Shortage of maintenance costs is the common concern of current highway maintenance management, thus requiring the rational allocation of government investment. However, the traditional annual maintenance cost prediction model was highly influenced by environmental and macroeconomic factors, leading to inferior performance in controlling the construction cost. Therefore, this study aims to propose a high-precision model for predicting the annual maintenance cost using time-series data. To this end, one database containing 7-year records for 15 highways in South China was constructed and preprocessed utilizing the mathematical statistics method. Specifically, input time-series data include maintenance costs and monetary inflation rates for previous years. Afterward, a moving forest model was first introduced by combining the techniques of a random forest regression and a sliding time window. Finally, taking the routine maintenance costs as an example, the impact significance of all input variables related to market prices and the environment were quantified and compared. As a result, the routine maintenance cost prediction model with a time window length of 3 has the best performance. At the same time, routine maintenance costs for previous years played the most important role. Overall, the findings give decision-makers insight into the required cost information, ultimately informing maintenance management strategies development.