Improved real-time prediction of zonally propagating extratropical intraseasonal oscillations in the upgraded IAP-CAS S2S system

Reliable subseasonal prediction of extratropical intraseasonal variability remains a major challenge, despite its importance for Eurasian high-impact weather. Here we evaluate real-time forecasts of three dominant zonally propagating extratropical intraseasonal oscillation (EISO) modes in the upgraded IAP-CAS subseasonal-to-seasonal (S2S) prediction system from 21 August 2024 to 30 April 2026. Compared with major operational S2S models, the upgraded IAP-CAS system reaches prediction day limits of 12, 14, and 19 days for EISO-SJE, EISO-PJE, and EISO-PJW, respectively, within 0–1 day of the leading system, and achieves mean real-time prediction scores of 74.8, 76.9, and 78.6. Process diagnostics show that this improvement is associated with reduced upper-tropospheric wind and geopotential-height errors along the subtropical and polar-front jet waveguides, particularly near the jet entrance and exit sectors, indicating a more accurate large-scale circulation environment for EISO propagation. The revised second-order exact sampling strategy also alleviates ensemble underdispersion and improves spread-error consistency in EISO phase space. Composite analyses further show that the upgraded system better captures the propagation properties and spatial structures of EISO-SJE, EISO-PJE, and EISO-PJW, although amplitude damping persists at subseasonal leads. These results identify a physically grounded pathway for improving real-time extratropical S2S prediction through targeted initialization and ensemble-design upgrades.

Subseasonal-to-seasonal (S2S) prediction bridges medium-range weather forecasting and seasonal climate outlooks (Vitart et al., 2017; White et al., 2022). This bridge is needed because deterministic weather prediction is fundamentally constrained by the chaotic nature of the atmosphere. Small uncertainties in the initial atmospheric state grow rapidly through non-linear interactions, imposing an intrinsic limit on deterministic predictability (Lorenz, 1963, 1969). Although numerical weather prediction has advanced substantially in recent decades, useful deterministic weather forecasts generally extend to approximately 2 weeks, with the exact limit depending on the variable, spatial scale, initial-condition uncertainty, model error, and prevailing circulation regime (Palmer, 2000; Bauer et al., 2015). Recent non-linear predictability studies further show that these limits are strongly state- and region-dependent. For example, summer surface air temperature over East Asia has an estimated upper predictability limit of about 12 days, with distinct error-growth dynamics between northern and southern East Asia (Li et al., 2026). Heatwave events over South China exhibit local predictability limits ranging from 4 to 12 days, reflecting event-to-event differences in non-linear error growth (Li et al., 2023). For the 2021 sudden stratospheric warming event, the practical local predictability limit was estimated to be about 17 days, with central Eurasia identified as a sensitive region for rapid forecast-error growth (Zhang et al., 2025). These studies highlight that prediction limits cannot be treated as a fixed universal number, but must be assessed for specific variables, regions and dynamical phenomena.

Beyond the deterministic weather-prediction range, forecast skill increasingly relies on slowly evolving boundary conditions and organized large-scale modes of atmospheric variability that retain memory beyond the synoptic timescale. S2S prediction therefore provides a critical window for early warning, preparedness, and climate-risk management (Vitart and Robertson, 2018; Domeisen et al., 2022; Vitart et al., 2025). Forecasts at lead times of 2–6 weeks are especially valuable for anticipating persistent high-impact events, including heatwaves, cold spells, heavy rainfall, and compound extremes (Pyrina and Domeisen, 2023; Yan et al., 2025; Zhu and Dai, 2026). Intraseasonal oscillations (ISOs) are central to this extended-range predictability because they provide a dynamical pathway through which coherent anomalies can persist, propagate, and modulate regional weather regimes for several weeks (Zhang, 2005; Waliser, 2011; Zhu et al., 2024). Accordingly, the prediction day limit of organized intraseasonal modes offers an intuitive and physically meaningful measure of how far useful forecast skill can be extended beyond the conventional deterministic weather-prediction range.

The Madden-Julian Oscillation (MJO) is the best-established tropical source of S2S predictability (Gottschalck et al., 2010; Jiang et al., 2020; Lin, 2022). Its planetary-scale convective and circulation anomalies affect tropical rainfall directly and influence the extratropics through Rossby-wave teleconnections, thereby shaping subseasonal variations in temperature, precipitation, and circulation regimes (Zhou et al., 2020; Deoras et al., 2023; Cui and Maloney, 2026). Advances in model physics, initialization, ensemble generation, and process understanding have substantially improved MJO prediction in modern operational systems, with useful skill often extending to 3–5 weeks (Vitart, 2017; Kim et al., 2018; Newman et al., 2020; Liu et al., 2024). Nonetheless, amplitude damping, propagation biases, and insufficient ensemble dispersion still constrain realized skill relative to potential predictability (Liu et al., 2017; Zeng et al., 2023). The progress and remaining limitations of MJO prediction illustrate a broader requirement for S2S development, i.e., verification must be connected to the dynamical structures that carry predictability.

