Leveraging probabilistic forecasts for dengue preparedness and control: The 2024 Dengue Forecasting Sprint in Brazil

Significance

The Infodengue-Mosqlimate Dengue Challenge 2024 (IMDC24) was organized by the Mosqlimate-Infodengue consortium, which aims to provide forecasting models as decision support tools for early warning systems, scenario assessments, and an empirical basis for resource allocation for mosquito-borne diseases. During IMDC24, six international teams, provided with dengue case, sociodemographic, and climate data, developed scenario forecasting models for the 2024 and 2025 dengue seasons in Brazil. In this study, we evaluated the performance of each model and built an ensemble model, considering the variation in performance of each model, especially during the 2024’s unprecedentedly intense season. Among the main applications of this work, we highlight the incorporation of the results into the Brazilian Ministry of Health’s nationwide dengue epidemic response agenda.

Abstract

Forecast models are a key decision-support tool for public health authorities in managing epidemics, feeding into early warning systems, scenario evaluations, and an empirical basis for resource allocation. In Brazil, improving dengue forecasting became a priority in response to the unprecedented increase in cases, which surpassed the total of the previous decade and expanded to new regions. The Infodengue-Mosqlimate consortium launched the Infodengue-Mosqlimate Dengue Challenge 2024 (IMDC24), or Dengue Forecast Sprint, bringing together six international teams provided with cases and climate covariates data to generate actionable forecasts for 2024 and 2025 seasons in five diverse Brazilian states, leveraging advanced machine learning and classical statistical models. This paper outlines the structure and findings of the IMDC24. The performance of the models varied between years and locations, and no single model consistently excelled, especially during 2024’s unprecedentedly large season. This performance variability highlighted the need for ensemble approaches. The ensemble models developed are presented as the main results of this collaborative development. As intended, the ensemble models have been adopted by Brazilian public health authorities to help with planning and response to the forecasted 2025 dengue epidemics across the country.

Forecast models are a key asset in the response to epidemics by public health authorities, informing their control strategies (1–3). Output from these models can provide early warning of emerging outbreaks, help with the evaluation of scenarios, and enhance the precision of resource allocation. Developing an appropriate forecasting model depends on specific research questions and public health objectives. Each model has unique strengths and weaknesses, and a single model is unlikely to encapsulate all relevant aspects of disease dynamics. For instance, while deep learning models are adept at capturing temporal structures and nonlinearities, classical statistical models like SARIMA can perform comparably well, especially in systems with pronounced seasonal patterns such as those observed in dengue. This underscores the need for rigorous benchmarking and model comparison using appropriate evaluation metrics across diverse modeling domains (4).

Numerous epidemic forecast initiatives have been proposed (3, 5–15), including probabilistic superensemble models that integrate seasonal climate forecasts and lagged disease cases (16) and ensemble machine learning methods that include advanced deep neural networks (17–19).

Combining predictions from various models in ensembles capture the main features of each model while guarding against overconfidence or catastrophic failure (2, 6, 7, 20–23). For instance, Johansson et al. (6) and Wu et al. (23) found that ensembles of dengue forecasting models tended to be better calibrated and have better point-prediction performance than individual models.

Dengue fever, a mosquito-borne viral disease, is a major public health challenge in tropical and subtropical regions worldwide. Previous studies have associated risk of dengue outbreaks with levels of urbanization, connectivity, and temperature suitability (24). Brazil, in particular, has witnessed a dramatic increase in dengue incidence in recent years, with the disease spreading to new geographical areas (25). In 2024, the number of cases surpassed the sum of the preceding decade (26). The geographical expansion of dengue to new areas in southern Brazil, combined with the high transmission in many cities, put a strain on public health systems, leading to increased lethality and hospitalization, especially among the elderly (27).

In anticipation of the 2025 dengue season, in June 2024, the Infodengue-Mosqlimate consortium launched a dengue forecast sprint called Infodengue-Mosqlimate Dengue Challenge 2024 (IMDC24) (summarized in Fig. 1). Infodengue is a dengue early warning system that has been active in Brazil since 2015, producing weekly reports for all municipalities (https://info.dengue.mat.br), while Mosqlimate (https://mosqlimate.org) is an open-source web-based platform designed to support the generation, comparison, and sharing of forecast models for dengue and other mosquito-borne diseases.

Fig. 1.

IMDC24 design and workflow. Harmonized surveillance, sociodemographic, and climate covariates are served via the Mosqlimate platform/API to six international teams, who generate weekly probabilistic forecasts for the 2024–2025 dengue seasons in five Brazilian states. Submissions (quantiles) are standardized via parametric approximations and evaluated with proper scoring rules (CRPS, log score, interval score). Two ensembles are built: $E_1$ (uniform linear mixture) and $E_2$ (log–linear/“logarithmic” pooling) with weights estimated by CRPS in 2023. See text for details. Outputs inform preseason planning at Brazil’s Ministry of Health. (Boxes denote components; arrows indicate data/forecast flow).

