Bisaloo
commented
3 years ago An simple option to start if we don't have anything better in the pipeline:
library(yaml)
library(purrr)
library(fs)
dir_ls("data-processed", recurse = TRUE, regexp = "/(.*)/metadata-\\1") %>%
purrr::map_dfr(read_yaml) %>%
knitr::kable(format = "markdown")| team_name | model_name | model_abbr | model_contributors | website_url | license | team_model_designation | methods | institution_affil | team_funding | twitter_handles | citation | methods_long | repo_url | data_inputs |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| BIOCOMSC | Gompertz | BIOCOMSC-Gompertz | Martí Català (Centre for Comparative Medicine and Bioimage (CMCiB), Germans Trias i Pujol Research Institute (IGTP), Barcelona, Spain) mcatala@igtp.cat, Enric Álvarez (Department of Physics, Universitat Politècnica de Catalunya, Barcelona, Spain) enric.alvarez@upc.edu, Sergio Alonso (Department of Physics, Universitat Politècnica de Catalunya, Barcelona, Spain) s.alonso@upc.edu, Daniel López (Department of Physics, Universitat Politècnica de Catalunya, Barcelona, Spain) daniel.lopez-codina@upc.edu, Clara Prats (Department of Physics, Universitat Politècnica de Catalunya, Barcelona, Spain) clara.prats@upc.edu | https://biocomsc.upc.edu/en/covid-19 | cc-by-4.0 | primary | Empirical model based on cases and deaths dynamics. | Universitat Politècnica de Catalunya and Institut de Recerca Germans Trias i Pujol. | European Commission, DG-CONNECT (CNECT/LUX/2020/LVP/0085, LC-01591965); Ministerio de Ciencia e Innovación, Gobierno de España (PGC2018-095456-B-I00) | BIOCOMSC1 | https://doi.org/10.1371/journal.pcbi.1008431 | NA | NA | NA |
| University of Cologne Covid Metrics | CovidMetrics-epiBATS | CovidMetrics-epiBATS | Tom Zimmermann (University of Cologne) tom.zimmermann@uni-koeln.de, Arne Rodloff (University of Cologne) arne.rodlloff@uni-koeln.de | https://tomz.shinyapps.io/coronaLandkreise/ | cc-by-4.0 | primary | Forecasts are based on TBATS - models (DeLivera, Hyndman und Snyder (2011)) and are updated daily for each German state. Final models are based on comparing multiple models and selecting the best-performing model based on AIC in historical data. | University of Cologne | NA | NA | https://tomz.shinyapps.io/coronaLandkreise/ | Forecasts are based on TBATS - models (DeLivera, Hyndman und Snyder (2011)) and are updated daily for each German state. Final models are based on comparing multiple models and selecting the best-performing model based on AIC in historical data. The models forecast the 7-day incidence. The forecasting model is the same that is used at our website below. | NA | NA |
| Priesemann Group, MPI-DS | Bayesian SIR | DSMPG-bayes | Sebastian B. Mohr (Max Planck Institute for Dynamics and Self-Organization)sebastian.mohr@ds.mpg.de, Jonas Dehning (Max Planck Institute for Dynamics and Self-Organization)jonas.dehning@ds.mpg.de, Viola Priesemann (Max Planck Institute for Dynamics and Self-Organization)viola.priesemann@ds.mpg.de | https://github.com/Priesemann-Group/covid19_inference_forecast | lgpl-3.0 | primary | Bayesian inference of SIR-dynamics | Max Planck Institute for Dynamics and Self-Organization | NA | ViolaPriesemann, JonasDehning | https://science.sciencemag.org/content/369/6500/eabb9789 | This model simulates SIR-dynamics with a log-normal convolutions of infections to obtain the delayed reported cases. Parameters of the model are sampled with Hamiltonian Monte-Carlo using the PyMC3 Python library. We assume that the infection rate can change every week, with a standard deviation that is also an optimized parameter. When new governmental restrictions are enacted or lifted, we include a small prior to the change of the infection rate. | https://github.com/Priesemann-Group/covid19-forecast-hub-europe | JHU CSSE (confirmed cases; reported fatalities) |
