Forecasts

Evaluations

Background

Community

Get in touch

Teams from anywhere in the world are invited to submit forecasts once a week for one or more of the countries. Take a look at the submission instructions and get in touch with any questions.

Here is the list of forecast models grouped by team:

European Centre for Disease Prevention and Control

ECDC

ARI Models

ARIMA model with seasonality

Auto ETS model.

Historical simplex pattern

Using historical data patterns with highest similarity to current data to foreast the future values
##### Contributors

##### Detailed description

Model consists of 1) taking 'm' latest data points, 2) find 'n' closest neighbours from all historical data using L2-norm, 3) use the next data time point from each historical data points, 4) use data from point 3 to fit a log-normal distribution , 5) use the log-normal distirbution to estimate the quantiles/distribution, 6) return to step 3 but use two data points ahead for estimations of horizon 2, etc for horizon 3 and 4, 7) find optimal values of 'n' and 'm for a given country, using past 4 weeks of forecasts.
##### Data inputs

Historical ILIARIRates data as provided by RespiCast and ECDC

Norrsken Piecewise Linear Bayesian Model

A Bayesian piece-wise square-root-linear model fit.
##### Contributors

##### Detailed description

Model consists of 1) scaling of the data (convert to square-root-scale), 2) assigning data points into groups such that neighbouring data points belong to the same group, and simultaneously in Stan 3) fitting piece-wise intercept-slope models in each group, 4) regularizing the difference between neighbouring slopes to be close to 0, 5) generating a future slope based on the last slope plus fitted noise.
##### Data inputs

ECDC ERVISS

Norrsken Piecewise Linear Bayesian Model

A Bayesian piece-wise log-linear model fit.
##### Contributors

##### Detailed description

Model consists of 1) scaling of the data (convert to log-scale), 2) assigning data points into groups such that neighbouring data points belong to the same group, and simultaneously in Stan 3) fitting piece-wise intercept-slope models in each group, 4) regularizing the difference between neighbouring slopes to be close to 0, 5) generating a future slope based on the last slope plus fitted noise.
##### Data inputs

ECDC ERVISS

ILI Models

Historical simplex pattern

Using historical data patterns with highest similarity to current data to foreast the future values
##### Contributors

##### Detailed description

Model consists of 1) taking 'm' latest data points, 2) find 'n' closest neighbours from all historical data using L2-norm, 3) use the next data time point from each historical data points, 4) use data from point 3 to fit a log-normal distribution , 5) use the log-normal distirbution to estimate the quantiles/distribution, 6) return to step 3 but use two data points ahead for estimations of horizon 2, etc for horizon 3 and 4, 7) find optimal values of 'n' and 'm for a given country, using past 4 weeks of forecasts.
##### Data inputs

Historical ILIARIRates data as provided by RespiCast and ECDC

Norrsken Piecewise Linear Bayesian Model

A Bayesian piece-wise log-linear model fit.
##### Contributors

##### Detailed description

Model consists of 1) scaling of the data (convert to log-scale), 2) assigning data points into groups such that neighbouring data points belong to the same group, and simultaneously in Stan 3) fitting piece-wise intercept-slope models in each group, 4) regularizing the difference between neighbouring slopes to be close to 0, 5) generating a future slope based on the last slope plus fitted noise.
##### Data inputs

ECDC ERVISS

Norrsken Piecewise Linear Bayesian Model

A Bayesian piece-wise square-root-linear model fit.
##### Contributors

##### Detailed description

Model consists of 1) scaling of the data (convert to square-root-scale), 2) assigning data points into groups such that neighbouring data points belong to the same group, and simultaneously in Stan 3) fitting piece-wise intercept-slope models in each group, 4) regularizing the difference between neighbouring slopes to be close to 0, 5) generating a future slope based on the last slope plus fitted noise.
##### Data inputs

ECDC ERVISS

Fjordhest

fjordhest

ARI Models

fjordhest-ensemble

An inverse-WIS weighted ensemble of 3 component models - an exponential trend smoothing (ETS) model, a quantile AR model, and a baseline model of random walk with drift.
##### Contributors

