One of the manifests of air-sea interactions is the change in sea level due to meteorological forcing through wind stress and atmospheric pressure. When meteorological conditions conducive to water level increase coincide with high tides during spring tides, the sea level may rise higher than expected and pose a flood risk to coastal low-lying areas. In Hong Kong, specifically when the northeast monsoon coincides with the higher spring tides in late autumn and winter, and sometimes even compounded by the storm surge brought by late-season tropical cyclones (TCs), the result may be coastal flooding or sea inundation. Aiming at forecasting such sea level anomalies on the scale of hours and days with local tide gauges using a flexible and computationally efficient method, this study adapts a data-driven method based on empirical orthogonal functions (EOF) regression of non-uniformly lagged regional wind field from ECMWF Reanalysis v5 (ERA5) to capture the effects from synoptic weather evolution patterns, excluding the effect of TCs. Local atmospheric pressure and winds are also included in the predictors of the regression model. Verification results show good performance in general. Hindcast using ECMWF forecasts as input reveals that the reduction of mean absolute error (MAE) by adding the anomaly forecast to the existing predicted astronomical tide was as high as 30% in February on average over the whole range of water levels, as well as that compared against the Delft3D forecast in a strong northeast monsoon case. The EOF method generally outperformed the persistence method in forecasting water level anomaly for a lead time of more than 6 h. The performance was even better particularly for high water levels, making it suitable to serve as a forecast reference tool for providing high water level alerts to relevant emergency response agencies to tackle the risk of coastal inundation in non-TC situations and an estimate of the anomaly contribution from the northeast monsoon under its combined effect with TC. The model is capable of improving water level forecasts up to a week ahead, despite the general decreasing model performance with increasing lead time due to less accurate input from model forecasts at a longer range. Some cases show that the model successfully predicted both positive and negative anomalies with a magnitude similar to observations up to 5 to 7 days in advance.