| European COVID-19 Forecast Hub | baseline | EuroCOVIDhub-baseline | European COVID-19 Forecast Hub team hugo.gruson@lshtm.ac.uk | https://covid19forecasthub.eu/ | cc-by-4.0 | other | An baseline model against which other models can be evaluated | NA | This model's development is funded by the European Centre for Disease Prevention and Control. | NA | NA | Baseline model automatically generated using past truth data via the script create-baseline.R in the present repository. Baseline generated using code at https://github.com/reichlab/covidModels/blob/master/R-package/R/quantile_baseline.R | https://github.com/epiforecasts/covid19-forecast-hub-europe | observed incident cases and deaths (from JHU). |
| European COVID-19 Forecast Hub | ensemble | EuroCOVIDhub-ensemble | Katharine Sherratt katharine.sherratt@lshtm.ac.uk, Nikos Bosse nikos.bosse@lshtm.ac.uk, Sebastian Funk sebastian.funk@lshtm.ac.uk | https://covid19forecasthub.eu/ | cc-by-4.0 | primary | An ensemble, or model average, of submitted forecasts to the European COVID-19 Forecast Hub. | NA | This model's development is funded by the European Centres for Disease Countrol. | kathsherratt, ftargument, sbfnk | NA | The current ensemble is specified in https://github.com/epiforecasts/covid19-forecast-hub-europe/blob/main/forecasthub.yml and a history of methods used for the ensemble in https://github.com/epiforecasts/covid19-forecast-hub-europe/blob/main/code/ensemble/EuroCOVIDhub/method-by-date.csv | https://github.com/epiforecasts/covid19-forecast-hub-europe | incident case and death forecasts from other models. |
| Frankfurt Institute for Advanced Studies & Forschungszentrum Jülich | Core Model, Germany | FIAS_FZJ-Epi1Ger | Maria V. Barbarossa barbarossa@fias.uni-frankfurt.de, \ Jan Fuhrmann j.fuhrmann@fz-juelich.de, \ Stefan Krieg s.krieg@fz-juelich.de, \ Jan H. Meinke j.meinke@fz-juelich.de | https://www.fz-juelich.de/SharedDocs/Meldungen/IAS/JSC/DE/2021/2021-01-covid-19.html;jsessionid=F4D5FB4027E871A6F4C2FCAF0F08FC35 | other | primary | An extended SEIR model with additional compartments for undetected cases | NA | NA | NA | https://doi.org/10.1101/2020.04.18.20069955 | An extended SEIR model with additional compartments for undetected cases is fit using multiple segments to the available data for reported cases and then to reported deaths. The parametrization of the last fit is used to forecast incidental and cumulative deaths. Confidence intervals are determined using Monte Carlo sampling of the parameter space. | NA | ECDC, confirmed cases, reported fatalities, population |
| Helmholtz Zentrum fuer Infektionsforschung | age-structured and extended SEIR model | HZI-AgeExtendedSEIR | Isti Rodiah Isti.Rodiah@helmholtz-hzi.de, Berit Lange Berit.Lange@helmholtz-hzi.de, Pratizio Vanella, Alexander Kuhlmann, Wolfgang Bock | https://www.helmholtz-hzi.de/en/nc/research/research-topics/bacterial-and-viral-pathogens/epidemiology/team/ | mit | primary | Deterministic SEIR type model | Helmholtz Zentrum fuer Infektionsforschung, Leibniz Universitaet Hannover, Technische Universitaet Kaiserslautern | NA | NA | NA | This model is an extended SEIR model with additional compartments for the hospitalized, intensive unit, long-COVID, and death. The model is also developed as an age-structured model. The model is fitted to reported cases and reported cumulative deaths. The fitted parameters are used to predict the number of cases and deaths. | NA | RKI (reported cases, reported deaths) |
| ICM / University of Warsaw | agentModel | ICM-agentModel | Rafał Bartczuk, Łukasz Górski, Magdalena Gruziel-Słomka, Artur Kaczorek, Jan Kisielewski, Antoni Moszyński, Karol Niedzielewski, Jędrzej Nowosielski, Maciej Radwan, Franciszek Rakowski, Marcin Semeniuk, Jakub Zieliński | https://covid-19.icm.edu.pl/en/model-description/ | other | primary | Agent-based model | NA | -- | NA | https://www.sciencedirect.com/science/article/pii/S0378437110003687 | The main idea behind our agent-based model is a representation of the social structure of population of Poland to the level of individual citizen and its social contacts. Our model has resolution in space and contexts. It means it can follow development of epidemic on geographical grid in 1km resolution as well as through physical contact in various characteristic contexts of social life. | NA | Polish Ministry of Health (general epidemic data: number of cases, deaths, etc.) and Polish National Institute for Public Health (Detailed epidemic data: time of hospitalization, age, symptoms, contact tracing etc.) |