Extratropical intraseasonal variability is less mature as an operational prediction target, despite its direct relevance to climate extremes. Over Eurasia, mid- to high-latitude circulation anomalies organized along the upper-tropospheric westerly jets modulate persistent temperature and precipitation extremes (Yang et al., 2023, 2024; Zhu et al., 2025; Mascolo et al., 2025). Recent studies have identified three dominant zonally propagating extratropical intraseasonal oscillation (EISO) modes: an eastward-propagating mode along the subtropical westerly jet (EISO-SJE), an eastward-propagating mode along the polar-front jet (EISO-PJE), and a westward-propagating mode along the polar-front jet (EISO-PJW) (Zhu et al., 2023b). These modes provide physically interpretable targets for monitoring and prediction because they are embedded in Rossby-waveguide environments and are expressed through upper-tropospheric meridional-wind and geopotential-height anomalies.

A key advance toward operational EISO prediction was the development of real-time EISO indices that isolate these modal signals from observations and forecasts without using future information. With these metrics, Zhu et al. (2024) provided the first systematic hindcast assessment of EISO prediction across current S2S operational models. They showed that the three modes are predictable but only within limited windows, with useful lead times of approximately 9–20 days depending on the mode and model. They also documented common amplitude underestimation, propagation errors, and strong state dependence of skill. In that benchmark, the IAP-CAS system developed at the Institute of Atmospheric Physics, Chinese Academy of Sciences, showed moderate EISO skill relative to other international systems. However, targeted sensitivity experiments indicated that moisture-aware initialization and larger ensembles can substantially improve EISO forecasts. These results identified a testable route for model development: better initial thermodynamic-dynamic balance and a more adequate representation of ensemble uncertainty should improve EISO prediction.

Motivated by these findings, the IAP-CAS fully coupled S2S prediction system was upgraded for real-time operations on 21 August 2024. The upgraded system incorporates additional moisture information in the initialization and uses a second-order exact sampling strategy to generate a fixed 49-member ensemble (Liu et al., 2026). This configuration was designed to better represent flow-consistent forecast uncertainty while remaining feasible for daily operational S2S prediction. A recent assessment showed that the upgraded system markedly improves MJO prediction (Liu et al., 2026), demonstrating the value of the new initialization and ensemble framework for a tropical intraseasonal mode. Whether the same operational upgrade improves extratropical intraseasonal prediction, and through which physical pathways, has not yet been established.

This gap matters because improved MJO skill does not necessarily imply improved EISO skill. Tropical convective modes and extratropical waveguide-trapped circulation anomalies differ in their dominant dynamics, error-growth pathways, and verification targets (Hoskins and Ambrizzi, 1993; Branstator, 2002). EISO prediction depends on the representation of the subtropical and polar-front jet waveguides (Branstator, 2002; Zhu and Yang, 2021; Zhu et al., 2023a), barotropic and baroclinic processes (Chang, 2001; Dai et al., 2021; Zhu and Yang, 2021), and multiscale energy interactions (Cai and Mak, 1990). Moreover, increasing the ensemble size is beneficial only if ensemble spread grows consistently with forecast error (Fortin et al., 2014). A credible real-time improvement should therefore be visible not only in higher skill scores, but also in reduced circulation errors along the jet waveguides, improved spread-error consistency, and more realistic EISO propagation and phase structure.

Here we evaluate real-time prediction of the three EISO modes in the upgraded IAP-CAS S2S system and benchmark its performance against major international operational S2S models. We first quantify whether the upgraded system has entered the leading group of real-time EISO prediction using bivariate correlation, real-time prediction score, amplitude error, and phase-angle error. We then diagnose the physical basis of the improvement by examining upper-tropospheric circulation errors along the subtropical and polar-front jet waveguides. We further test whether the second-order exact sampling strategy improves ensemble reliability by reducing the mismatch between ensemble spread and ensemble-mean error in EISO phase space. Finally, we use lag-longitude and phase-longitude composites to determine whether the upgraded system better captures EISO propagation and spatial structure. By linking real-time operational verification with process-level diagnosis, this study extends previous EISO work from hindcast assessment to real-time model evaluation and extends recent IAP-CAS system development from tropical MJO prediction to extratropical intraseasonal variability.

Datasets and model configurations

The fifth-generation European Center for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalysis (ERA5; Hersbach et al., 2020) was used as the observational reference throughout the study. Daily zonal wind (U), meridional wind (V), and geopotential height (Z) were obtained at upper-tropospheric pressure levels. The primary fields were U200, V200, and Z200, which were used for circulation-error diagnosis and EISO index construction following the real-time EISO framework of Zhu et al. (2024). All observational and forecast fields were interpolated to a common 1.5° × 1.5° grid. ERA5 also supplied the pre-initialization observed segments needed to construct real-time EISO indices at short forecast leads.

