Yiyi Guo: Distributional Bias Correction for Madden-Julian Oscillation Forecast
Introduction: Subseasonal forecasts are critical for weather- and climate-sensitive decisions. The Madden-Julian Oscillation (MJO) is a key source of predictability at these timescales, but operational MJO forecasts remain limited by systematic model errors and imperfect uncertainty representation. Existing post-processing methods improve deterministic trajectories, but do not provide calibrated uncertainty estimates for risk-informed use.
Methodology: We develop a distributional bias correction framework for operational MJO forecasts. The method uses a Transformer encoder with simultaneous quantile regression to learn state-dependent predictive quantiles for the Real-time Multivariate MJO indices, RMM1 and RMM2, from dynamical forecast trajectories. By estimating multiple non-crossing quantiles in a single model, it produces both a bias-corrected median forecast and prediction intervals without assuming Gaussian forecast errors.
Results: Across BoM, JMA and CNRM reforecasts, the method reduces continuous ranked probability score by 18.3-24.7% relative to the raw ensemble, outperforms EMOS, Bayesian model averaging and Bayesian neural network baselines, and maintains near-nominal 90% interval coverage of 0.895-0.907. It also reduces RMSE and bivariate mean squared error, delays the loss of useful bivariate correlation by one to two days, and improves phase correction. These gains are consistent across lead times, Wheeler-Hendon phases and seasons.
Audience: This framework provides operational forecasters and climate-risk users with more accurate MJO trajectories and calibrated uncertainty estimates, supporting more reliable subseasonal decision-making.
Keywords: Madden-Julian Oscillation, Subseasonal Prediction, Distributional Bias Correction, Probabilistic Forecasting, Quantile Regression, Uncertainty Quantification.