The choice of the Madden-Julian Oscillation (MJO) as the centerpiece of the study is no accident. The MJO is a vast, eastward-propagating envelope of tropical convection that circles the planet at roughly 5 meters per second, with a characteristic cycle of 40 to 50 days—a fact first documented by Roland Madden and Paul Julian in 1972. It is the dominant source of weather variability on weekly-to-monthly timescales in the tropics and, through atmospheric teleconnections, exerts a powerful influence on precipitation over regions as distant as the western United States. For many forecast applications, accurately simulating the MJO is a prerequisite for skillful S2S precipitation prediction. Yet despite five decades of research, the MJO remains notoriously difficult for even the most advanced Earth system models to capture, and disagreements about its fundamental generation mechanisms persist.

The team targeted three well-observed MJO events that occurred during the 2011 Dynamics of the Madden-Julian Oscillation (DYNAMO) field campaign, spanning mid-October through December 2011. These events offer an exceptionally well-documented framework for evaluating how models represent MJO-related convection and precipitation. The experimental design was elegantly layered. A control simulation, launched on 1 September 2011, used ERA5 atmospheric fields—generated through the HICCUP tool—for its atmospheric initial conditions and land states derived from an offline E3SM Land Model run forced by ERA5 data stretching back to 1979. Ten additional experiments, initialized at approximately five-day intervals from 5 September to 20 October, formed an ensemble whose mean the authors call E_ERA5. Against these, the team ran two counterfactuals: one in which only the land initial conditions came from an equilibrium state, and one—the fully equilibrium experiment—in which both the atmosphere and land began from a balanced state obtained after roughly twenty years of coupled model simulation.

The results are striking in their consistency. When precipitation from the experiments was compared against the CMORPH satellite-based precipitation dataset over the period from 20 October to 31 December 2011, the equilibrium-initialized simulation showed a time-averaged precipitation difference of 0.89 mm/day in the MJO region, while the ERA5-initialized control showed only −0.32 mm/day. Globally, the gap was smaller but still clear: 0.55 mm/day for the equilibrium run versus just 0.02 mm/day for the ERA5-initialized experiment. Root-mean-square errors told the same story—14.56 mm/day in the MJO region for the ERA5-initialized control versus 15.01 mm/day for the equilibrium case, and 9.45 versus 9.73 mm/day globally. In other words, the traditional equilibrium approach, prized because it minimizes the model’s adjustment shock, produced systematically worse precipitation simulations once the model had settled down.

Perhaps the most consequential finding concerns how long these initialization effects endure. By tracking the evolution of precipitation errors as a function of lead time across the eleven ERA5-initialized experiments, the researchers found that the simulations generally reached their optimal performance around 40 days of integration in the MJO region and around 50 days in the global domain. The control run, initialized from realistic conditions, actually showed larger early errors than the equilibrium experiment during the first weeks—evidence of the so-called initialization shock as the model adjusts toERA5-imposed atmospheric fields—but its errors declined below those of the equilibrium run after approximately 40 days. The authors note that the roughly ten-day lag between the global and MJO-region persistence timescales mirrors the 5-to-9-day lag with which MJO influences propagate to the midlatitudes, an intriguing hint that the memory of initial conditions is physically coupled to the oscillation itself.

To place these results on firm statistical footing, the team conducted additional ten-member ensemble experiments, perturbing initial atmospheric temperatures with small-amplitude white noise. They then evaluated MJO prediction skill using the bivariate anomaly correlation coefficient (ACC) applied to the real-time multivariate (RMM) index, a standard metric constructed from observed outgoing longwave radiation (OLR) and zonal winds at 850 and 200 hPa following Wheeler and Hendon’s widely used methodology. An ACC above 0.5 is conventionally regarded as useful forecast skill. The ERA5-initialized control maintained ACC values above 0.5 for roughly the first 65 lead days during the first MJO event—longer than reported in many previous studies, a result the authors attribute partly to the relatively weak MJO activity in that period. More tellingly, Student’s t-tests across the ensembles showed that differences between equilibrium and ERA5 land initializations were statistically significant at the 95 percent confidence level across lead days 44–54, and that differences attributable to atmospheric initial conditions remained significant across lead days 44–58. ERA5-based land initial conditions also reduced the ACC spread by an average of 0.15 over that window.

