Windows of opportunity for predicting seasonal climate extremes highlighted by the Pakistan floods of 2022

The study addresses how to provide confident and actionable seasonal predictions of rare climate extremes, using the summer 2022 Pakistan floods as a case study. Seasonal outlooks typically communicate tercile probabilities, which can mask extreme risk. The authors posit that strong, predictable climate drivers (notably La Niña) create windows of opportunity where predictive skill for extremes is elevated beyond average hindcast-based estimates. The purpose is to demonstrate that examining forecast distributions beyond terciles, assessing physical drivers, and employing targeted experiments can increase confidence in forecasts of extremes and improve early warning value.

Previous work has linked Pakistan’s extreme monsoon rainfall, notably in 2010, to strong La Niña conditions and a westward-shifted West Pacific Subtropical High (WNPSH), alongside teleconnections from ENSO that affect South Asian monsoon variability. Extratropical influences via upper-tropospheric circulation anomalies have also been implicated, including links to the 2010 Russian heatwave. Seasonal forecast systems have shown skill in representing ENSO teleconnections to South Asia and in predicting related circulation features such as the Somali jet and WNPSH. Prior case studies used perturbation experiments to identify influences of specific SST patterns (e.g., North Atlantic tripole, Indian Ocean Dipole) on regional extremes, suggesting that targeted experiments can establish causality and bolster forecast confidence.

- Prediction systems and datasets: Primary analyses use the Met Office DePreSys3 (DP3) near-term prediction system initialized 1 May with a 40-member ensemble and a 43-year hindcast (1979–2021). Observations include GPCP and GPCC precipitation and ERA5 reanalysis winds/MSLP. Eight additional operational seasonal systems from the Copernicus Climate Change Service (C3S) archive (ECMWF SEAS5, Météo-France System 8, CMCC SPS3.5, DWD GCFS2.1, NCEP CFSv2, JMA CPS, and UKMO GloSea) are used to compare quantile probabilities for 2022. - Beyond tercile probabilities: The 1 May forecasts for JJA 2022 Pakistan rainfall are assessed across quantiles (upper tercile, quintile, decile) relative to hindcast climatologies. Time series of forecasted probabilities are compared across years to contextualize 2022. - Skill assessment and teleconnections: Average hindcast skill is quantified via correlation between DP3 ensemble mean Pakistan rainfall and observations. Circulation diagnostics compare forecast and observed 850 hPa and 250 hPa wind anomalies, and historical correlations between Pakistan rainfall and circulation (ERA5 and DP3) to assess consistency with known teleconnections. Gridpoint skill maps are computed for zonal wind fields. - Perturbation experiments (attribution): To isolate La Niña’s influence, tropical Pacific ocean temperature and salinity are nudged to a near-neutral ENSO state (2021) to create 1 May 2022 initial conditions without La Niña; a 40-member ensemble is run in parallel to the original. A reciprocal experiment nudges 2022 tropical Pacific initial conditions into the 1 May 2021 state. La Niña impacts are diagnosed as differences between original and perturbed ensembles, scaled to account for partial SST signal realization due to weak ocean relaxation (approximate factor-of-two inflation based on Niño3.4 amplitude ratios). - Empirical modeling: Two large-scale indices are constructed: WNPSH strength (MSLP over 115–150E, 15–25N) and a subtropical jet meridional shift (STJshift) index (difference in U250 averaged over boxes in Fig. 3). Multiple linear regression (MLR) using WNPSH and STJshift predicts Pakistan rainfall anomalies; correlations are computed with and without 2022. - Extremes identification tool: Global land is divided into 237 equal-area regions (Stone), and standardized ensemble mean rainfall anomalies are mapped to flag extreme predicted signals (>±2σ using semi-standard deviations for wet/dry tails). An evaluation over 1979–2022 assesses how often flagged extremes correspond to observed outer quantile outcomes (quintile, decile, 5th percentile) using GPCC, applying filters by minimum hindcast correlation (e.g., r>0.25).