The upgraded IAP-CAS operational S2S system, hereafter IAP-CASnew, was evaluated for 21 August 2024–30 April 2026. The same period was used for all available operational S2S models in the benchmark: the European Center for Medium-Range Weather Forecasts (ECMWF), UK Met Office (UKMO), National Centers for Environmental Prediction (NCEP), China Meteorological Administration (CMA), Japan Meteorological Agency (JMA), Korea Meteorological Administration (KMA), Environment and Climate Change Canada (ECCC), Hydrometeorological Center of Russia (HMCR), Institute of Atmospheric Sciences and Climate, National Research Council of Italy (ISAC-CNR), and IAP-CASnew. Model configurations are summarized in Table 1.

IAP-CAS officially switched to IAP-CASnew in real-time operations on 21 August 2024 and has since run daily forecasts. Relative to the previous operational configuration, the upgraded system includes two major changes. First, atmospheric moisture information is incorporated into the initialization through reanalysis nudging and forecast nudging. Second, the ensemble perturbation strategy was revised using second-order exact sampling, with the ensemble size fixed at 49 members. Unlike the previous time-lagged perturbation approach, which constructs ensemble perturbations from differences between forecasts initialized at different times, SOES generates dynamically balanced perturbations from a statistically representative sample of the model's intrinsic variability while enforcing second-order consistency constraints. Consequently, the initial perturbations evolve more consistently with the model dynamics, providing a more realistic representation of forecast uncertainty and reducing the systematic underdispersion commonly found in the previous operational configuration. The use of offline perturbation generation also allows the ensemble size to be flexibly adjusted without increasing the computational cost of the operational forecast integration. A detailed description of the SOES algorithm is available in Liu et al. (2026). The previous operational configuration, hereafter IAP-CASold, is used as the internal reference system for diagnostic comparisons. Detailed descriptions of the new and old systems, including initialization, assimilation, running frequency, and ensemble configuration, are provided in earlier model papers and in the S2S database documentation (Bao et al., 2019; Zeng et al., 2023; Liu et al., 2024, 2026; Lamichhane et al., 2025; https://confluence.ecmwf.int/display/S2S/Models).

For the multi-model benchmark, all available ensemble members from each operational system were averaged to obtain the ensemble-mean forecast unless otherwise stated. For the spread-error analysis, individual ensemble members were retained to calculate ensemble spread in EISO phase space.

Methods

Real-time EISO indices

The three EISO modes were quantified using the real-time metric framework of Zhu et al. (2024). For each model and forecast lead, anomalies were calculated by subtracting the corresponding model daily climatology as a function of calendar date and lead time. Following Zhu et al. (2024), the variables and domains used for index construction were V200 over the subtropical-jet domain (25°-55°N, 15°W−150°E) for EISO-SJE, V200 over the polar-front-jet domain (55°-85°N, 25°-150°E) for EISO-PJE, and Z200 over the same polar-front-jet domain for EISO-PJW.

Variability slower than the target intraseasonal bands was removed by subtracting the mean of the most recent T days from the anomalous field. T was set to 9, 11, and 15 days for EISO-SJE, EISO-PJE, and EISO-PJW, respectively, following the real-time EISO metric design. At short leads, this recent-T-day mean combines observations before initialization with forecast fields after initialization. For example, a 1-day lead uses the previous T-1 observed days plus the first forecast day, whereas a 5-day lead uses the previous T-5 observed days plus the first five forecast days. A previous-3-day running mean was then applied to suppress synoptic-scale noise.

The processed anomalies were zonally weighted and projected onto the observed EISO EOF patterns to obtain two standardized projection coefficients for each mode. These coefficients define the two-dimensional real-time EISO indices: EISO-SJE = [SJE1, SJE2], EISO-PJE = [PJE1, PJE2], and EISO-PJW = [PJW1, PJW2]. The EOF patterns and normalization standard deviations were taken from the observed EISO metric framework.

The EISO amplitude was defined as: A (t) = P C 1 2 (t) + P C 2 2 (t). Strong EISO events were defined as days with A > 1.5, following Zhu et al. (2023b). The eight EISO phases were determined from the signs and relative magnitudes of the two standardized projection coefficients, consistent with the phase definitions in the real-time EISO metric framework.

Forecast skill metrics

Four complementary metrics were used to evaluate EISO forecast quality: bivariate correlation (BCOR), real-time prediction score (RPS), amplitude error (ERRa), and phase-angle error (ERRp).

BCOR measures the ability of a forecast system to follow the observed EISO trajectory in two-dimensional phase space and was calculated as:

where O 1 (t) and O 2 (t) are the observed EISO indices at verification time t, and P 1 (t, τ) and P 2 (t, τ) are the corresponding forecasts at lead time τ. The prediction day limit (PDL) was defined as the first lead day when BCOR falls below 0.5.