The purpose of the Sprint was to generate forecasts that could provide actionable information to the Brazilian Ministry of Health (MoH), guiding resource allocation for effective prevention and control measures. The typical dengue season in Brazil runs from October to May, and the Ministry of Health needed to know the expected magnitude of the following season, using the information available up to June. In addition, results needed to be delivered in a timely manner, before the season starts, to facilitate evidence-based decision-making and strategic planning for the next epidemic season. Thus, to ensure alignment with the objectives of the Brazilian health policy, the forecasting targets were collaboratively established with the Ministry of Health.

This pioneering initiative in Brazil aimed to bridge the gap between research and real-world application, drawing inspiration from previous endeavors such as the 2015 CDC Dengue Forecasting Challenge (6). To this end, we assembled a comprehensive public dataset that included case data, climatic, and demographic variables, with a custom API for easy data retrieval (28).

A total of six teams from four countries participated in the IMDC24. At the calibration phase, they were invited to produce forecasts for two previous seasons (2023 and 2024), for five federative units (FU)(states), one for each Brazilian macroregion. Thus, a total of 10 predictions were submitted by each team, which were scored using different methods (described in Data). All teams were invited to produce forecasts for 2025 (true forecasts).

With all the forecasts and scores at hand, we can then produce ensembles that can, in turn, be used for decision-making. One of our aims was to produce ensembles that are easier for decision-makers to interpret by guaranteeing unimodality. To this end, we propose a methodology which combines parametric approximations of the probabilistic forecasts—in the form of log-normal distributions—with log–linear (logarithmic) pooling (29, 30). Details are provided in SI Appendix. We use forecasts from all groups (see below) to build ensemble forecast models to predict the 2025 season.

Materials and Methods

The IMDC24 lasted approximately three months. The announcement was released on 20 June 2024, with a forecast submission deadline of 16 August 2024. The tight schedule was necessary to accommodate the needs of the decision-makers. A document describing the rationale, challenges, links to the data and rules was provided to all participants (31). To participate, each team was required to fill out an online form, read and accept the instructions and rules and agree to use the training and validation as indicated. Then, a link to the sprint environment was provided.

The five Brazilian states included in the IMDC24 have unique climatic and socioeconomic characteristics, urbanization levels, and dengue history (SI Appendix, Fig. S1). These states span a latitudinal gradient from the equatorial north in Amazonas (AM) to the subtropical south in Paraná (PR), with distinct seasonal temperature patterns. Additionally, the longitudinal range, from Goiás (GO) in the west to Minas Gerais (MG) and Ceará (CE), in the east, is marked by widely varying humidity levels. All selected states have undergone multiple dengue outbreaks in the last 20 y, with the 2024 season being the most severe in Minas Gerais, Paraná, and Goiás (SI Appendix, Fig. S1). SI Appendix, Fig. S1 delineates the two primary forecasting exercises proposed to the teams: The first challenge was to train models using data up to 2022 (white background) and test their forecasts against 2023 data (shaded blue), while the second challenge involved training models up to 2023 and testing them using data from the shaded green period (2024).

Data.

The IMDC24 dataset and its documentation were made available through the Mosqlimate platform (28) and the interested reader is referred to Ganem et al. (32), which describes how the mosqlimate platform works. This comprehensive dataset included administrative, demographic, epidemiological, and climate data for all cities included in the challenge and additionally for all 27 Brazilian states and 5,570 municipalities.

Administrative and demographic data.

Within each state, municipalities are organized into health districts. The names and codes of both administrative divisions were provided for each municipality. The estimated population size per municipality was obtained from the Brazilian Institute of Geography and Statistics–IBGE.

Epidemiological data and case definition.

Case notification data of probable dengue cases at the date of symptom onset were made available aggregated by epidemiological week and municipality. A probable dengue case is confirmed through laboratory testing or clinical-epidemiological criteria, or is still under investigation. The Brazilian Ministry of Health uses this case definition for arboviruses surveillance. The dengue epidemiological year (here referred to as epiyear) is defined as the period spanning from the epidemiological week (EW) 41 of a calendar year to the epidemiological week 40 of the following year. We refer the reader to the published datasets, for the actual starting dates of these EWs, however, they typically start in early October.

