The paper introduces MOIRAI, a Masked Encoder-based Universal Time Series Forecasting Transformer designed to address the challenges of universal forecasting in time series data. Traditional deep learning approaches for time series forecasting are limited to a one-model-per-dataset framework, which restricts their ability to leverage large pre-trained models. MOIRAI aims to overcome these limitations by proposing novel enhancements to the conventional time series Transformer architecture. These enhancements include multi-patch size input projection layers, Any-variate Attention, and a mixture of parametric distributions to handle varying frequencies, multivariate data, and flexible predictive distributions. To train MOIRAI, the authors introduce the Large-scale Open Time Series Archive (LOTSA), a collection of over 27 billion observations across nine domains. LOTSA provides a diverse and large-scale dataset for pre-training, addressing the need for a large-scale training dataset in universal forecasting. MOIRAI is trained in three sizes—Small, Base, and Large—and evaluated on both in-distribution and out-of-distribution settings, demonstrating competitive or superior performance compared to full-shot models. The paper also includes a detailed experimental evaluation, showing that MOIRAI achieves strong performance in probabilistic and long sequence forecasting. Ablation studies and further analyses highlight the importance of key components in MOIRAI's architecture and training methodology. The authors discuss limitations and future work, emphasizing the need for efficient tuning techniques, more flexible cross-frequency learning, and support for high-dimensional time series data.The paper introduces MOIRAI, a Masked Encoder-based Universal Time Series Forecasting Transformer designed to address the challenges of universal forecasting in time series data. Traditional deep learning approaches for time series forecasting are limited to a one-model-per-dataset framework, which restricts their ability to leverage large pre-trained models. MOIRAI aims to overcome these limitations by proposing novel enhancements to the conventional time series Transformer architecture. These enhancements include multi-patch size input projection layers, Any-variate Attention, and a mixture of parametric distributions to handle varying frequencies, multivariate data, and flexible predictive distributions. To train MOIRAI, the authors introduce the Large-scale Open Time Series Archive (LOTSA), a collection of over 27 billion observations across nine domains. LOTSA provides a diverse and large-scale dataset for pre-training, addressing the need for a large-scale training dataset in universal forecasting. MOIRAI is trained in three sizes—Small, Base, and Large—and evaluated on both in-distribution and out-of-distribution settings, demonstrating competitive or superior performance compared to full-shot models. The paper also includes a detailed experimental evaluation, showing that MOIRAI achieves strong performance in probabilistic and long sequence forecasting. Ablation studies and further analyses highlight the importance of key components in MOIRAI's architecture and training methodology. The authors discuss limitations and future work, emphasizing the need for efficient tuning techniques, more flexible cross-frequency learning, and support for high-dimensional time series data.