The mechanism work is where the study makes its most novel contribution. In the MJO region, the improved precipitation skill turned out to be tightly linked to a better representation of outgoing longwave radiation, the satellite-observed signature of deep tropical convection. The correlation between precipitation and OLR errors in the E_ERA5 ensemble reached 0.5 over lead days 11 to 73, and the OLR error itself reached its minimum around day 40—mirroring the precipitation error evolution almost exactly. Zonal wind errors, after an initial spike from the initialization shock, stabilized near day 40 at both 200 and 850 hPa. At the global scale, by contrast, the controlling variable was surface latent heat flux—the evaporation-driven energy exchange between surface and atmosphere—whose errors correlated with precipitation errors at 0.72. The authors interpret this as evidence that regional tropical precipitation is fundamentally a convection problem, while global precipitation is more strongly constrained by the atmospheric energy budget.

The land component of the story centers on the Maritime Continent, the archipelagic region of Indonesia and surrounding islands where MJO convection often stalls or reorganizes. When the team compared their ERA5-land and equilibrium-land ensembles during 15–25 October 2011—the window of maximum divergence—they found that the equilibrium-land simulation systematically underestimated surface latent heat flux over equatorial land areas relative to ERA5. That deficit limits the transport of moisture from the land surface into the lower troposphere, drying the planetary boundary layer and creating hostile conditions for deep convection: rising convective parcels suffer enhanced entrainment dilution and evaporative cooling, suppressing convective development. The ERA5-land experiment, with more realistic moisture fluxes, sustained a moister lower troposphere and more robust deep convection. The equilibrium-land run also displayed a widespread warm bias in surface temperature over equatorial land, which the ERA5-land initialization substantially reduced, restoring more realistic land-surface energy partitioning and boundary-layer thermodynamics.

These surface improvements propagated upward into the convection and circulation fields. The equilibrium-land run overestimated OLR near 10°N, 90°E and underestimated it near 150°E—hallmarks of misplaced convective activity—while the ERA5-land experiment sharply reduced those biases, particularly near 150°E. At 850 hPa, the equilibrium-land run’s overestimate of zonal winds near 130°E, reflecting distorted low-level convergence, was likewise corrected. Together, the chain of evidence—surface fluxes to boundary-layer moisture, moisture to convection, convection to OLR and circulation—demonstrates that land initial conditions act as a genuine, statistically significant, though secondary, source of MJO predictability through their modulation of land–atmosphere coupling.

The findings carry practical weight for the forecasting community. S2S prediction occupies a critical gap in operational meteorology: weeks three through six, beyond the reach of deterministic weather prediction but before the slow ocean drivers of seasonal climate dominate. Current forecast systems, from ECMWF’s SEAS5 to NOAA’s unified forecast system, depend heavily on initialization quality, and the results here suggest that investing in realistic land-surface initialization—alongside atmospheric data assimilation—can yield measurable gains at timescales where skill is scarce. The study also highlights an unresolved tension: a 2024 analysis of the Community Earth System Model version 2 by Richter and colleagues found no significant impact of land initial state on S2S skill, a discrepancy the authors suggest may reflect how differently models represent atmosphere-land interactions, and how those differences become magnified during active MJO periods.

The authors are careful to frame their work as a case study, albeit one whose robustness is supported by eleven experiments initiated across September and October, and they acknowledge limitations: ten-member ensembles cannot establish strong statistical significance on their own, and ocean initial conditions—the most obvious missing ingredient, given that MJO events predominantly occur over the ocean—were not considered. Ocean coupling, they note, warrants investigation as a next step, ideally through coupled data assimilation approaches that would initialize atmosphere, land, and ocean coherently. In an era when exascale computing is making such ambitious initialization strategies feasible, this study offers both a technical roadmap and a clear demonstration of the payoff: better initial conditions, even imperfect ones, remember themselves far longer than the field has generally assumed—long enough, in fact, to matter for the forecasts that people most need.