The dataset comprises 14 seasons of case notification data between epidemiological years 2010–2024 in addition to some derived measures, such as the effective reproductive number ($R_t$) for dengue per epiweek and municipality, provided by Infodengue (33). A second group of descriptors came from the Episcanner tool (34). This tool fits Richard’s model to the dengue incidence data and estimates the basic reproduction number ($R_0$), the peak week, the total outbreak size, and the outbreak duration. These metrics were available only for municipalities and years that had reported dengue transmission, defined as a minimum of 3 wk with $R_t>1$, and at least 50 cases reported since 2010.

Climate data.

Meteorological variables were obtained from the ERA5 reanalysis product, encompassing temperature, pressure, humidity, and precipitation (35). Hourly data were aggregated to epidemiological week for each municipality and the average, maximum, and minimum statistics were computed. The diurnal temperature range was also calculated and summarized in the same way. Furthermore, monthly aggregated climate anomaly indices—the El Niñno Southern Oscillation (ENSO), the Indian Ocean Dipole (IOD), and the Pacific Decadal Oscillation (PDO)—were obtained from the National Oceanic and Atmospheric Administration (NOAA) (https://psl.noaa.gov/enso/mei/).

Other data.

A set of variables fixed in time were also provided for each municipality: the predominant Koppen climate type, the predominant biome, and the average altitude. These variables were obtained from the Brazilian Institute of Geography and Statistics–IBGE.

Participants were allowed to use additional data as long as those data were made available to all participants. A description of all data is available in SI Appendix, section 4, Sprint data and parameters description.

Challenges.

A series of forecasting challenges were presented to the teams. The first two exercises utilized historical data (see details below) and served as benchmarks for validation and comparative analysis of predictions. These are depicted as shaded regions in SI Appendix, Fig. S1. Then, teams were invited to provide forecasts for 2025 (i.e., forward-looking forecasts). For each challenge, the teams were instructed to provide the predictions for each of the selected states, compute point estimates and the (90%) uncertainty intervals, and upload them to the Mosqlimate platform.

First challenge: predict the weekly number of dengue cases for each selected state in the 2023 epiyear (EW 41 2022–EW 40 2023), using data covering the period from EW 1 2010 to EW 25 2022. Overall, the observed dengue incidence during this epiyear was not as high as in 2024, thus more comparable with historical levels (AM–percentile 53, CE–percentile 33, GO–percentile 40, MG–percentile 73, and PR–percentile 86).

Second challenge: predict the weekly number of dengue cases for each selected state in the 2024 epiyear (EW 41 2023–EW 40 2024), using data covering the period from EW 1 2010 to EW 25 2023. Dengue incidence in this epiyear was extremely high, posing a challenge for the predictive models. It included the largest case peaks observed historically in the states of MG (percentile 100), PR (percentile 100), and GO (percentile 100). In CE (percentile 13), the pattern was similar to that of 2023, while in AM (percentile 86), although incidence was higher than in 2023, comparable outbreaks had been recorded in previous years.

Third challenge: The teams were tasked with generating forecasts for the 2025 epiyear using the same models employed for the first two targets. The forecast window was set to be from EW 41 2024 to EW 40 2025.

Parametric Approximation and Scoring.

As previously mentioned, the predictions were presented as point estimates accompanied by uncertainty intervals, to aid tractability, we decided to approximate the predictive distribution as a log-normal distribution, this approximation provided a good representation of the median and the upper prediction interval—see SI Appendix, for a detailed description of the approximation procedure.

The models’ performances were evaluated using three scores: logarithmic score, continuous ranked probability score (CRPS), and interval score, available in the Python scoringrules package (36). The scores for each model were computed by state and EW, then averaged across epiyears and within a three-week window centered on the peak week, which is defined as the week with the maximum number of cases within the epiyear.

The formula for the CRPS for a log-normal distribution is given below (Eq. 1):

$$\begin{array}{c}\hfill \text{CRPS}(\text{LN}({\mu }_i,{\sigma }_i),y_i)=y_i\left[2\mathrm{\Phi }(y_i)-1\right]-\\ \hfill 2exp\left({\mu }_i+\frac{{\sigma }_i^2}{2}\right)\left[\mathrm{\Phi }({\omega }_i-{\sigma }_i)+\mathrm{\Phi }\left(\frac{{\sigma }_i}{\sqrt{2}}\right)\right],\end{array}$$ [1]

where $\mathrm{\Phi }$ is the cumulative distribution function (CDF) of the standard normal distribution and ${\omega }_i=\frac{\text{log}{y}_i-{\mu }_i}{{\sigma }_i}$, in which $y_i$ refers to the cases observed in week $i$, ${\mu }_i$ is the forecast for week $i$ and ${\sigma }_i$ is the SD of the forecast in week $i$. The CRPS can be interpreted as a generalization of the absolute error for probabilistic forecasts (37).

