Data aggregation has been developed to take benefits from either spatial or temporal correlation of collected data. In wireless sensor networks, data aggregation is a key mechanism to reduce energy consumption and to increase capacity. Most current aggregation functions are designed for specific network topology or data pattern, with prior assumptions. Concretely, these functions are developed with fixed models or parameters, so they fail to adapt against non-anticipated data variations or dynamic scenarios. In this work, we propose a dynamic forecasting function: Agnostic Aggregation (A²), which can predict values with self-tuned model based on temporal correlation. With A², sensor nodes are able to adjust model to fit the latest time series better. Compared with previous works, the proposed A² can deliver packets to the sink with higher accuracy, and is more robust to independent and consecutive errors. We also investigate the influence of system parameters on the performance of A², and highlight the optimal choices. Comparison results verify that A² improves the recovery fidelity and reduces model updates than other aggregation functions.