RPS was calculated following QX/T 638-2022, the Chinese meteorological industry standard for ISO real-time verification (China Meteorological Administration, 2022). Unlike BCOR, which summarizes average phase-space association across many initializations, RPS emphasizes the quality of individual real-time forecasts, namely whether a single run follows the phase progression and intensity evolution of an EISO event. The RPS calculation is:

where I CORr is the real-time correlation skill score for a certain EISO forecast, O 1 (T) and O 2 (T) are the observed EISO indices on day T, and P 1 (T) and P 2 (T) are the corresponding predicted value. M is the length of forecast days considered for the real-time verification, with the recommended value of 30 days following QX/T 638-2022. Using the real-time correlation skill score I CORr, the final R P S = 1 + I C O R r 2 × 100. The value range of RPS is from 0 to 100. The larger RPS means the complete EISO process can be accurately captured.

ERRa was used to quantify the relative error in predicted EISO amplitude:

Positive ERRa denotes amplitude overestimation, whereas negative ERRa denotes amplitude underestimation.

Unlike the signed phase error used in some previous studies, we use a wrapped absolute phase-angle error to quantify the magnitude of phase displacement without distinguishing whether the predicted propagation is too fast or too slow. The predicted and observed phase angles are first calculated as:

The absolute phase difference is then wrapped onto [0, π]:

The phase-angle error is then:

ERRp is therefore non-negative and represents the mean phase-angle error in degrees.

Diagnosis of circulation errors

To determine whether the upgraded system improves the large-scale state relevant to EISO prediction, we compared circulation errors between IAP-CASnew and IAP-CASold. Root-mean-square error (RMSE) was calculated as:

where P i (τ) and O i are forecast and observed values at grid point i, and w i = cos(ϕ i) is the latitude weight.

Ensemble spread and spread–error consistency

To assess ensemble reliability in EISO phase space, we compared ensemble spread with the RMSE of the ensemble mean for IAP-CASnew and IAP-CASold. The RMSE of the ensemble mean in two-dimensional phase space was calculated as:

The ensemble spread was calculated as:

where P 1, num and P 2, num are the predicted EISO indices for member num, and NUM is the ensemble size of the corresponding forecast system. For a statistically reliable ensemble, the ensemble spread should approximately match the RMSE of the ensemble mean. A systematic spread smaller than RMSE indicates underdispersion and overconfidence (Whitaker and Loughe, 1998; Fortin et al., 2014).

Statistical uncertainty

Unless otherwise stated, forecast skill and diagnostic quantities were calculated using all available real-time initializations during the verification period. Sampling uncertainty was estimated with bootstrap resampling over initialization dates. For paired comparisons between IAP-CASnew and IAP-CASold, the same initialization dates were resampled jointly to preserve the paired structure. A total of 1,000 bootstrap samples were used; differences were considered robust when the 95%, 99%, or 99.9% confidence interval did not include zero, as specified for each diagnostic.

The upgraded IAP-CAS system becomes a top-tier operational model for real-time EISO prediction

We first benchmark the upgraded IAP-CAS operational system (IAP-CASnew) against other state-of-the-art real-time S2S operational models during 21 August 2024–30 April 2026 (Figures 1–4). Forecast quality is evaluated with the real-time EISO metric suite (Methods): BCOR and PDL quantify phase-space tracking skill, RPS measures case-wise forecast quality, and ERRa and ERRp separate amplitude and phase errors.

BCOR decreases with lead time for all three EISO modes in all models (Figure 1). EISO-PJW generally has higher skill than the two eastward-propagating modes, EISO-SJE and EISO-PJE, consistent with the hindcast benchmark of Zhu et al. (2024). Specifically, PDLs span 9–13 days for EISO-SJE, 10–15 days for EISO-PJE, and 13–19 days for EISO-PJW. IAP-CASnew reaches PDLs of 12, 14, and 19 days for EISO-SJE, EISO-PJE, and EISO-PJW, respectively, placing it within 0–1 day of the best-performing system in this real-time window (ECMWF: 13, 15, and 19 days).

RPS provides a complementary event-oriented perspective (Figure 2). IAP-CASnew has a high central tendency and fewer low-score cases, indicating that its improvement is not confined to a small subset of easy forecasts. Mean RPS values for EISO-SJE range from 69.6 to 75.3, with IAP-CASnew reaching 74.8, second only to ECMWF. For EISO-PJE, values range from 70.7 to 77.3, with IAP-CASnew reaching 76.9. For EISO-PJW, values range from 73.5 to 79.2, with IAP-CASnew reaching 78.6.