The logarithmic score (log score) is given by

$$\begin{array}{cc}\hfill \text{LogS}(\text{LN}({\mu }_i,{\sigma }_i),y_i)& =log{y}_i+log{\sigma }_i+\frac{1}{2}log(2\pi )\hfill \\ \hfill & +\frac{{(log{y}_i-{\mu }_i)}^2}{2{\sigma }_i^2}.\hfill \end{array}$$ [2]

The (log score) penalizes forecasts that assign a low probability density to the observed outcome. In particular, it tends to penalize light-tailed distributions more heavily when observations fall in the tails, even if the central tendency of the forecast is close to the observed value. This is because lighter tails assign lower probability to extreme values, resulting in a lower score when the true value lies far from the mode (37).

The interval score is computed directly from the prediction intervals, without requiring any parametric approximation. This metric is calculated using Eq. 3:

$$\begin{array}{c}\hfill S_{\alpha }^{\mathit{int}}(l_i,u_i;y_i)=u_i-l_i+\frac{2}{\alpha }(l_i-y_i)I\{y_i<l_i\}\\ \hfill +\frac{2}{\alpha }(y_i-u_i)I\{y_i>u_i\},\end{array}$$ [3]

where $I$ is the indicator function, $\alpha$ is the (significance) level of the interval, $u_i$ the upper value of the interval at week $i$, and $l_i$ the lower value. The interval score penalizes forecasts whose observed values fall outside the predicted interval, with the penalty increasing proportionally to the distance between the observation and the nearest bound of the interval. Furthermore, when comparing two prediction intervals that both contain the observed value, the interval score assigns a higher penalty to the wider interval, thus favoring sharper predictions.

Ensemble Models Construction.

Combining predictive models into ensembles is a well-established technique to improve forecast accuracy (7, 20, 38). To construct an ensemble from all the models submitted, we adopted a methodology based on the logarithmic pooling of the predictive probability distributions for each model (30) (see SI Appendix, section 2 for details). Two ensemble models ($E_1$ and $E_2$) were constructed: The first ($E_1$) used equal weights and a linear mixture of predictive distributions as a baseline, while the second ($E_2$) had its weights optimized to minimize the CRPS score of a logarithmic pooling of the ensemble predictions (SI Appendix, Fig. S6).

To assess the performance of the ensemble $e$ models in comparison with each individual model $m$, we computed a skill score based on their CRPS score, defined as

$$\begin{array}{c}\hfill {\text{SS}}_{e,m}=1-\frac{{\text{CRPS}}_{\text{e}}}{{\text{CRPS}}_{\text{m}}}.\end{array}$$ [4]

Results

Six teams submitted probabilistic forecasts for the five Brazilian states. One of them submitted two models, totaling seven models ($M_1$-$M_7$) (Table. 1). The specific structures of each model and how they used the available datasets are described in SI Appendix, section 1. Model types ranged from classical regression models to deep neural networks, with very diverse approaches.

Table 1.

Submitted forecast models

Model	ID	Description	Spatial scale	Uses climate data
Dobby Data (DD)	M 1	LSTM model	Health district	Yes
Global Health Resilience (GHR)	M 2	Bayesian spatiotemporal model	Health district	Yes
GeoHealth (GH)	M 3	LSTM and prophet models	State	Yes
Ki-Dengu Peppa (KDP)	M4, M5	Time series decomposition models	State	No
BB-M (BBM)	M 6	Bayesian baseline model	Health district	No
DS_OKSTATE (OK)	M 7	CNN-LSTM model	State	No

The spatial scale column indicates the smallest spatial unit of data aggregation used during model training.

Individual Models Performance.

Overall, forecast accuracy and precision varied considerably among models for the 2023 and 2024 epiyears. Fig. 2 show the forecasts for AM (a northern state with fewer cases and more endemic-like behavior), CE (a northeastern state that has exhibited seasonal peaks annually since 2010), GO (a Midwestern state showing a time series pattern comparable to that of CE), MG (a southeastern state that experienced larger seasonal peaks in the past), and PR (a southern state with increased outbreak intensity observed since 2019). Fig. 3A displays the CRPS values for each state and model in 2023 and 2024. Interval scores are provided in SI Appendix, Fig. S7, and logarithmic scores in SI Appendix, Fig. S8. In Amazonas (AM), where the 2023 dengue epiyear was below average, two models overestimated the epidemic curve ($M_4$ and $M_7$), while the remaining models aligned well with the observed data. Among them, $M_6$ and $M_1$ were the most accurate based on the CRPS and the interval scores (Fig. 3A and SI Appendix, Fig. S7). Based on the logarithmic score, the best models are $M_6$ and $M_2$ (SI Appendix, Fig. S8). In terms of the CRPS computed around the peak, the top models were $M_3$ and $M_6$ (SI Appendix, Fig. S9). In 2024, during a period of above-average seasonal incidence, models $M_5$ and $M_6$ exhibited the best performance based on the CRPS score, while models $M_6$, $M_3$, and $M_7$ were optimal according to the interval score. As observed in 2023, models $M_6$ and $M_2$ continue to outperform others based on the logarithmic score. Also, in terms of the CRPS computed around the peak, the top models in 2024 were $M_1$ and $M_7$.

