Real-Time Forecasting Using Mixed Frequency Time-Series Data

This paper addresses the challenge of forecasting time-series data, including mixed-frequency data. The proposed method integrates high-frequency features with low-frequency outputs by capturing essential dynamics and enhancing real-time decision-making. A novel objective is introduced to learn and ensemble a set of models with identical structures and parameters. By utilizing this base model within a larger framework, the method processes lagged inputs from various frequencies to generate accurate predictions. The model's performance was evaluated using three datasets: the Electricity Transformer Dataset, Individual Household Electric Power Consumption, and the Max Planck Weather Dataset. Results demonstrate that the proposed method outperforms state-of-the-art models in both high-frequency and low-frequency forecasting scenarios. Bayesian Hyperparameter Optimization was employed to fine-tune the model, further improving its performance with reduced computational effort. This approach opens new avenues for handling mixed-frequency data in various forecasting applications.