<p>Real-time anomaly detection in time series data is crucial for domains including system monitoring, health care, and manufacturing. Here, real-time capability encompasses both minimal detection delays and quick training, enabling timely intervention and issue mitigation. However, existing sliding window-based methods often suffer from delayed detection, and deep learning techniques impose training inefficiencies. To address these challenges, we propose the Mahalanobis Distance of Reservoir States (MD-RS) method. A reservoir refers to a fixed, randomly initialized recurrent neural network, which nonlinearly encodes input time series into high-dimensional reservoir states. The reservoir encoding suppresses detection delays as it processes inputs sequentially, with past inputs exponentially fading away. Furthermore, its fixed recurrent weights significantly reduce training costs. Unlike standard reservoir-based methods, which use fluctuating reconstruction errors as the anomaly score, MD-RS leverages Mahalanobis distance to measure deviations from the learned distribution of reservoir responses to normal data, enabling temporally stable anomaly scoring. Additionally, our method introduces a reservoir composed of mixed slow and fast neurons. Slow neurons excel at capturing long-term dependencies, while fast neurons rapidly return to normal states after anomalies. By leveraging both neuron types, the mixed reservoir enhances flexibility, facilitating effective real-time anomaly detection. Through comprehensive performance evaluation on univariate and multivariate datasets, we demonstrate that MD-RS outperforms state-of-the-art methods in terms of real-time capability. This makes MD-RS a promising new standard approach for real-time anomaly detection in time series data.</p>