Fig. 2.

Forecasted (median and 90% uncertainty intervals) versus observed probable dengue cases by week for the Brazilian states selected for the challenge [Amazonas–AM, Ceará–CE, Goiás–GO, Minas Gerais–MG, and Paraná (PR)] and epiyears 2023 and 2024 produced by individual models ($M_1$ to $M_7$) (first column) and ensemble methodologies ($E_1$ and $E_2$) (second column). The panels in each row share the same y-axis scale.

Fig. 3.

Forecasting models performance based on the continuous ranked probability score (CRPS) and the Skill Score (SS). (A) CRPS values for the individual models by state for epiyears 2023 and 2024, with lower values indicating better performance. The CRPS values represent the average score over the entire prediction window. (B) SS comparing the ensemble model $E_2$ with the individual models and the baseline ensemble model $E_1$, by state. For the year 2023, the predictions from $E_2$ are in-sample, as its weights were computed using both the 2023 predictions and the corresponding observed dengue data. In contrast, the predictions for 2024 are out-of-sample, as the same weights derived from the 2023 data were applied directly, without further adjustment. Positive SS values are interpreted as superior predictive skill of the ensemble model $E_2$.

In Minas Gerais (MG), both years were characterized by above-average dengue activity (Fig. 2). In 2023, $M_4$ had the best CRPS scores (Fig. 3A), but not the best interval and logarithmic scores (SI Appendix, Figs. S7 and S8, model $M_6$). Moreover, $M_4$ was among the top-performing models around the peak, alongside $M_1$ (SI Appendix, Fig. S9). In 2024, none of the models predicted the earlier onset of the epidemic curve, and overall, the models tended to underestimate the cases. However, $M_5$ performed best according to the CRPS both for the entire season and around the peak, while $M_3$ and $M_2$ achieved the highest accuracy based on the interval score and logarithmic score, respectively.

Ceará (CE) experienced below-average dengue epiyears in 2023 and 2024 (Fig. 2). In 2023, all models tended to overestimate the epidemic curve, while in 2024, $M_5$ and $M_1$ aligned well with the data, with considerably better CRPS scores both for the entire season and around the peak, as well as more accurate interval scores compared to the other models (Fig. 3 and SI Appendix, Figs. S7 and S9). According to the logarithmic score, models $M_6$ and $M_2$ demonstrated the best overall performance (SI Appendix, Fig. S8).

For Goiás (GO), all models tended to overestimate the epidemic curve in 2023, a below-average epiyear (Fig. 2). The estimates were consistent across models, except $M_5$, which exhibited poorer performance. The 2024 epiyear in this state was exceptionally high and no model successfully captured it. Only two models ($M_2$ and $M_3$) generated uncertainty intervals that encompassed the observed level.

In Paraná (PR), the 2023 epiyear was moderate while 2024 was the highest epiyear in the study period. While the best models were $M_4$ and $M_1$, in 2023, according to the CRPS scores, they greatly underestimated the peak (Figs. 2 and 3A). Regarding the peak specifically, the top-performing models were $M_2$ and $M_5$ in both 2023 and 2024 epiyears (SI Appendix, Fig. S9). Meanwhile, $M_4$ and $M_6$ were the best according to the interval score (SI Appendix, Fig. S7), while $M_6$ and $M_2$ were top-ranked according to the logarithmic score. For 2024, all models underestimated the severity of the 2024 epiyear, with $M_5$ and $M_2$ as the best models according to the CRPS score (Fig. 3), while $M_3$ and $M_2$ were the top performers on the other two score methods (SI Appendix, Figs. S7 and S8).

We report the weekly values of the scores in SI Appendix, section 3.1 and Figs. S10–S12. In general, scores were worse when the incidence was higher. According to the CRPS score (Fig. 3A), the best-performing models in 2023 were $M_6$ and $M_4$, while in 2024, $M_2$ and $M_5$ demonstrated the highest performance.

The Ensemble Models.

