TSI-Bench is a comprehensive benchmark for time series imputation using deep learning techniques. It provides a standardized framework for evaluating imputation algorithms across diverse domains and missingness scenarios. The benchmark includes 28 algorithms, 8 datasets (air quality, traffic, electricity, healthcare), and covers various missingness patterns (point, subsequence, block) and rates (10%, 50%, 90%). TSI-Bench enables researchers to compare imputation methods fairly and identify the impact of missingness on model performance. It also offers a systematic approach to adapt forecasting models for imputation tasks. The benchmark includes extensive experiments across 34,804 trials, revealing that no single algorithm performs best across all settings. Different missing patterns significantly affect imputation performance, with block and subsequence missingness leading to higher errors. Forecasting models can be adapted for imputation, showing improved performance. Traditional methods like LOCF and Linear interpolation also perform well in some cases. TSI-Bench provides insights into the importance of model selection based on missingness rates and patterns, and highlights the need for efficient and robust imputation methods. The benchmark supports downstream tasks like classification, regression, and forecasting, demonstrating the significance of imputation in improving these tasks. TSI-Bench aims to advance time series imputation research by providing a standardized, open-source platform for evaluation and experimentation.TSI-Bench is a comprehensive benchmark for time series imputation using deep learning techniques. It provides a standardized framework for evaluating imputation algorithms across diverse domains and missingness scenarios. The benchmark includes 28 algorithms, 8 datasets (air quality, traffic, electricity, healthcare), and covers various missingness patterns (point, subsequence, block) and rates (10%, 50%, 90%). TSI-Bench enables researchers to compare imputation methods fairly and identify the impact of missingness on model performance. It also offers a systematic approach to adapt forecasting models for imputation tasks. The benchmark includes extensive experiments across 34,804 trials, revealing that no single algorithm performs best across all settings. Different missing patterns significantly affect imputation performance, with block and subsequence missingness leading to higher errors. Forecasting models can be adapted for imputation, showing improved performance. Traditional methods like LOCF and Linear interpolation also perform well in some cases. TSI-Bench provides insights into the importance of model selection based on missingness rates and patterns, and highlights the need for efficient and robust imputation methods. The benchmark supports downstream tasks like classification, regression, and forecasting, demonstrating the significance of imputation in improving these tasks. TSI-Bench aims to advance time series imputation research by providing a standardized, open-source platform for evaluation and experimentation.