ConvTimeNet is a deep hierarchical fully convolutional network designed for multivariate time series analysis. It addresses limitations of traditional convolutional networks by introducing adaptive segmentation of time series into sub-series level patches, treating these as fundamental modeling units. This approach avoids the sparsity semantics of raw point-level time steps. Additionally, a fully convolutional block is designed by integrating deepwise and pointwise convolution operations, inspired by Transformer encoders. This backbone network effectively captures global sequence and cross-variable dependencies, leveraging both Transformer architecture and convolutional properties. Multi-scale representations are learned by controlling kernel sizes. Extensive experiments on forecasting and classification tasks show that ConvTimeNet outperforms strong baselines in most cases. The code is publicly available. ConvTimeNet introduces a deformable patch embedding layer to transform raw time series into patch embeddings, enabling adaptive selection of time points based on input features. This module allows for adaptive adjustment of patch positions and scales, mitigating semantic loss from fixed splitting. A fully convolutional block is designed using deepwise and pointwise convolutions, enabling efficient modeling of temporal and cross-variable dependencies. The network's hierarchical architecture and large kernels allow for global receptive fields and multi-scale representation learning. ConvTimeNet outperforms traditional convolutional and Transformer methods in time series forecasting and classification tasks, demonstrating superior performance and efficiency. However, it requires careful tuning of hierarchical hyperparameters and may benefit from self-supervised pre-training for further improvement. The network's effectiveness is validated through experiments on various time series datasets, showing its potential as a versatile model for time series analysis.ConvTimeNet is a deep hierarchical fully convolutional network designed for multivariate time series analysis. It addresses limitations of traditional convolutional networks by introducing adaptive segmentation of time series into sub-series level patches, treating these as fundamental modeling units. This approach avoids the sparsity semantics of raw point-level time steps. Additionally, a fully convolutional block is designed by integrating deepwise and pointwise convolution operations, inspired by Transformer encoders. This backbone network effectively captures global sequence and cross-variable dependencies, leveraging both Transformer architecture and convolutional properties. Multi-scale representations are learned by controlling kernel sizes. Extensive experiments on forecasting and classification tasks show that ConvTimeNet outperforms strong baselines in most cases. The code is publicly available. ConvTimeNet introduces a deformable patch embedding layer to transform raw time series into patch embeddings, enabling adaptive selection of time points based on input features. This module allows for adaptive adjustment of patch positions and scales, mitigating semantic loss from fixed splitting. A fully convolutional block is designed using deepwise and pointwise convolutions, enabling efficient modeling of temporal and cross-variable dependencies. The network's hierarchical architecture and large kernels allow for global receptive fields and multi-scale representation learning. ConvTimeNet outperforms traditional convolutional and Transformer methods in time series forecasting and classification tasks, demonstrating superior performance and efficiency. However, it requires careful tuning of hierarchical hyperparameters and may benefit from self-supervised pre-training for further improvement. The network's effectiveness is validated through experiments on various time series datasets, showing its potential as a versatile model for time series analysis.