Multidata Causal Discovery for Statistical Hurricane Intensity Forecasting

Abstract Improving statistical forecasts of tropical cyclone (TC) intensity is limited by complex nonlinear interactions and difficulty in identifying relevant predictors. Conventional methods prioritize correlation or fit, often overlooking confounding variables and limiting generalizability to unseen TCs. To address this, we leverage a multidata causal discovery framework with a replicated dataset based on the Statistical Hurricane Intensity Prediction Scheme (SHIPS) using the fifth generation ECMWF atmospheric reanalysis (ERA5) dataset. We conduct experiments to identify and select predictors causally linked to TC intensity changes. We then train multiple linear regression models to compare causal feature selection with correlation, random forest feature importance, and no feature selection, across five forecast lead times from 1 to 5 days (24–120 h). Causal feature selection consistently outperforms on unseen test cases, especially for lead times shorter than 3 days. Top causal features include vertical shear, midtropospheric potential vorticity, and surface moisture conditions, which are physically significant yet often underutilized in TC intensity predictions. We build an extended predictor set (SHIPS+) by adding selected features to the standard SHIPS predictors. SHIPS+ yields increased short-term predictive skill at lead times of 24, 48, and 72 h. Adding nonlinearity using a multilayer perceptron further extends skill to longer lead times, despite our framework being purely regional and not requiring global forecast data. Operational SHIPS tests confirm that three of the six added causally discovered predictors improve forecast skill, with the largest gains at longer lead times. Our results demonstrate that causal discovery improves TC intensity prediction and pave the way toward more empirical forecasts. Significance Statement As data-driven tropical cyclone (TC) forecasting tools become more sophisticated, it is critical to ensure that they make accurate forecasts for the right reasons, especially for TCs not included during training. Here, we use a technique called multidata causal discovery, which identifies a unique set of predictors driving TC intensity changes over the next 5 days by analyzing multiple realizations of the same underlying phenomenon: TCs in the North Atlantic. Adding these predictors to an operationally used statistical TC intensity prediction scheme improves accuracy of forecasts. This shows that selecting predictors based on cause and effect, not just correlation, can result in improved forecasts that are more generalizable as well as interpretable, since the causally discovered predictors are physically consistent.