The second column of Fig. 2 shows the forecasts produced by the two ensemble models, considering an implementation with equal weights ($E_1$) and the logarithmic pool with CRPS informed weights ($E_2$), for the mandatory states. Fig. 3B shows the results of the skill score of the $E_2$ model compared to $E_1$ and the individual models ($M_1$ to $M_7$). The ensemble model $E_2$ outperformed $E_1$ and most individual models in 2023 (in-sample $E_2$ forecasts, as the weights were computed using both the predictions and the observed dengue data from 2023), except $M_6$ for GO. For the out-of-sample forecasts, the results were mixed. The $E_2$ model outperformed most models for AM, CE, and PR. However, for GO and MG, the $E_2$ ensemble model did not show a clear improvement over individual models when predicting the highly atypical seasons of 2024 (out-of-sample $E_2$ forecasts, as the prediction of the 2024 season relied on weights derived from both the 2023 predictions and observed dengue data in 2023).

A notable contribution of the ensemble models is the reduction of the forecast uncertainty. This is clear for all target states and, in general, is more pronounced when applying the $E_2$ methodology.

Forecasting the 2025 Dengue Epiyear.

Fig. 4A shows, for each of the five states, three forecast curves for the 2025 epiyear. The curves represent ensemble forecasts generated using the $E_1$ methodology, which applies linear pooling to the individual predictions, and the $E_2$ methodology, which employs logarithmic pooling. The ensemble $E_1$, shown in yellow, employs linear pooling with equal weights assigned to each individual prediction. Conversely, the ensemble $E_2(2023)$, shown in green, uses the pooling weights to minimize the CRPS of the models in the 2023 epiyear, while the ensemble $E_2(2024)$, in blue, uses pooling weights to minimize the CRPS of the models in the 2024 epiyear. The black dashed line represents the number of probable cases up to epidemiological EW 25 2025, reported up to EW 31 2025. For all states except PR, the 2024-trained ensemble produced higher epidemic curves than the 2023-trained ensemble, possibly capturing the expectation of a worse season after a sequence of increasing waves. The time series forecasts produced by the individual models and ensembles for the 2023, 2024, and 2025 epiyears, together with the historical data since 2010, are shown in SI Appendix, Fig. S2. SI Appendix, Fig. S3 presents the forecasts of individual models and ensembles exclusively for the 2025 epiyear.

Fig. 4.

Performance of the ensemble models for the 2025 season. (A) Forecasting of weekly probable dengue cases (median and 90% uncertainty interval) for the 2025 epiyear across the five Brazilian states included in the IMDC24. The green line ($E_2(2023)$) corresponds to the forecast generated by the ensemble model with weights calibrated using the 2023 predictions, while the blue line ($E_2(2024)$) corresponds to the forecast generated by the same model using weights calibrated on 2024 predictions. The yellow line represents the baseline ensemble generated using the $E_1$ methodology. The dashed black line represents the observed data for the 2025 season. (B) CRPS values for the individual models and ensembles for 2025, with lower values indicating better performance. In the figure, $E_1$ represents the baseline ensemble model, constructed as a linear log-normal mixture with equal weights assigned to each component model. (C) SS comparing the ensemble model $E_2(2023)$ with the individual models and the ensemble model $E_2(2024)$ and $E_1$. Positive SS values are interpreted as superior predictive skill of the ensemble model $E_2(2023)$.

Based on the CRPS values presented in Fig. 4B, the ensemble $E_2(2023)$ outperformed the individual models in the states of AM and CE. In contrast, model $M_6$ yielded the best performance in GO, MG, and PR. Furthermore, as shown in Fig. 4C, the performance of $E_2(2023)$ was inferior to that of models $M_6$ and $M_7$ in GO, and to models $M_6$, $M_7$, $E_1$ (the baseline ensemble), and $M_1$, $M_4$, and $M_7$ in other states. This weaker performance is consistent with the weights assigned to individual models in $E_2(2023)$ and $E_2(2024)$, as shown in SI Appendix, Fig. S4. In the states of GO, MG, and PR, models $M_2$, $M_4$, and $M_5$ received the highest weights, despite being among the worst-performing models for the 2025 season based on CRPS values (Fig. 4B). This misalignment can be explained by the fact that these models performed well in 2023 and 2024 (Fig. 3A), which influenced their selection in the ensemble. However, they significantly overestimated the magnitude of the 2025 outbreak, as illustrated in SI Appendix, Fig. S2.

