Novel Design of Compact Data Learning Frameworks for Time-Series Forecasting

This paper focuses on optimizing machine learning-based time-series forecasting models by constructing compact data. Compact Data Learning for Time Series (CDL-TS) is a novel framework aimed at minimizing forecasting model errors. By utilizing reduced sampling and robust comparison procedures, CDL-TS addresses the challenges of forecasting models on extensive real-time data systems. By strategically minimizing data size while maintaining accuracy, CDL-TS presents an innovative framework that facilitates robust predictions and enhances operational efficiency. The combined bivariate performance measure-based optimization effectively balances sampling frequency with the mean square error (MSE) to improve the trade operation performance in stock markets. Through a series of empirical applications, particularly involving the M/M/1 queuing system and various stock trade optimizations with global high-tech companies, the CDL-TS framework has proven its effectiveness by significantly minimizing both forecasting errors and operational costs. This accomplishment highlights the robust capabilities of CDL-TS in enhancing predictive accuracy while facilitating operation cost savings across different domains, including complex systems like queues and real-time financial markets.