Monthly Sea Level Forecast

Monthly observations are first converted to anomalies by removing a fixed 1993-2016 monthly climatology. A linear trend is then fitted to those monthly anomalies over 1993-2025. The fitted slope and intercept are evaluated over every displayed month, including months after 2025, and subtracted from the anomaly field. Data after 2025 do not change the fitted trend. A minimum finite-sample requirement is applied before a trend is retained.

The skill system uses IB-inclusive SSHa. Tide-gauge observations contain the inverse-barometer response naturally, while the gridded altimetry product includes its atmospheric-pressure-related sea-level component. Tide-gauge observations use the same anomaly and detrending convention. Stations are mapped to the 1-degree observation grid for gridded display and verification, and station-indexed companion products are used where available. Remaining tide-gauge gaps may be filled only after conversion to detrended anomaly space, using a nearest-altimetry proxy.

Dynamical model forecasts are bias-corrected using model-specific annual-cycle and linear-trend correction components derived from each model's retained-member ensemble mean, and their SLP-derived inverse-barometer component is included before skill is evaluated. The trend fit period is 1993-2025, and the corrected outputs are saved as linearly detrended, IB-inclusive SSHa. For damped persistence, independently trained no-IB and IB-component hindcasts are recombined. ST-LIM and CS-LIM use their no-IB hindcasts plus the damped-persistence IB-component hindcast. XRO uses the ORAS5 SST/D20 gridded no-IB hindcast plus the same damped-persistence IB-component hindcast. The displayed leads are months 1-6.

The site includes dynamical models, statistical models, and multi-model ensemble products when their monthly files are available for the forecast cycle.

Equal-weight ensemble means are computed across finite available model forecast means. The ensemble spread is the standard deviation of an equal-weight Gaussian mixture using each model's SSHa mean and standard deviation; if only one model contributes, the spread is that model's standard deviation. Displayed 68% and 95% ensemble intervals are exact Gaussian-mixture CDF quantiles rather than mean ± mixture standard deviation.

The Top-5 MME product selects candidate models independently at each location and displayed lead using retrospective anomaly correlation coefficient (ACC). Damped persistence is included as a candidate for these top-model products. The selected finite forecasts are then averaged with equal weights.

Retrospective skill is evaluated by aligning each forecast valid month with the matching observed linearly detrended, IB-inclusive SSHa over 1993-2025. The website displays ACC, root-mean-square error (RMSE), RMSE skill score (RMSESS), and probabilistic CRPSS. RMSESS is calculated as 1 − RMSE/RMSE_ref, where RMSE_ref is the error of a zero-SLA forecast. ACC uses a lag-1 adjustment for effective sample size. RMSE is in meters. Individual-model CRPSS uses the Gaussian forecast distribution; multi-model products use the exact equal-weight Gaussian-mixture distribution. Both are referenced to the fixed full-period empirical distribution of IB-inclusive observations for the same target calendar month.

Daily water-level outlooks are not detrended anomalies. For each tide-gauge station, the workflow uses the observed station trend components fitted from monthly station levels. The trend is evaluated at the midpoint of each forecast month and compared with the midpoint of the 2002-2020 tidal datum epoch. That trend change since the datum-epoch midpoint is added to the monthly detrended forecast anomaly before creating daily water-level values. If station trend components are missing, the add-back is set to zero and the product records that status.

Daily maximum water level is then computed as:

Total forecast = predicted astronomical tide daily maximum relative to 2002-2020 MHHW + observed station trend change since the 2002-2020 datum-epoch midpoint + detrended forecast monthly anomaly.

For observed daily maxima, the same display separates the long-term trend change, the observed monthly anomaly, and a daily residual term, so storm-surge and other non-tidal daily departures are not folded into the trend component.

The daily tide is estimated from a UTide harmonic reconstruction fit to hourly station data over 2002-2020 when enough data are available; otherwise a daily-maximum tide climatology fallback is used. The displayed reference is a station MHHW datum defined as the mean daily maximum water level over 2002-2020. For the Local calendar day option, UTC hourly timestamps are converted with the station's UHSLC timezone_name before daily aggregation, including daylight-saving transitions where applicable.

The default threshold is an empirical station threshold: the 98th percentile of 2002-2020 observed daily maximum water levels, expressed relative to the 2002-2020 MHHW datum. Where a NOAA minor-flood threshold is available, that threshold is converted to the same 2002-2020 MHHW reference and can be selected in the high-tide-flooding panel.

The high-frequency uncertainty option follows the structure of the seasonal high-tide-flooding approach of Dusek et al., 2022: hourly non-tidal residuals are estimated as observed water level minus astronomical tide, monthly means are removed, and residual distributions are conditioned by calendar month and tide height decile. Daily exceedance probability is computed from hourly exceedance probabilities with an autocorrelation adjustment. The alternative monthly uncertainty option uses Gaussian mean ± 1 and ± 1.96 standard deviations for an individual model, and exact equal-weight Gaussian-mixture quantiles for an MME, around the same tide-plus-monthly-anomaly daily maximum centerline. The combined option assumes independence between the monthly model uncertainty and the zero-centered high-frequency residual. For an individual model their variances are added; for an MME the high-frequency variance is added to every Gaussian component before exact mixture quantiles are calculated. Hourly bounds are converted to daily maxima, while the centerline remains unchanged.

The Forecast Discussion is prepared with assistance from IDEA, the Intelligent Data Exploring Assistant. IDEA consults the Monthly Sea Level Forecast Discussion Skill, reads the latest dashboard data and recent context, drafts a concise discussion, and revises the text in response to feedback from the human forecaster in the loop. The final discussion is reviewed by the forecaster before publication and is not treated as an autonomous forecast.

This workflow follows the general IDEA framework described by Widlansky and Komar, 2025, in which a tool-using AI assistant supports transparent, reproducible geoscience data exploration while retaining human scientific judgment. Published discussions may include an IDEA conversation link when available so users can inspect the preparation context.