Learnable multi-scale decomposition with integrated autocorrelation and confidence-aware time series forecasting

In this work, we introduce a time series forecasting architecture that incorporates autocorrelation while leveraging dual encoders operating at multiple scales. Unlike traditional models that rely on predefined trend and seasonal components, our architecture applies learnable decomposition in the frequency domain and employs two separate encoders per scale: one focusing on low-pass filtering to capture trends and the other utilizing high-pass filtering to model seasonal variations. This learnable filtering allows the model to dynamically adapt and isolate trend and seasonal components directly in the frequency domain. A central and distinguishing innovation in our approach is the introduction of an integrated confidence scoring mechanism based on the reconstruction error of the same encoder backbone used for forecasting. This confidence score provides a direct, learned metric of prediction certainty, which we use to dynamically adjust the final forecasting output, effectively mitigating risks from uncertain predictions. We achieve the integration of autocorrelation by computing lagged differences in temporal processing, enabling the model to capture dependencies across time more effectively. In our architecture, each encoder processes the input through fully connected layers to handle temporal and channel interactions, and the same encoder backbone is used for both reconstruction and forecasting outputs. By combining frequency-domain filtering, autocorrelation-based temporal modeling, channel-wise transformations, and proposed confidence scoring mechanism, our model not only accurately captures long-term dependencies and fine-grained patterns, but also operates more efficiently compared to other state-of-the-art methods. Its lightweight design ensures faster processing while maintaining high precision in forecasting across diverse datasets and time horizons.