| IEM Health | CovidProjections | IEM_Health-CovidProject | Brad Suchoski (IEM) bradley.suchoski@iem.com, Steve Stage (IEM) steve.stage@iem.com, Heidi Gurung (IEM) heidi.gurung@iem.com, Sid Baccam (IEM) sid.baccam@iem.com | https://iem-modeling.com/ | cc-by-4.0 | primary | SEIR model projections for daily incident confirmed COVID cases and deaths by using AI to fit actual cases observed. | NA | NA | NA | NA | IEM is currently using an AI model to fit data from various sources and project new cases of COVID-19. We do not assume the average number of secondary infections (R-value) stays the same over time. Our model makes projections for over 375 individual counties in the U.S., all 50 states, the District of Columbia, 3 U.S. territories, and 32 European countries. For each jurisdiction, our model runs 4 million simulations to find the best R value that would allow the SEIR model to fit the data for confirmed cases. Our model assumes that the R value changes quite rapidly over time due to changes in human behavior. Our model uses a sliding window to fit the data and find the best R values for each of the windows. | NA | JHU reported daily confirmed cases and deaths |
| ILM | EKF | ILM-EKF | Stefan Heyder stefan.heyder@tu-ilmenau.de, Thomas Hotz thomas.hotz@tu-ilmenau.de | https://github.com/Stochastik-TU-Ilmenau | cc-by-4.0 | primary | Extended Kalman filter based on reproduction equation | NA | NA | NA | NA | We use the reproduction equation to obtain a state space model whose states comprise incidences, the current reproduction number, the generation time distribution, as well as fraction of deaths by delay. An extended Kalman filter is used to fit this model and obtain forecasts. Quantiles for the latter are obtained from a log-normal distribution with mean and variance given by the prediction steps. | NA | JHU |
| ITWW | county_repro | ITWW-county_repro | Przemyslaw Biecek przemyslaw.biecek@gmail.com, Viktor Bezborodov integral2008-1@mail.ru, Marcin Bodych bodychmarcin@gmail.com, Jan Pablo Burgard burgardj@uni-trier.de, Stefan Heyder stefan.heyder@tu-ilmenau.de, Thomas Hotz thomas.hotz@tu-ilmenau.de, Tyll Krüger tyll.krueger@googlemail.com | https://github.com/Stochastik-TU-Ilmenau | mit | secondary | Forecasts of county level incidence based on regional reproduction numbers. | NA | ECDC | NA | NA | Using county level incidence data we estimate regional reproduction numbers with a small area estimation approach. These estimates are the basis for simulations of future incidences which are then aggregated to the state or national level. Deaths are predicted by age groups and use estimated case fatality ratios. | NA | RKI, ECDC, Polish patient data, county level incidences by age group, distribution for time to death and proportion of deaths in each age group |
| Imperial College London | Death to cases (ascertainr) | Imperial-DeCa | Sangeeta Bhatia (Imperial College London) s.bhatia@imperial.ac.uk, Pierre Nouvellet (Sussex University) p.nouvellet@imperial.ac.uk | https://mrc-ide.github.io/covid19-short-term-forecasts | cc-by-4.0 | primary | Uses both cases and deaths to estimate an observed CFR. Projections are based on the estimated CFR. | NA | NA | PNouvellet sangeeta0312 | NA | NA | https://github.com/mrc-ide/covid19-forecasts-orderly | Cases and deaths from WHO dashboard |
| Imperial College London | RtI0 (jointlyr) | Imperial-RtI0 | Sangeeta Bhatia (Imperial College London) s.bhatia@imperial.ac.uk, Pierre Nouvellet (Sussex University) p.nouvellet@imperial.ac.uk | https://mrc-ide.github.io/covid19-short-term-forecasts | cc-by-4.0 | secondary | Jointly estimates initial incidence and reproduction number | NA | NA | PNouvellet sangeeta0312 | NA | NA | https://github.com/mrc-ide/covid19-forecasts-orderly | Deaths from WHO dashboard |
