural characteristics

The recurring architectural traits attributed to ModernTCN are consistent across the provided sources. At the broadest level, it is a pure convolutional model for time-series analysis, with no recurrent state update and no mandatory attention mechanism. Its design separates temporal, channel-wise, and variable-wise interactions, and it relies on large kernels to enlarge the temporal receptive field (Hao et al., 4 Aug 2025). In the controlled financial benchmark, these same ideas are instantiated as large-kernel depthwise temporal convolutions combined with pointwise convolutions, temporal patch embedding, hierarchical downsampling, structural reparameterisation, and a direct forecasting head $\hat{y} \in \mathbb{R}^h$ for multi-step output (Saidd, 27 Feb 2026).

The benchmarked forecasting formulation is explicitly direct rather than recursive: $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$ with $X_t$ an OHLCV window and $h \in \{4,24\}$ in the financial experiments (Saidd, 27 Feb 2026). This direct multi-step formulation also aligns with the trajectory study’s observation that ModernTCN predicts all future steps in one forward pass, so inference time is almost constant as the prediction horizon varies (Hao et al., 4 Aug 2025).

Task-specific instantiations differ substantially, which indicates that published uses of ModernTCN often adapt the original design to local constraints rather than reproducing a single canonical hyperparameter set.

Study	Reported ModernTCN instantiation	Reported role
(Hao et al., 4 Aug 2025)	Original-paper settings reused; internal hyperparameters not listed	Strongest CNN-type baseline for trajectory prediction
(Fan et al., 27 Oct 2025)	3 residual blocks, channels $[32,32,16]$, dilations $2^i$, $SEQ\_LEN=20$, $PRED\_STEP=5$	Backbone forecaster for a degradation-sensitive indicator
(Saidd, 27 Feb 2026)	Patch size $8$ or $\hat{y} \in \mathbb{R}^h$0, channel dims $\hat{y} \in \mathbb{R}^h$1, RevIN, dropout and head dropout $\hat{y} \in \mathbb{R}^h$2–$\hat{y} \in \mathbb{R}^h$3, parameter count $\hat{y} \in \mathbb{R}^h$4	Best-performing architecture in a controlled multi-horizon benchmark

This heterogeneity is consequential. In SARNet, the paper states that normalization, activations, and any gating or depthwise structure are inherited from the original ModernTCN design and are not redefined, whereas in the financial study the implementation details are surfaced explicitly and become part of the hyperparameter search space (Fan et al., 27 Oct 2025, Saidd, 27 Feb 2026).

3. Optimization, objectives, and computational profile

ModernTCN is used primarily as a supervised forecaster trained with pointwise regression losses. In the financial benchmark, all models, including ModernTCN, minimize mean squared error on future price vectors,

$\hat{y} \in \mathbb{R}^h$5

with model selection and early stopping based on validation MSE (Saidd, 27 Feb 2026). In SARNet, the ModernTCN backbone is likewise optimized with MSE using Adam at learning rate $\hat{y} \in \mathbb{R}^h$6, batch size $\hat{y} \in \mathbb{R}^h$7, and $\hat{y} \in \mathbb{R}^h$8 epochs, but its target is not RUL itself; rather, it forecasts the univariate degradation-sensitive feature FFT_bin_2_H (Fan et al., 27 Oct 2025).

The trajectory study adds a systems-level comparison. There, ModernTCN is assigned asymptotic time complexity $\hat{y} \in \mathbb{R}^h$9, matching MMCTP, TCN, TSMixer, LSTM, and Seq2Seq LSTM, while Informer is listed as $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$0 and PatchTST as $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$1. Empirically, ModernTCN’s per-sample inference time is reported as similar to MMCTP’s and positioned between Informer and TSMixer; when prediction length varies at fixed input length, its inference time remains almost constant because all steps are produced in parallel (Hao et al., 4 Aug 2025).

The financial benchmark reinforces the computational interpretation. Pure convolutions make ModernTCN parallelisable and memory-efficient on GPUs, while its parameter count remains moderate relative to attention-heavy alternatives. In that study, its selected configurations have approximately $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$2 parameters, smaller than Autoformer and iTransformer but larger than DLinear (Saidd, 27 Feb 2026). This suggests that ModernTCN’s computational identity is not merely “CNN instead of Transformer,” but a specific trade-off: moderate capacity, wide temporal filtering, and direct multi-step decoding.

4. Use in applied research

ModernTCN has been used in the provided literature in three notably different ways: as a baseline forecaster in trajectory prediction, as the sole deep backbone in a hybrid prognostics system, and as the strongest standalone architecture in a large-scale financial benchmark.

In user trajectory prediction on GeoLife, ModernTCN serves as the most competitive baseline model. It is evaluated under the same train/validation/test split and the same MSE, MAE, and inference-time criteria as the proposed MMCTP model. For the $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$3 s sampling interval dataset with input length $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$4 and prediction length $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$5, ModernTCN attains $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$6 and $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$7, while MMCTP attains $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$8 and $$f_{\theta}: \mathbb{R}^{w \times 5} \rightarrow \mathbb{R}^{h}, \qquad \hat{y}_{t+1:t+h} = f_\theta(X_t),$$9, corresponding to the headline $X_t$0 MSE reduction and $X_t$1 MAE reduction relative to ModernTCN at similar inference time. Across all reported input lengths and horizons, MMCTP consistently outperforms ModernTCN, but ModernTCN remains the principal comparison point and the strongest baseline in that study (Hao et al., 4 Aug 2025).

