The sliding window is commonly used to detect the changing causal relationships between time series, but its performance is sensitive to window sizes. However, little research has been conducted for appropriate window sizes. We propose Detection Windows based on Hidden Markov Model (HMMDW) for time-varying causal discovery of time series. Firstly, a sliding window moves along the time series, an autoregressive model with an external variable and time lags (called full model) are established in each window. A series of adjusted R-squares are calculated by the full model fitted to the time series within each window. Subsequently, these R-squares are as the observations of the hidden Markov model with two states (e.g., 1 & 0), which imply good or poor full models. After the hidden Markov model is solved, the scope of consecutive 1 or 0 corresponds to a detection window size, so a series of detection windows are obtained. Finally, the Granger causality test is used to discover the causal relationships between time series within every detection window. The experimental results show that our method is effective, and its comprehensive performance outperforms the comparative methods in terms of accuracy. Code is available: https://github.com/wkjwang/hmm_window