The error-partition metrics identify the main remaining limitations. ERRa becomes increasingly negative with lead time after approximately 1 week in all models, indicating systematic amplitude damping at subseasonal leads (Figure 3), consistent with Zhu et al. (2024). By the end of week 4, ERRa ranges from about −1.14 to −0.45 for EISO-SJE, −1.24 to −0.62 for EISO-PJE, and −1.21 to −0.63 for EISO-PJW. For IAP-CASnew, ERRa at lead 28 days is approximately −1.10, −1.19, and −1.10 for the three modes, respectively. ERRp increases smoothly with lead time as phase uncertainty accumulates (Figure 4): the multi-model range increases from about 19.26° at lead 1 to 93.37° at week 4 for EISO-SJE, from 17.72° to 95.15° for EISO-PJE, and from 19.00° to 90.03° for EISO-PJW. IAP-CASnew generally lies near the lower-error envelope, with lead-28 ERRp values of approximately 84.25°, 88.28°, and 82.03° for EISO-SJE, EISO-PJE, and EISO-PJW.

Together, these results show that IAP-CASnew has entered the leading group of real-time EISO prediction systems, with PDLs close to the best operational model and consistently strong event-level scores. At the same time, amplitude damping and phase-error growth remain substantial. The following sections therefore examine two mechanisms that can account for the gain: improved waveguide-region circulation associated with the upgraded initialization and improved ensemble reliability associated with the revised perturbation strategy.

Improved initialization substantially reduces circulation errors over the Eurasian jet waveguides

We next examine whether higher real-time EISO skill is associated with a more accurate upper-tropospheric circulation state over the Eurasian jet waveguides that host and guide the three EISO modes (Hoskins and Karoly, 1981; Branstator, 2002; Zhu et al., 2023b). The upgraded initialization incorporates moisture information. Although moisture is not a direct factor triggering and maintaining EISOs, improved moisture analyses can modify diabatic heating and divergent circulation, which then feed back onto balanced winds and geopotential height (Branstator, 2002; Wu et al., 2020; Zeng et al., 2023). Because EISOs are Rossby-wave-like anomalies embedded in the subtropical and polar-front jet waveguides, smaller errors in upper-level winds and geopotential height should directly benefit phase evolution and spatial-structure prediction.

At the initial stage, IAP-CASnew shows widespread RMSE reductions in U200, V200, and Z200 relative to IAP-CASold across Eurasia (Figures 5c, f, i), with the most coherent improvements in U200 and Z200. The reductions are organized along the subtropical and polar-front jet waveguides. The old-minus-new RMSE differences indicate that initial U200 errors decrease by about 0.25–1.50 m s −1 over broad areas, with reductions exceeding 1.0 m s −1 along the southern flank and entrance/exit sectors of the subtropical jet (Figure 5c). V200 errors decrease by about 0.08–0.40 m s −1 across much of Eurasia, with locally stronger improvements near the entrance and exit sectors of the two jets (Figure 5f). Z200 errors decrease by 3.00–10.00 gpm over much of the domain, with the largest improvements over East Asia, the western North Pacific, and parts of subtropical Eurasia (Figure 5i).

It is worth noting that the largest improvements occur near jet entrance and exit regions, where both IAP-CAS versions also have relatively large errors (Figures 5a, b, d, e, g, h). This co-location is dynamically important. Jet entrance and exit sectors favor eddy growth, wave-packet amplification, and energy conversion along the waveguide (Hoskins and Ambrizzi, 1993; Hare and James, 2001; Chang, 2001). For EISO modes, whose predictability depends on coherent propagation along the subtropical and polar-front jets, improved circulation in these sectors is therefore especially consequential.

These circulation improvements persist into the short-range forecast. Mean RMSE differences for lead times 1–5 days show patterns broadly consistent with the initial-stage differences (Figure 6). U200 reductions remain widespread and are typically 0.50–1.20 m s −1, again with stronger reductions near the subtropical-jet entrance and exit sectors (Figure 6c). V200 improvements are larger than at initialization, with coherent positive differences over the waveguide regions and East Asia (Figure 6f). Z200 improvements remain widespread, mostly in the range of 2.67–10.00 gpm (Figure 6i). Because these lead-1–5-day maps mainly diagnose the immediate influence of the upgraded initialization, we further examine whether the error reductions are maintained throughout the subseasonal forecast range that is directly relevant to EISO prediction.

