Adaptive bias correction for improved subseasonal forecasting

Abstract

Subseasonal forecasting-predicting temperature and precipitation 2 to 6 weeks ahead-is critical for effective water allocation, wildfire management, and drought and flood mitigation. Recent international research efforts have advanced the subseasonal capabilities of operational dynamical models, yet temperature and precipitation prediction skills remain poor, partly due to stubborn errors in representing atmospheric dynamics and physics inside dynamical models. Here, to counter these errors, we introduce an adaptive bias correction (ABC) method that combines state-of-the-art dynamical forecasts with observations using machine learning. We show that, when applied to the leading subseasonal model from the European Centre for Medium-Range Weather Forecasts (ECMWF), ABC improves temperature forecasting skill by 60-90% (over baseline skills of 0.18-0.25) and precipitation forecasting skill by 40-69% (over baseline skills of 0.11-0.15) in the contiguous U.S. We couple these performance improvements with a practical workflow to explain ABC skill gains and identify higher-skill windows of opportunity based on specific climate conditions.

Fig. 1. Average forecast skill for dynamical models (red) and their adaptive bias correction (ABC) counterparts (blue).

Across the contiguous U.S. and the years 2018–2021, ABC provides a pronounced improvement in skill for each SubX or ECMWF dynamical model input and each forecasting task (a–d). The error bars display 95% bootstrap confidence intervals. Models without forecast data for weeks 5–6 are omitted from the bottom panels.

Fig. 2. Spatial skill distribution of dynamical models and their adaptive bias corrections.

Across the contiguous U.S. and the years 2018–2021, dynamical model skill drops precipitously at subseasonal timescales (weeks 3–4 and 5–6), but adaptive bias correction (ABC) attenuates the degradation, doubling or tripling the skill of CFSv2 (a, b) and boosting ECMWF skill 40–90% over baseline skills of 0.11–0.25 (c, d). Taking the same raw model forecasts as input, ABC also provides consistent improvements over operational debiasing protocols, tripling the precipitation skill of debiased CFSv2 and improving that of debiased ECMWF by 70% (over a baseline skill of 0.11). The average temporal skill over all forecast dates is displayed above each map.

Fig. 3. Fraction of contiguous U.S. with 2018–2021 spatial skill above a given threshold.

For each forecasting task and dynamical model input (a–d), adaptive bias correction (ABC) consistently expands the geographic range of higher skill over raw and operationally debiased dynamical models. The error bars display 95% bootstrap confidence intervals.

Fig. 4. Spatial distribution of model bias over the years 2018–2021.

Across the contiguous U.S., adaptive bias correction (ABC) reduces the systematic model bias of its dynamical model input for both precipitation (a, c) and temperature (b, d).

Fig. 5. Impact of the first 500 hPa geopotential heights principal component (hgt_500_pc1) on adaptive bias correction (ABC) skill improvement.

a To summarize the impact of hgt_500_pc1 on ABC-ECMWF skill improvement for precipitation weeks 3–4, we divide our forecasts into 10 bins, determined by the deciles of hgt_500_pc1, and display above each bin map the probability of positive impact in each bin along with a 95% bootstrap confidence interval. The highest probability of positive impact is shown in blue, and the lowest probability of positive impact is shown in red. We find that hgt_500_pc1 is most likely to have a positive impact on skill improvement in decile 1 which features a positive Arctic Oscillation (AO) pattern, and least likely in decile 9, which features AO in the opposite phase. b The forecast most impacted by hgt_500_pc1 in decile 1 is also preceded by a positive AO pattern and replaces the wet debiased ECMWF forecast with a more skillful dry pattern in the west.

Fig. 6. Impact of the Madden–Julian Oscillation phase (mjo_phase) on adaptive bias correction (ABC) skill improvement.

a To summarize the impact of mjo_phase on ABC-ECMWF skill improvement for precipitation weeks 3–4, we compute the probability of positive impact and an associated 95% bootstrap confidence interval in each lagged MJO phase bin and adopt the methodology of ref.  to create an MJO phase space diagram. The highest probabilities of positive impact (those falling within the confidence interval of the highest probability overall) are shown in blue and the lowest probability of positive impact is shown in red. We find that positive impact on skill improvement is most common in phases 2, 4, 5, and 8 and least common in phase 1. b The forecast most impacted by mjo_phase in phases 2, 4, 5, and 8 avoids the strongly negative skill of the debiased ECMWF baseline.

Fig. 7. Defining windows of opportunity for opportunistic adaptive bias correction (ABC) forecasting.

Here we focus on forecasting precipitation in weeks 3–4. a When more explanatory variables fall into high-impact deciles or bins (e.g., the blue bins of Figs. 5 and 6), the mean skill of ABC-ECMWF improves, but the percentage of forecasts using ABC declines. b The overall skill of opportunistic ABC is maximized when ABC-ECMWF is deployed for target dates with two or more high-impact variables and standard debiased ECMWF is deployed otherwise.

Fig. 8. Schematic of the opportunistic adaptive bias correction (ABC) workflow.

Opportunistic ABC uses historical ABC and baseline forecasts and a candidate set of explanatory variables to identify windows of opportunity for selective deployment of ABC in an operational setting.