##### Detailed description

Quantile autoregression model uses quantgen R package, and is similar to CMU-Timeseries model from 2022/23 FluSight season. Random walk and ETS models use fable R package. To build an ensemble, the quantile distributions of the component models are weighted (location- and target-specific) by the mean of inverse-WIS scores over the last 3 weeks that could be evaluated. The estimates are solely the responsibility of the contributors and do not represent, nor are endorsed by, the Norwegian Institute of Public Health.
##### Citation/s

Tibshirani R, Brooks L (2020). quantgen: Tools for generalized quantile modeling; https://github.com/cmu-delphi/flu-hosp-forecast/; O'Hara-Wild M, Hyndman R, Wang E (2022). fable: Forecasting Models for Tidy Time Series. R package version 0.3.2.
##### Data inputs

ECDC ERVISS
##### License

CC-BY_SA-4.0

ILI Models

fjordhest-ensemble

An inverse-WIS weighted ensemble of 3 component models - an exponential trend smoothing (ETS) model, a quantile AR model, and a baseline model of random walk with drift.
##### Contributors

##### Detailed description

Quantile autoregression model uses quantgen R package, and is similar to CMU-Timeseries model from 2022/23 FluSight season. Random walk and ETS models use fable R package. To build an ensemble, the quantile distributions of the component models are weighted (location- and target-specific) by the mean of inverse-WIS scores over the last 3 weeks that could be evaluated. The estimates are solely the responsibility of the contributors and do not represent, nor are endorsed by, the Norwegian Institute of Public Health.
##### Citation/s

Tibshirani R, Brooks L (2020). quantgen: Tools for generalized quantile modeling; https://github.com/cmu-delphi/flu-hosp-forecast/; O'Hara-Wild M, Hyndman R, Wang E (2022). fable: Forecasting Models for Tidy Time Series. R package version 0.3.2.
##### Data inputs

ECDC ERVISS
##### License

CC-BY_SA-4.0

ISI Foundation

ISI

ARI Models

AriABCaster

EpiNowARI

Generalized Logistic Model for ARI

ILI Models

EpiNow

FluABCaster

GLEAM

Generalized Logistic Model

IPSICast

An exogenous autoregressive model that integrates information on digital surveillance data from the Influweb participatory system.

ItaLuxColab

ItaLuxColab

ILI Models

Epidemic projections by SIRS+EKF

Stochastic SIRS model with automatic data integration by the extended Kalman filter (EKF) with adaptive hyperparameter estimation, applied separately on each region (github.com/AtteAalto/EpiEKF)

PoliTo Complex System Lab

CSL_PoliTo

ILI Models

Metapopulation SEINR model

An epidemic model based on the susceptible-exposed-infectious-noninfectious-removed dynamic, with meta-population and activity-driven modelling, class subdivision, Bayesian parameter optimization

QMUL

QMUL

RespiCast

respicast

ARI Models

Hub Ensemble

Quantile Baseline

A simple baseline model that samples future target changes from historical increments.
##### Contributors

##### Detailed description

The baseline model consistently predicts as median value the last data point within the calibration period, while its confidence intervals are estimated on past data increments. Similar baseline models have become a standard neutral benchmark providing a simple reference for all models in the context of collaborative forecasting hubs, see Cramer et al. "Evaluation of individual and ensemble probabilistic forecasts of COVID-19 mortality in the United States", 2022 and Sherratt et al. "Predictive performance of multi-model ensemble forecasts of COVID-19 across European nations", 2023.
##### License

CC-BY-4.0

ILI Models

Hub Ensemble

Quantile Baseline

A simple baseline model that samples future target changes from historical increments.
##### Contributors

##### Detailed description

The baseline model consistently predicts as median value the last data point within the calibration period, while its confidence intervals are estimated on past data increments. Similar baseline models have become a standard neutral benchmark providing a simple reference for all models in the context of collaborative forecasting hubs, see Cramer et al. "Evaluation of individual and ensemble probabilistic forecasts of COVID-19 mortality in the United States", 2022 and Sherratt et al. "Predictive performance of multi-model ensemble forecasts of COVID-19 across European nations", 2023.
##### License

CC-BY-4.0