The increasing volume of financial transactions and the dynamic nature of fraud patterns pose significant challenges for real-time fraud detection. Traditional methods often struggle to detect anomalies in streaming data due to their static nature and inability to adapt to evolving patterns. To address these challenges, we propose a novel hybrid model, ‘aSFD’ (Autoencoder-based Streaming Fraud Detection), which integrates an autoencoder for latent feature extraction with an ExtraTrees classifier, specifically tailored for streaming learning and real-time performance. The autoencoder is trained on non-fraudulent data to effectively identify outliers in real time, enhancing fraud detection capabilities. A sliding window mechanism is employed to enable continuous updates, ensuring adaptive decision-making and efficient management of concept drift, which distinguishes our model from static alternatives. Our main contribution is a hybrid model combining an autoencoder and ExtraTrees classifier, specifically adapted for streaming data environments with a sliding window mechanism to handle concept drift effectively. Experimental evaluations on the Kaggle and PaySim fraud datasets showed improved performance, with precision, recall, and F1-scores of 0.95, 0.88, 0.92 (Kaggle) and 0.93, 0.80, 0.86 (PaySim), surpassing state-of-the-art algorithms.