Discussion and Directions for Future Research

In alignment with major international health organizations, Brazil has been investing in preventive measures to reduce arboviral diseases and mitigate the health impacts of the climate crisis through coordination among health authorities and partner sectors (education, environment, civil defense, social assistance, and others), civil society, and federal, state, and municipal governments. As part of this effort and anticipating the 2025 seasonal peak of arboviral infections, the Brazilian Ministry of Health launched in mid 2024 the Action Plan for Dengue and Other Arboviral Disease Reduction, aimed at guiding the phased implementation of new vector control technologies, revising protocols for monitoring key surveillance and healthcare indicators for early detection, and strengthening the integrated response capacity to reduce hospitalizations and preventable deaths. The IMDC24 was designed to forecast the most likely 2025 epidemiological scenarios using all available information and a suite of predictive models, to contribute to this effort.

This paper describes the IMDC24 initiative, the first to be implemented in Brazil, presenting lessons learned and directions for future endeavors. First of all, IMDC24 demonstrates that organizing a multimodel forecast sprint aligned with decision-making requirements, is feasible and can deliver timely results, provided certain conditions are met: access to comprehensive and harmonized data, robust infrastructure for model submission and evaluation, and a well-established community of modelers. This was possible due to a long term collaboration between the research community and the Brazilian MoH through the Infodengue early warning system. In operation since 2015, Infodengue harmonizes climate and disease data, estimates incidence using nowcast models, and delivers weekly reports for all Brazilian municipalities. This collaboration provided the background epidemiological data, and knowledge for deploying the Sprint at short notice. The Mosqlimate platform supplemented the Infodengue dataset with comprehensive climate data, and the model comparison and scoring tools to run the Sprint.

To maximize the usefulness of the IMDC24 initiative, the predicted scenarios were published in September as a technical report (in Portuguese) addressed to the Brazilian Ministry of Health, ensuring it reached key decision-makers in the country on time for the implementation of preventive measures (39). Additionally, several webinars were organized with key stakeholders during the Sprint and afterward to elucidate and discuss the methods and results. For example, one of the webinars, organized by the Global Health Network LAC, reached online attendance of almost a thousand people, including a large number of public health professionals from different Brazilian cities (40).

In the calibration phase of IMDC24, when models provided predictions for the 2023 and 2024 seasons, no single model consistently excelled across all forecast targets. While most models performed well for the 2023’s relatively typical seasons, neither individual models nor ensemble models provided accurate results for the extreme epidemic observed during the 2024 epiyear. This season was atypically severe in Brazil, and more generally, in Latin America, a phenomenon that has been partially attributed to climate change, which has altered environmental conditions, and expanding the regions with dengue transmission suitability (41).

Another potential factor that may have hampered models in 2024 is related to the contamination of dengue notification data with cases of other diseases with similar symptoms. This year observed a substantial increase and geographical expansion of Oropouche fever, which presents symptoms similar to dengue (42) and could have been reported as dengue cases. Similar problems have been reported in the past, like the misclassification of Zika during epidemics of chikungunya and dengue (43). Addressing these uncertainties requires better laboratory confirmation, although this poses a challenge in a country as large as Brazil.

Overall, models $M_1$ and $M_6$ were the top performers across states in the 2023 epiyear, while in the 2024 epiyear, models $M_2$ and $M_5$ tended to rank higher according to the CRPS score, as shown in Fig. 3A. In 2025 epiyear, as in the 2023 epiyear, the models $M_6$ and $M_1$ generally achieved higher rankings according to CRPS, as shown in Fig. 4B. Notably, $M_2$, $M_5$, and $M_6$ are Bayesian inference models, underscoring the advantages of this methodology for generating probabilistic forecasts compared to machine learning and deep learning models.

Within the selected models, model $M_6$—which relies solely on historical data—appears to be a suitable candidate for defining expected dengue seasons based on activity thresholds, as proposed in ref. 44. This is supported by the fact that $M_6$ achieved the best performance according to the CRPS in 2025 for the states of AM, GO, PR, and MG. However, $M_6$ lagged behind other models when forecasting MG and PR in 2023 and 2024. In those years, model $M_4$ was better in 2023 and model $M_5$ in 2024. Also, in 2024, $M_6$ did not rank as the top-performing model in any state. These results suggest that models $M_2$, $M_4$, and $M_5$ are able to capture patterns that deviate from typical seasonal trends. Furthermore, as illustrated in SI Appendix, Fig. S1, 2023 in MG and PR was characterized by high activity levels but with some precedents in the historical data, while 2024 presented atypical patterns with no similar historical events, potentially contributing to the lower predictive performance observed in that year.

Among the submitted models, $M_2$ and $M_3$ used meteorological variables directly; $M_1$ included a global climate index (ENSO). These models did not consistently achieve higher CRPS rankings (Figs. 3A and 4B), suggesting that the inclusion of meteorological variables did not consistently improve model performance. According to the dengue modeling literature, the contribution of climate variables to dengue incidence varies spatially (45). Models without climate variables seem more parsimonious in certain locations and epiyears for the long prediction window target proposed by IMDC24.