In SARNet, ModernTCN is not used as an end-to-end RUL regressor. It first forecasts the scalar degradation-sensitive indicator FFT_bin_2_H; an adaptive consecutive-spike rule then validates sustained onset, after which a stacked RF–LGBM regressor produces final RUL estimates. The full SARNet system, which includes the ModernTCN backbone, reports $X_t$2, $X_t$3, $X_t$4, and Coverage $X_t$5. By contrast, non-spike variants built on the same ModernTCN backbone, such as MTCN + LGBM, report $X_t$6, $X_t$7, $X_t$8, and Coverage $X_t$9. The ablation therefore shows that ModernTCN is a competent indicator forecaster whose outputs become markedly more useful once embedded in spike-aware validation and nonlinear post-onset regression (Fan et al., 27 Oct 2025).

In multi-horizon financial forecasting, ModernTCN is the top-ranked architecture in a controlled comparison of nine models over cryptocurrency, forex, and equity index markets. Its mean RMSE rank is $h \in \{4,24\}$0, with a $h \in \{4,24\}$1 first-place rate over $h \in \{4,24\}$2 asset-horizon evaluation slots. Category-level RMSE aggregates are $h \in \{4,24\}$3 for crypto, $h \in \{4,24\}$4 for forex, and $h \in \{4,24\}$5 for indices, each the best among the tested architectures. At the harder $h \in \{4,24\}$6 horizon, it is best on all $h \in \{4,24\}$7 forex assets and all $h \in \{4,24\}$8 indices. This is the strongest positive empirical evidence for ModernTCN in the provided corpus (Saidd, 27 Feb 2026).

5. Comparative position, strengths, and limitations

ModernTCN is consistently presented as stronger than older recurrent baselines in settings where large receptive fields and parallel temporal modeling are advantageous, but the evidence in the provided sources does not support the claim that it is universally dominant.

Its principal strengths are architectural. The trajectory study emphasizes its ability to model temporal, channel-wise, and variable-wise relationships separately while using large kernels for broad receptive fields (Hao et al., 4 Aug 2025). The financial benchmark adds that local, hierarchical convolutions and RevIN-based normalization appear well matched to noisy, non-stationary OHLCV sequences, where ModernTCN outperforms Transformer, MLP, and RNN families on RMSE and MAE (Saidd, 27 Feb 2026). This suggests that its core inductive bias is especially effective when multi-scale local structure matters more than explicit token-to-token attention.

At the same time, the supplied evidence identifies several limitations. In trajectory prediction, a model that explicitly separates global and local temporal information and fuses them with cross-attention surpasses ModernTCN across all reported settings, with the gap widening at coarser sampling intervals and longer contexts (Hao et al., 4 Aug 2025). In RUL estimation, ModernTCN alone is insufficiently robust around degradation onset; SARNet improves performance by wrapping it in adaptive thresholding, consecutive validation, targeted feature engineering, and an RF–LGBM regressor (Fan et al., 27 Oct 2025). In finance, although ModernTCN is the best model under MSE-trained point forecasting, directional accuracy remains near $h \in \{4,24\}$9 across all categories and horizons, specifically $[32,32,16]$0 and $[32,32,16]$1 for crypto at $[32,32,16]$2 and $[32,32,16]$3, $[32,32,16]$4 and $[32,32,16]$5 for forex, and $[32,32,16]$6 and $[32,32,16]$7 for indices. The benchmark explicitly interprets this as evidence that MSE-trained models, including ModernTCN, lack directional skill at hourly resolution (Saidd, 27 Feb 2026).

A common misconception is therefore that ModernTCN’s superior point-forecasting accuracy automatically transfers to all downstream objectives. The provided studies contradict that view. In some settings it is the strongest standalone forecaster; in others it is best used as a backbone inside a larger domain-specific system; and in at least one application it is outperformed by an architecture that adds explicit global-local decomposition and cross-attention (Hao et al., 4 Aug 2025, Fan et al., 27 Oct 2025, Saidd, 27 Feb 2026).

6. Relation to the broader evolution of temporal convolutional models

ModernTCN belongs to a broader shift from earlier TCN designs toward architectures that preserve temporal resolution, enlarge receptive fields efficiently, and specialize temporal processing. A closely related but distinct line is exemplified by MS-TCN++, which is not the same model as Luo and Wang’s ModernTCN but is described as embodying “modern” TCN principles for action segmentation: full-resolution temporal convolutions, large receptive fields via dilation, multi-stage refinement, explicit temporal consistency losses, and architectural specialization of stages (Li et al., 2020).

Other domain-specific “modern TCN” systems show parallel developments. WiFi-TCN employs a WaveNet-style causal and dilated TCN with large kernels, residual-style blocks, and a causal-masked attention module for WiFi human interaction recognition, reaching $[32,32,16]$8 accuracy on the HHI dataset while remaining computationally efficient relative to a BiLSTM baseline (Lin et al., 2023). DualTCN, in marine CSEM inversion, uses a dual-branch TCN encoder with physics-constrained outputs, achieving a $[32,32,16]$9 loss reduction over baseline models and $2^i$0 ms inference per sample on an A100 GPU (Ahmed et al., 6 May 2026). These systems are not ModernTCN itself, but they indicate the surrounding design trajectory in which contemporary temporal convolutional models increasingly combine large-kernel convolutions, multi-scale structure, task-aware heads, and domain constraints.

Within that landscape, ModernTCN occupies a specific position: a general-purpose, pure-convolution backbone whose appeal lies in large receptive fields, dependency decoupling, and efficient direct forecasting. The available evidence shows both why it became a strong modern baseline and why later systems often preserve it as a backbone while adding application-specific mechanisms around it (Hao et al., 4 Aug 2025, Fan et al., 27 Oct 2025, Saidd, 27 Feb 2026).