Figure 7 presents the lead-time evolution of regional-mean RMSE from day 1 to 30 over the subtropical-jet and polar-front-jet waveguide regions. The improvement is not confined to the initial state or the first few forecast days. Over the subtropical-jet waveguide, U200 RMSE decreases from 4.69 to 4.15 m s −1 at day 1 and from 13.93 to 12.14 m s −1 by day 30, a reduction of 0.54–1.79 m s −1 across the lead range (Figure 7a). Over the polar-front-jet waveguide, U200 RMSE decreases from 3.68 to 3.37 m s −1 at day 1 and from 11.49 to 9.94 m s −1 by day 30, a reduction of 0.31–1.55 m s −1 (Figure 7b). V200 reductions are smaller but persistent: from 4.58 to 4.37 m s −1 at day 1 and from 12.68 to 10.94 m s −1 by day 30 over the subtropical-jet waveguide (Figure 7c), and from 4.13 to 4.05 m s −1 at day 1 and from 12.32 to 10.72 m s −1 by day 30 over the polar-front-jet waveguide (Figure 7d). Z200 reductions are substantial in absolute magnitude, from 33.90 to 30.88 gpm at day 1 and from 124.09 to 107.61 gpm by day 30 over the subtropical-jet waveguide (Figure 7e), and from 39.49 to 36.49 gpm at day 1 and from 182.62 to 158.09 gpm by day 30 over the polar-front-jet waveguide (Figure 7f). The RMSE reductions increase rapidly during the first 10–15 forecast days and then become quasi-stationary, indicating that the upgraded system improves not only the starting state and early forecast evolution but also the large-scale waveguide circulation throughout the S2S prediction range. This sustained reduction in subseasonal circulation errors provides a physically interpretable pathway from improved waveguide representation to the enhanced EISO phase-space skill shown in Section 3.1.

Second-order exact sampling improves ensemble consistency by reducing underdispersion

Because the revised SOES strategy was designed to provide a more realistic representation of initial forecast uncertainty than the previous time-lagged perturbation method, we next evaluate whether this design translates into improved ensemble reliability. In an ideal ensemble system, ensemble spread should be comparable to the RMSE of the ensemble mean; persistent spread smaller than RMSE indicates underdispersion and overconfidence. Figure 8 compares ensemble-mean RMSE and ensemble spread in EISO phase space for IAP-CASold and IAP-CASnew.

IAP-CASold is strongly underdispersive during weeks 1–2 for all three modes. At lead 10 days, the spread-to-RMSE ratio is 0.26 for EISO-SJE (RMSE = 1.13; spread = 0.29; Figure 8b), 0.25 for EISO-PJE (RMSE = 1.17; spread = 0.29; Figure 8d), and 0.20 for EISO-PJW (RMSE = 1.33; spread = 0.26; Figure 8f), corresponding to RMSE-spread gaps of 0.84, 0.88, and 1.07. Underdispersion remains pronounced at lead 15 days, with gaps of 0.77 (EISO-SJE: 1.59–0.82; Figure 8b), 0.76 (EISO-PJE: 1.54–0.78; Figure 8d), and 0.72 (EISO-PJW: 1.53–0.81; Figure 8f). These deficits show that the old ensemble substantially under-represents forecast uncertainty during the critical week-2 window, when phase errors begin to accumulate.

IAP-CASnew shows much better spread-error consistency. At lead 10 days, the spread-to-RMSE ratio increases to 0.81 for EISO-SJE (RMSE = 1.11; spread = 0.90; Figure 8a), 0.77 for EISO-PJE (RMSE = 1.19; spread = 0.92; Figure 8c), and 0.72 for EISO-PJW (RMSE = 1.41; spread = 1.02; Figure 8e), reducing the RMSE-spread gaps to 0.21, 0.27, and 0.39. At lead 15 days, the gaps decrease from 0.77 (IAP-CASold; Figure 8b) to 0.11 (IAP-CASnew; Figure 8a) for EISO-SJE, from 0.76 (IAP-CASold; Figure 8d) to 0.17 (IAP-CASnew; Figure 8c) for EISO-PJE, and from 0.72 (IAP-CASold; Figure 8f) to 0.08 (IAP-CASnew; Figure 8e) for EISO-PJW. By lead 20 days, spread and RMSE are nearly matched for EISO-SJE and EISO-PJW in the new system, with gaps of 0.05 and 0.04 (Figures 8a, e), whereas the old system still has gaps of 0.43 and 0.51 (Figures 8b, f). Even at lead 30 days, IAP-CASnew remains closer to perfect consistency for EISO-PJW (gap = 0.02 vs. 0.25; spread/RMSE = 0.99 vs. 0.86; Figures 8e vs. 8f).

These results indicate that the revised SOES perturbation strategy substantially mitigates underdispersion and produces a more statistically reliable ensemble in EISO phase space. By providing a more realistic representation of forecast uncertainty, the revised ensemble better captures the growth of EISO prediction errors and improves the consistency between ensemble spread and ensemble-mean error. This enhanced ensemble reliability provides an important foundation for the improved real-time prediction skill documented in Section 3.1.

The upgraded system better captures the propagation and phase structure of EISO

Finally, we test whether the improvements documented above translate into a more faithful representation of EISO dynamics. Two composite diagnostics are used for strong observed events: lag-longitude composites of the 15-day evolution after strong P1 events, which diagnose propagation (Figure 9), and phase-longitude composites for strong P1–P8 events, which diagnose spatial structure (Figure 10). Conditioning on observed strong phases ensures that the two model versions are evaluated against the same dynamically coherent EISO events. Although the propagation is diagnosed using strong P1 events here, the conclusions are not sensitive to the phases.