- Observed extremity: Pakistan JJA 2022 rainfall totaled 415 mm, 260% of climatology and 4.9σ above the mean, exceeding the 2010 event (2.4σ). - Forecast extremity: DP3 predicted a Pakistan area-mean ensemble-mean anomaly of +3.5σ for JJA 2022, unprecedented in the 44-year record. The upper tercile probability was 67.5% (highest in 44 years); upper quintile probability 52.5% (2.6× climatology); upper decile probability 35% (3.5× climatology). Several C3S systems also showed unprecedented high upper-decile probabilities; in three systems 50–80% of members were in the upper decile. - Skill context: Average hindcast correlation for Pakistan rainfall in DP3 is r=0.50 (p<0.001); excluding 2022 gives r=0.27 (p=0.08). Across eight C3S systems, hindcast correlations ranged from r≈−0.02 to 0.38. ENSO (Niño3.4) forecast skill is high (r=0.86, p<0.001). - Physical consistency: Forecast 850 hPa and 250 hPa wind anomalies in 2022 closely match ERA5 and historical teleconnections, including anomalous easterlies extending from the Bay of Bengal to Pakistan, a strengthened/northwest track of monsoon depressions, intensified Somali jet moisture transport, a westward-intensified WNPSH, and a poleward-shifted subtropical Asian jet with anomalous easterlies over the southern Tibetan Plateau. - Attribution via perturbations: Nudging experiments confirm that the predicted strong summer La Niña substantially contributed to the Pakistan rainfall forecast signals and associated circulation anomalies in 2022. - Empirical model: WNPSH is highly predictable (r=0.72, p<0.001); STJshift predictability is significant (r=0.48, p=0.001). MLR using WNPSH and STJshift predicts Pakistan rainfall with r=0.67 (p<0.001), and r=0.53 (p<0.001) excluding 2022; both drivers were in phase in 2022, yielding a >2σ predicted anomaly. - Extremes tool evaluation: For 77 regions with hindcast skill r>0.25, 134 predicted extreme events (>±2σ) during 1979–2022 correspond to observed outcomes far exceeding climatological outer quantile frequencies: over half were outer quintile, more than one-third outer decile, and over one-fifth outer 5% extremes. For summer 2022, among 16 skilled regions with predicted >±2σ anomalies, 12 (75%) realized correct-sign outer quintile outcomes, 8 (50%) were outer decile, and 4 (25%) were outer 5% (including Pakistan).

The findings demonstrate that 2022 presented a window of opportunity for confident seasonal prediction of an extreme event due to a strong, predictable La Niña that organized large-scale circulation features conducive to Pakistan rainfall extremes. Standard tercile-focused outlooks obscured the severity of risk, whereas examining higher quantiles revealed unprecedented probabilities across multiple systems. Physical consistency between forecast and historical teleconnections, along with targeted perturbation experiments and an empirical MLR using key circulation indices, provided convergent evidence to increase confidence in extreme forecasts. The proposed extremes identification tool further shows potential to systematically flag and assess extreme signals and their drivers, thereby improving early warnings. These results underscore the importance of state-dependent predictability: when major drivers such as ENSO or IOD are active, actionable and more confident extreme-event forecasts are feasible. Effective communication and consideration of false alarms remain essential for uptake.

Strong, physically grounded signals for the extreme Pakistan rainfall in summer 2022 were present in modern seasonal prediction systems but were not fully conveyed by standard tercile products. By probing the full forecast distribution, leveraging physical teleconnections, conducting perturbation experiments to attribute drivers (notably La Niña), and applying simple empirical models, forecast confidence can be enhanced in windows of opportunity. The authors propose routine monitoring tools to highlight extreme signals, provide mechanistic context, and, where appropriate, deploy targeted experiments in real time. Future work should extend real-time perturbation capabilities, consider additional drivers (e.g., North Atlantic SSTs), refine regional thresholds and skill filters, and integrate social science to optimize risk communication and decision support.

- Limited hindcast sample sizes (typically 20–30 years per system; 44 years for DP3) constrain robust evaluation of state-dependent skill and joint-driver effects for rare extremes. - Teleconnection analyses establish consistency but not strict causality; attribution relies on perturbation experiments with partial realization of desired SST signals due to weak ocean relaxation and unchanged atmospheric nudging, requiring signal inflation. - Average skill estimates may underrepresent conditional skill during strong-driver years and vary across models (some with low or near-zero correlations). - Other potential drivers (e.g., North Atlantic variability) were not comprehensively tested for 2022; combined influences remain uncertain. - Tool evaluation depends on observational datasets and methodological choices (e.g., semi-standard deviations, region definitions, skill thresholds) and may filter out or miss certain regions/events; false alarms remain a concern.