Low-cost non-dispersive infrared (NDIR) CO2 sensors are widely used for continuous air quality monitoring, but their long-term reliability is hindered by instrumental drift. This paper presents an empirical and theoretical investigation into the drift behavior of such sensors, based on long-term measurement data. After isolating environmental effects, we observe that the residual instrumental drift exhibits meanreverting characteristics. We model this behavior using the Ornstein-Uhlenbeck process, a stochastic differential equation well-suited for capturing such dynamics. Maximum likelihood estimation is employed to fit the model to real sensor data, revealing sensor-specific long-term means and reversion rates. The results have implications for the design of more effective calibration strategies, particularly for one-point methods like automatic baseline correction. Limitations of the OU model are also discussed, motivating directions for future modeling improvements.