The propagation composites show that IAP-CASnew more realistically reproduces the observed phase progression of EISO wave trains (Figure 9). For EISO-SJE, observations show a clear eastward-propagating sequence of alternating upper-tropospheric V200 anomalies, with coherent tilted bands persisting for nearly 2 weeks (Figure 9a). IAP-CASnew captures the banded structure and downstream tilt more faithfully than IAP-CASold, which shows more distorted phase bands and an overly strong early anomaly stripe (Figures 9b, c). For EISO-PJE, the observed evolution is smoother and more large-scale (Figure 9d). Both systems reproduce the gross sign transitions, but IAP-CASnew better preserves the coherent tilted structure, whereas IAP-CASold is more diffuse and loses coherence earlier (Figures 9e, f). For EISO-PJW, the observed Z200 composite shows a westward-evolving large-scale geopotential-height structure (Figure 9g). Both forecasts reproduce the broad negative anomaly band, but IAP-CASold produces a stronger and more eastward-centered negative anomaly during the early stage (Figure 9i); IAP-CASnew is closer to the observed distribution, although a negative bias remains (Figure 9h).

The phase composites further show improved phase organization in IAP-CASnew (Figure 10). For EISO-SJE, the observed phase-longitude composite has a well-organized wave-train sequence from P1 to P8 (Figure 10a). IAP-CASnew reproduces this progression and continuity more cleanly than IAP-CASold, which shows more localized phase distortions (Figures 10b, c). For EISO-PJE, the observed composite has a coherent large-scale anomaly triplet over 30°-120°E and a clear phase reversal between early and late phases (Figure 10d); IAP-CASnew better captures both the placement and continuity of this pattern (Figures 10e, f). For EISO-PJW, observations show a broad negative band during P2–P5 and a reversed positive band during P6–P8 (Figure 10g). IAP-CASnew captures the overall banded structure and phase transition more clearly than IAP-CASold, although amplitude biases persist (Figures 10h, i).

The propagation composites may appear less degraded than the rapid ERRp growth shown in Section 3.1 because the diagnostics sample different parts of the EISO distribution. ERRp is calculated over all phases and amplitudes. Weak-amplitude phases have lower signal-to-noise ratios and larger case-to-case variability, leading to faster accumulation of absolute phase-angle errors. By contrast, Figures 9, 10 condition on strong observed phases, for which propagation and phase organization are more coherent. This behavior is consistent with the hindcast benchmark showing that strong EISO events are more predictable than weak events (Zhu et al., 2024). The conditional composites therefore isolate the dynamically organized component of EISO evolution and provide process-level support for the improved real-time skill.

Overall, Figures 9, 10 show that the upgraded system represents EISO propagation and spatial structure more faithfully, whereas the amplitude damping documented in Section 3.1 remains a key limitation for maintaining subseasonal anomaly intensity.

This study provides a real-time, process-based assessment of how targeted upgrades to a fully coupled S2S prediction system can improve extratropical intraseasonal variability forecasts. Building on hindcast evidence that initialization and ensemble configuration are key controls on EISO skill, we evaluate the upgraded IAP-CAS operational system, which incorporates moisture-aware initialization and a revised ensemble generation strategy. By combining multi-model real-time benchmarking, circulation-error diagnosis, ensemble reliability assessment, and dynamical composites, we show not only that the upgraded system is competitive with leading operational S2S models, but also why the improvement emerges.

The real-time benchmark shows that IAP-CASnew has entered the leading group of operational EISO prediction systems. During 21 August 2024–30 April 2026, its PDLs reach 12, 14, and 19 days for EISO-SJE, EISO-PJE, and EISO-PJW, respectively, within 0–1 day of the best-performing operational system in this verification window. Its event-level quality is also high, with mean RPS values of 74.8, 76.9, and 78.6 for the three modes. The improvement is therefore evident not only in mean phase-space tracking but also in individual real-time forecasts. Note that the real-time skills of most models are lower than those reported in historical hindcasts, underscoring the added difficulty of operational prediction, where forecasts depend on near-real-time analyses with greater uncertainty and lower completeness than retrospective reanalyses (Vitart et al., 2017; Shao et al., 2023). This distinction is central: hindcasts provide a controlled estimate of model capability, whereas real-time forecasts measure the information actually available for operational decisions.

A central result is that improved EISO skill is linked to a more accurate Eurasian upper-tropospheric waveguide environment. The upgraded initialization directly incorporates thermodynamic moisture information rather than the dynamical fields used to define the EISO indices, yet the circulation response is consistent with coupled thermodynamic-dynamic adjustment. Improved moisture analyses can modify diabatic heating and divergent circulation, which then feed back onto balanced winds and geopotential height. In IAP-CASnew, RMSE reductions in U200, V200, and Z200 are organized along the subtropical and polar-front jet waveguides and persist from the initial state through the 1–30-day forecast range. A more accurate waveguide provides a direct physical pathway for maintaining the phase coherence, downstream development, and spatial organization of Rossby-wave-like intraseasonal anomalies.

