DriftMind: A Self-Adaptive, Cold-Start Framework for Time Series Forecasting and Anomaly Detection in Fast Data Streams

The proliferation of streaming data from IoT devices, financial markets, and industrial sensors has made real-time forecasting a critical capability. While deep learning models have emerged as state-of-the-art for this challenge, their heavy computational requirements, dependence on GPU hardware, and need for initial “warm-up” periods limit their applicability in resource-constrained or true cold-start environments, where no prior training data is available. This paper introduces DriftMind, a novel, lightweight forecasting engine specifically designed for such scenarios. DriftMind operates on the principle of online pattern clustering, dynamically identifying, learning, and adapting to evolving data patterns without any model retraining. We benchmark DriftMind against OneNet, a leading deep learning architecture, using the publicly available ETTh2 and ETTm1 time series datasets. The results show that DriftMind achieves comparable predictive performance (in terms of MAE and MSE) while being up to 140 times faster. On the ETTh2 dataset, DriftMind delivers forecasts in seconds on a standard CPU architecture, whereas OneNet requires nearly an hour of computation on a GPU. We conclude that DriftMind presents a practical, scalable, and highly efficient alternative for real-time forecasting applications where speed, adaptability, and cold-start capability are essential.