The numerical deviations of the existing pavement performance prediction models for project-level forecasting make it not suitable for maintenance and repair (M&R) funding planning. Project-level forecasting lacks multi-feature parameters, and high-frequency maintenance makes it lack sufficient time series data. We have constructed a prediction framework to overcome the problem of insufficient data and large forecast bias. The performance of the time series status data-based prediction model is stable. Time series status data play important roles in improving the prediction accuracy, and the effect is greater than that of the dataset size. The difference in prediction accuracy between the machine learning algorithms is not significant. The variable selection conclusions obtained by the mean decrease impurity (MDI) sorting can effectively support random forest (RF), gradient boosting decrease tree (GBDT), and extreme gradient boosting (XGboost) prediction models. The dimensionality of feature data is greatly reduced. The 1–5 year prediction deviations of ride quality index (RQI) and rutting depth index (RDI) are basically within ± 2, which demonstrates the framework is an effective project-level forecasting method for pavement performance.