Weather Adds Little Edge When Hospitals Predict Emergency Department Crowds

Every morning, hospital administrators face the same question: how many patients will walk through the emergency department doors today? Getting that answer roughly right determines whether the night shift is adequately staffed, whether beds are available, and whether waiting times spiral into dangerous overcrowding. A new multicenter study from Türkiye now offers a sobering answer to a long-running debate in emergency medicine and operations research: the weather, contrary to intuition and decades of speculation, is largely a sideshow compared with the relentless structure of the calendar and the momentum of recent attendance history.

The research, published in the journal Air Quality, Atmosphere & Health, was led by İbrahim Sarbay of Gaziosmanpaşa Training and Research Hospital in İstanbul together with colleagues at İzmir Katip Çelebi University. The team assembled daily emergency department visit counts from two tertiary hospitals, one in İstanbul and one in İzmir, spanning October 2023 through June 2025. After careful temporal alignment, 625 daily observations were available at each site. Alongside attendance figures, the researchers matched five meteorological variables: temperature, atmospheric pressure, humidity, wind speed, and precipitation. The goal was simple and ambitious: to test rigorously whether weather data add meaningful predictive power to forecasting models, or whether simpler, calendar-driven approaches suffice.

Methodologically, the study was designed as a genuine bake-off between statistical and machine-learning approaches. The authors first compared univariate forecasting models that used only the historical series of daily visits, then built multivariable models in four flavors: SARIMAX, a seasonal autoregressive integrated moving average framework that can absorb external predictors; dynamic regression, which models the target series as a function of calendar and weather covariates; and two ensemble tree methods, random forest and XGBoost, which are widely used in applied machine learning for tabular forecasting problems. Accuracy was measured with a battery of standard metrics, including root mean square error, mean absolute error, and the percentage-based measures MAPE, sMAPE, and MASE, which normalize error against a naive benchmark.

The statistical verdict was delivered with Diebold-Mariano tests, a formal method for deciding whether one forecast is significantly more accurate than another, rather than merely better by a rounding error. Robustness checks added further depth: rolling-origin validation, which repeatedly re-estimates models as new data arrive, mimicking real operational deployment; hyperparameter sensitivity analysis; ensemble averaging; and a cross-center generalization test in which models trained at one hospital were applied to the other.

The headline result is striking in its simplicity. Among univariate models, a neural network autoregression approach performed best in İstanbul, while the Theta method, a decomposition-based statistical forecaster, won in İzmir. But when calendar and weather covariates entered the picture, models relying on calendar structure alone delivered the best holdout accuracy in both cities. Dynamic regression achieved a mean absolute percentage error of 5.88 percent in İstanbul, and SARIMAX reached 4.42 percent in İzmir, both with calendar-only inputs. These are clinically meaningful accuracies, precise enough to inform staffing rosters and bed planning.

Weather augmentation, meanwhile, behaved erratically. In İstanbul, adding meteorological variables significantly worsened the SARIMAX and dynamic regression forecasts, with p-values of .013 and .015 respectively. The sole bright spot was XGBoost, where weather inputs produced a significant improvement, with p = .005. In İzmir, weather made no significant difference for any model. The authors are candid that the effect of weather is model- and center-specific, and that no universal weather dividend emerged.

Equally consequential is the finding that models do not travel. When a model tuned at one hospital was deployed at the other, errors ballooned severalfold, with mean absolute percentage errors climbing to roughly 39 percent. Local attendance culture, catchment demographics, hospital-specific referral patterns, and idiosyncratic scheduling apparently imprint themselves so deeply on each site’s demand curve that a model calibrated elsewhere is nearly useless. For health systems hoping to buy a single, generic forecasting product, this is a cautionary datapoint: locally tuned systems outperform imported ones.

There is an important caveat the authors themselves flag. In the holdout analyses where weather appeared to help, the models were fed observed future meteorological values as exogenous inputs. In a real operational setting, tomorrow’s weather is not observed but forecast, and forecast error would erode any weather advantage. The weather-augmented results therefore represent an idealized upper bound on the incremental value of meteorological information, not a fully prospective simulation of a live deployment.

The broader scientific context makes the finding less paradoxical than it first appears. Epidemiological literature has long documented links between ambient temperature and morbidity, seasonal patterns in respiratory viral infections, and correlations between weather and trauma admissions. Weather does influence health. But those influences are largely entangled with the seasons, and the calendar already encodes the seasons. Day-of-week effects, holidays, paydays, and seasonal disease waves explain most of the predictable variance in emergency demand, leaving little residual signal for humidity or wind speed to capture once the calendar is modeled properly.

For hospital operations, the practical message is clear and refreshingly frugal. Daily emergency department attendance can indeed be forecast with an accuracy useful for staffing and resource planning, with errors around four to six percent at both Turkish centers. But the winning ingredients are not exotic weather feeds or elaborate machine-learning pipelines; they are each hospital’s own attendance history and the disciplined use of calendar structure. The most valuable weather forecast for an emergency department, it turns out, may simply be the knowledge that the calendar rarely lies.

Subject of Research: Time-series forecasting of daily emergency department visit volume using calendar and meteorological variables

Article Title: Forecasting emergency department visit volume using meteorological variables: a multicenter time-series study

Article References: Sarbay, İ., Coban, B., Efgan, M. G., Kanter, E., Özel, Z., Kaymaz, E., Bıyıklı, S., Çalık, M., & Tekindal, M. A. (2026). Forecasting emergency department visit volume using meteorological variables: a multicenter time-series study. Air Quality, Atmosphere & Health, 19(9), Article 205. https://doi.org/10.1007/s11869-026-02097-z

Image Credits: AI Generated

DOI: 10.1007/s11869-026-02097-z

Keywords: emergency department, forecasting, time series, meteorological variables, machine learning, SARIMAX, XGBoost, dynamic regression, calendar effects, hospital operations, Türkiye, emergency