This study employed three different scoring methods to evaluate models, each with its own characteristics. Our results indicate that using multiple scores is important to measure different aspects of the fit. Scoring remains an important avenue for future development, in particular with regard to the development of scores that are fine-tuned to measure key aspects in public health decision-making (46).

The performance of the IMDC24 ensemble models could have been enhanced by incorporating additional models which were not explored during this initiative. For example, some of the submitted models did not incorporate any assumptions related to the biological dynamics of disease transmission. This underscores the ongoing challenge of integrating domain-specific epidemiological knowledge into statistical and machine learning models.

In particular, ensemble $E_2$, our best ensemble, did perform worse than many individual models in GO and MG in 2024. In this epiyear, we used weights computed based on the 2023 predictions and observed data. These weights are presented in the Right panel of SI Appendix, Fig. S4. In the case of GO, the $E_2$ ensemble assigned weights exclusively to models $M_2$ and $M_1$, and in 2024, $M_1$ was among the worst performing models for that state according to CRPS. Similarly, for MG, the methodology selected only model $M_4$ for inclusion in the ensemble. As shown in Fig. 3B, this resulted in $E_2$ performing worse than models $M_1$, $M_2$, $M_3$, and $M_5$, as its CRPS score was equal to that of model $M_4$.

In addition, for the 2025 forecast, we propose the ensembles $E_2(2023)$ and $E_2(2024)$. The weights for the former were derived using the 2023 predictions and observed dengue data, while those for the latter were based on the 2024 predictions and data. The ensemble weights assigned to each model are presented in SI Appendix, Fig. S4. In the states of GO, MG, and PR—where ensemble performance was poorer (as shown in Fig. 4)—the models $M_2$, $M_4$, and $M_5$ received the highest weights, based on their 2023 and 2024 performances, but did not perform well in 2025 as indicated by their (higher) CRPS values (Fig. 4B). Consequently, they significantly overestimated the magnitude of the 2025 outbreak, as illustrated in SI Appendix, Fig. S2, which presents forecasts for the 2023, 2024, and 2025 epiyears.

In addition to new models, improving the weight calculation from available models could be investigated in future endeavors. Due to constraints in the IMDC24 rules, weights for the 2024 ensembles were derived solely from data from the year 2023, while the 2025 ensemble forecasts weights were derived from data from both 2023 and 2024. Extending the number of epiyears used for weight calculation is recommended in future efforts.

Some of the choices made for the challenge posed constraints for the extrapolation of results to other settings. For instance, selecting states with a history of high dengue endemicity facilitated model development due to the availability of long time series and outbreak sequences for training. However, we did not assess model performance in areas with low endemicity or recent disease emergence—as seen, for instance, in Brazil’s southernmost states—thus neglecting the important epidemiological task of predicting potential outbreaks or an unprecedented increase in historically low areas.

It is also important to note that short-term forecasts play a very important and complementary role to the long-term projections we present here. As the work of Wu et al. (23) shows, ensembles have both retrospective and prospective performance for shorter, one to three months ahead predictions for dengue. We envision the use of these shorter term forecasts in conjunction with the longer-term projections we provide. The season-long predictions could be updated conditional on short term information and the shorter-term projections could condition on the longer forecasts. Efficient integration between short and long term forecasts is an important topic for future research.

Finally, the models presented here do not cover all possibly relevant sources of information for forecasting dengue. For example, they do not incorporate virological surveillance data due to its limited availability across all states.

Conclusion

The IMDC24 initiative underscores that planning for future epidemics necessitates multimodel approaches to account for epistemic and probabilistic uncertainties. The quality of models can be enhanced by robust surveillance systems, high data quality, and a deep understanding of causal links. We conclude that forecast scenarios are vital component of the public health toolkit, providing an additional piece of epidemiological intelligence alongside other decision-support tools. Models should be viewed as tools to explore scenarios for action, built on incomplete knowledge, reinforcing, as a recommendation, the importance of joint work, of engagement with health professionals who work in surveillance. As we finalize this article for publication, the second edition of IMDC, IMDC25 is ongoing, with a larger number of groups working to produce their best forecasts for 2026. This is evidence of the relevance of the Mosqlimate platform and the community of practice that grew around it to foster collaborative efforts to better understand the relationship between dengue dynamics and climate.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The data underlying this article are available for download from Zenodo (28) as well as on the Mosqlimate platform from where it can be freely downloaded through its API. The code used to generate the analyses presented here is provided as SI Appendix on the following GitHub repository: https://github.com/Mosqlimate-project/dengue_sprint_paper/tree/main.