| Imperial College London | sbkp (apeestim) | Imperial-sbkp | Sangeeta Bhatia (Imperial College London) s.bhatia@imperial.ac.uk, Pierre Nouvellet (Sussex University) p.nouvellet@imperial.ac.uk Kris V Parag k.parag@imperial.ac.uk | https://mrc-ide.github.io/covid19-short-term-forecasts | cc-by-4.0 | proposed | Optimises the window over which reproduction number is assumed to be constant. | NA | NA | PNouvellet sangeeta0312 krisparag1 | NA | NA | https://github.com/mrc-ide/covid19-forecasts-orderly | Deaths from WHO dashboard |
| KITmetricslab | bivar_branching | KITmetricslab-bivar_branching | Johannes Bracher (Karlsruhe Institute of Technology) johannes.bracher@kit.edu | https://github.com/jbracher/branching_process_delta | cc-by-4.0 | primary | Delta-variant and other cases are modelled as independent branching processes, with weekly growth\ \ rates following random walks. Forecasts for 3 and 4 wk are likely unreliable. | NA | Helmholtz Innovation and Data Science Project "SIMCARD" | NA | NA | The total weekly incidence is modelled as the sum of two independent overdispersed branching processes (delta / non-delta cases; may be updated to other pairs of variants later), with the weekly growth rates following multiplicative random walks. Sequencing data are included via an additional binomial observation process with the probabilities for the two variants proportional to their occurrence in the two latent branching processes. Posterior samples are generated using the JAGS software. Priors were chosen as 'uninformative' uniform distributions, but may be specified in a more informative fashion in the future. In order to be included in the ensemble forecasts are generated up to 4 wk into the future, but given the simple model structure, three and four-week-ahead forecasts should be interpreted with caution. | NA | JHU (confirmed cases), RKI sequencing data (variants) |
| Karlen Working Group | python Population Modeller | Karlen-pypm | Dean Karlen (University of Victoria and TRIUMF) karlen@uvic.ca | https://pypm.github.io/home/ | gpl-3.0 | primary | Discrete-time difference equations with long periods of constant transmission rate | NA | none | NA | https://arxiv.org/abs/2007.07156 | NA | NA | JHU (cases and deaths), Google open-data (vaccinations), GISAID (genomic data) |
| Los Alamos National Labs | GrowthRate | LANL-GrowthRate | Dave Osthus dosthus@lanl.gov, Sara Del Valle, Carrie Manore, Brian Weaver, Lauren Castro, Courtney Shelley, Manhong (Mandy) Smith, Julie Spencer, Geoffrey Fairchild, Travis Pitts, Dax Gerts, Lori Dauelsberg, Ashlynn Daughton, Morgan Gorris, Beth Hornbein, Daniel Israel, Nidhi Parikh, Deborah Shutt, Amanda Ziemann | https://covid-19.bsvgateway.org/ | other | primary | This model makes predictions about the future, unconditional on particular intervention strategies. Statistical dynamical growth model accounting for population susceptibility. | NA | U.S. Department of Energy | NA | NA | This model makes predictions about the future, unconditional on particular intervention strategies. The model consists of two processes. The first process is a statistical model of how the number of COVID-19 infections changes over time. The second process maps the number of infections to the reported data. We model the growth of new cases as the product of a dynamic growth parameter and the underlying numbers of susceptible and infected cases in the population at the previous time step, scaled by the size of the state's starting susceptible population. Change 2020-10-29: The growth parameter can be thought of as the transmissibility of the virus in that state on that date and is a weighted regression between the trend in the growth rate over the past 42 days and a growth rate that would keep the number of new daily confirmed cases constant. The weights of these two components are dynamically tuned to the observed data. To model new deaths in the population, we assume that a fraction of the 1,2,3,4, or 5-week moving average of the daily confirmed cases will die. The model learns both the moving average window and the case fatality fraction that best fits the historical observations. | NA | JHU (confirmed cases; reported fatalities), population |
