This paper introduces a novel generative pretrained hierarchical transformer (GPHT) for time series forecasting. The proposed model addresses two critical issues in existing methods: limited dataset scale and the one-step generation schema. GPHT constructs a mixed dataset under the channel-independent assumption, enabling the model to learn commonalities across diverse time series data. It also employs an auto-regressive forecasting approach, allowing for flexible forecasting at arbitrary horizons without a customized forecasting head. The model is evaluated on eight benchmark datasets, demonstrating superior performance compared to both self-supervised and supervised methods across various fine-tuning and zero/few-shot learning settings. The results show that GPHT achieves significant improvements in forecasting accuracy, particularly in long-term forecasting tasks. Additionally, the model exhibits strong generalization capabilities, even when applied to unseen datasets. The paper also presents ablation studies and zero-shot evaluations, further validating the effectiveness of GPHT. The proposed method is a generative self-supervised pretraining approach that effectively addresses the challenges of dataset scale and forecasting schema, offering a promising solution for time series forecasting.This paper introduces a novel generative pretrained hierarchical transformer (GPHT) for time series forecasting. The proposed model addresses two critical issues in existing methods: limited dataset scale and the one-step generation schema. GPHT constructs a mixed dataset under the channel-independent assumption, enabling the model to learn commonalities across diverse time series data. It also employs an auto-regressive forecasting approach, allowing for flexible forecasting at arbitrary horizons without a customized forecasting head. The model is evaluated on eight benchmark datasets, demonstrating superior performance compared to both self-supervised and supervised methods across various fine-tuning and zero/few-shot learning settings. The results show that GPHT achieves significant improvements in forecasting accuracy, particularly in long-term forecasting tasks. Additionally, the model exhibits strong generalization capabilities, even when applied to unseen datasets. The paper also presents ablation studies and zero-shot evaluations, further validating the effectiveness of GPHT. The proposed method is a generative self-supervised pretraining approach that effectively addresses the challenges of dataset scale and forecasting schema, offering a promising solution for time series forecasting.