The ensemble diagnosis provides a complementary explanation. IAP-CASold is strongly underdispersive at subseasonal leads, indicating that it substantially under-represents forecast uncertainty during the period when phase errors begin to accumulate. IAP-CASnew shows much better spread-error consistency, with ensemble spread nearly matching ensemble-mean RMSE for several modes and leads. The revised perturbation strategy therefore does more than increase ensemble size or smooth the ensemble mean; it provides a more realistic representation of uncertainty in EISO phase space. This is operationally important because overconfident ensembles can be misleading for risk-based decisions, whereas improved spread-error consistency supports more trustworthy event-level guidance.

The dynamical composites confirm that these statistical gains correspond to a more faithful representation of EISO behavior. For strong observed phases, IAP-CASnew better reproduces the eastward propagation of EISO-SJE and EISO-PJE, the westward evolution of EISO-PJW, and the longitudinal structures of all three modes. Thus, the gains in BCOR and RPS are not merely artifacts of phase-space verification; they reflect improved maintenance of the dynamical skeleton of EISO wave trains. At the same time, the composites and ERRa both show persistent amplitude damping. The upgrade improves phase organization and propagation more clearly than it improves anomaly-amplitude maintenance.

This distinction between phase/structure prediction and amplitude prediction is central to interpreting the results. BCOR and RPS show that the upgraded system better follows the coherent evolution of EISO events, especially during strong phases. ERRa shows that forecast anomalies still decay too strongly at longer leads. Such damping may arise from several non-exclusive sources, including excessive diffusion of upper-tropospheric wave anomalies (Zhu et al., 2023a), imperfect wave-mean-flow interaction (Bühler, 2009), biases in diabatic forcing and transient eddy feedbacks (Winkler et al., 2001; Fang and Yang, 2016), and insufficient maintenance of energy conversion along the jet waveguides (Zhu et al., 2023a,b). Because EISO amplitude is closely tied to downstream impacts on persistent temperature and precipitation anomalies (Yang et al., 2023, 2024), reducing this damping is essential for translating improved EISO tracking into improved high-impact weather guidance.

The contrast between coherent propagation composites and rapid all-sample ERRp growth also reveals state-dependent predictability. The composites are conditioned on strong observed phases, when the EISO signal-to-noise ratio is high and propagation is dynamically organized. ERRp, by contrast, is calculated across all phases and amplitudes. Weak-amplitude phases are more susceptible to synoptic noise, phase ambiguity, and case-to-case variability, so absolute phase-angle errors grow faster when these cases are included. Strong events therefore provide a clearer window of opportunity for subseasonal prediction, whereas weak or marginal events remain a major source of uncertainty.

Several limitations and open questions remain. First, the operational upgrade combines multiple changes, most notably moisture-aware initialization and revised ensemble perturbation generation. The diagnostics presented here identify a coherent pathway linking improved waveguide circulation and better ensemble reliability to enhanced EISO skill, but they do not fully isolate the causal contribution of each component. Targeted sensitivity experiments, including moisture-initialization on/off experiments, perturbation-strategy comparisons, and controlled ensemble-size experiments under the same initialization, are needed to quantify their relative roles.

Second, the real-time verification period is still short relative to conventional hindcast archives. Although the results are consistent across three EISO modes and multiple diagnostics, the sample may not fully represent modulation by ENSO, Arctic variability, land-atmosphere states, or other slowly varying boundary conditions. Extending the real-time archive and repeating the diagnosis across additional years will be necessary to determine whether the present leading performance is stable across different climate backgrounds and flow regimes.

Third, residual errors near jet entrance and exit regions require targeted process investigation. These sectors are where EISO disturbances can grow and organize, but they are also regions where models often struggle with baroclinic development, ageostrophic adjustment, and wave-packet evolution. The upgraded initialization reduces errors there but does not eliminate them. Future work should diagnose energy conversion, wave activity flux, eddy feedback, and balance adjustment in these sectors to identify the processes that limit further improvement.

Finally, persistent amplitude damping remains a broader challenge for operational S2S prediction. Better EISO phase evolution is necessary but not sufficient for downstream applications, because subseasonal heat, precipitation, drought, and compound-risk guidance depends strongly on whether models maintain realistic anomaly intensity. Addressing this limitation will require improvements in model physics, initialization balance, stochastic perturbation design, and post-processing. Physically constrained hybrid approaches may be particularly useful if they correct amplitude and phase biases without disrupting the dynamically coherent EISO structure captured by the upgraded system. Such approaches should build on the dynamical foundation identified here: a well-initialized jet waveguide and a reliable ensemble are prerequisites for further statistical or machine-learning correction.