keepseatAI HANDBOOK
← All models
architecture

ModernTCN

advancedsupervised
parametricforecastingclassificationanomaly_detectionrepresentationtime_seriessequences

ModernTCN is a pure convolutional architecture that uses patches, large depthwise kernels, and separate temporal and cross-variable mixing for general time-series analysis.

Mechanisms

depthwise convolutionlarge kernelspatchingbackpropagation

Properties

nonlinearrepresentation learning

Constraints

requires large datarequires scalingsensitive to tuning

Practical profile

Explainability
low
Training cost
high
Inference cost
medium
Data appetite
high
Concepts to understand
Forecasting BaselinesRolling-Origin Backtesting