| Universitaet Leipzig IMISE/GenStat | SECIR | LeipzigIMISE-SECIR | Yuri Kheifetz Yuri.Kheifetz@imise.uni-leipzig.de, Holger Kirsten holger.kirsten@imise.uni-leipzig.de, Markus Scholz Markus.Scholz@imise.uni-leipzig.de | https://github.com/holgerman/covid19-forecast-hub-europe | mit | primary | SECIR type model | NA | NA | GenStatLeipzig | NA | We integrate an adapted mechanistic epidemiologic model of the SECIR type into Input-Output Non-Linear Dynamical Systems (IO-NLDS) serving as hidden layers, i.e. the true dynamics cannot directly be observed. Thereby, we include an asymptomatic compartment, a compartment of patients requiring intensive care, and subdivide most of the compartments into three sub-compartments to model time delays. Changing factors of the system due to non-pharmaceutical interventions, changing age-structure of infected population, and changes in testing policy are imposed as inputs to the system. We then estimate parameters by a knowledge synthesis process considering parameter ranges derived from external studies and public data. Specifically, we use Bayesian inference for the parameters’ estimation, which can also be time-dependent. Public data is translated to model outputs not identical but related to hidden states of the model. The model is fitted to data by a full information approach. | https://github.com/holgerman/covid19-forecast-hub-europe | RKI and DIVI, ECDC, age distributions for cases and deaths from RKI as described in methods, case & ICU & death counts |
| Faculty of Mathematics, Informatics, and Mechanics, University of Warsaw | StochSEIR | MIMUW-StochSEIR | Anna Gambin, Krzysztof Gogolewski, Blażej Miasojedow, Ewa Szczurek, Daniel Rabczenko, Magdalena Rosińska | https://covid19.mimuw.edu.pl | other | primary | Extended SEIR model | NA | tba | NA | https://arxiv.org/abs/2007.07156 | SEIR model with extensions: introduction of the undiagnosed compartment; testing limits influencing number of diagnosed cases; stochastic perturbations of time-dependent contact rate | NA | tba |
| CovidAnalytics at MIT | DELPHI | MIT_CovidAnalytics-DELPHI | Michael Lingzhi Li (Model Creator and Maintainer), Hamza Tazi Bouardi (Maintainer), Dimitris Bertsimas (Advisor). For a full list of general contributors to the team, see https://www.covidanalytics.io/team. | https://www.covidanalytics.io/ | apache-2.0 | primary | This model makes predictions for future cases based on a heavily modified SEIR model taking into account underdetection and government intervention. Current interventions are assumed to continue. | NA | NA | NA | https://www.covidanalytics.io/DELPHI_documentation_pdf | This model makes predictions for future cases based on a heavily modified SEIR model. New states are added to the SEIR model to account for cases that went undetected, and an explicit death state is included. The infection rate is corrected with a nonlinear curve that represents the governmental and societal response (which is assumed) to continue until the end of the pandemic). Key parameters for the disease are fixed using a metanalysis conducted by the CovidAnalytics group of over 150 parameters while epidemiological parameters are fitted to hisorical death counts and detected cases. | https://github.com/COVIDAnalytics/DELPHI | JHU, New York Times |
| MOCOS group | agent1 | MOCOS-agent1 | Marek Bawiec, Marcin Bodych, Tyll Krueger, Tomasz Ozanski, Barbara Pabjan, Agata Migalska, Przemyslaw Biecak, Viktor Bezborodov, Ewa Szczurek, Ewaryst Rafajłowicz, Ewa Rafajłowicz, Wojciech Rafajłowicz | https://mocos.pl/ | other | primary | Agent-based microsimulation model | NA | ECDC | NA | https://www.medrxiv.org/content/10.1101/2020.03.25.20043109v2 | Agent based model. Continuous time stochastic microsimulation on the base of census data for household composition , age distribution, working places etc.). Model includes contact tracing - classic and app based- , testing and quarantine . All relevant duration times like incubation time , time till hospitalization, time till testing are sampled fom distributions based on empirical data | NA | tba |
| Masaryk University | ARIMA | MUNI-ARIMA | Andrea Kraus, David Kraus david.kraus@mail.muni.cz | https://krausstat.shinyapps.io/covid19global/ | cc-by-4.0 | primary | Seasonal ARIMA model with outlier detection fitted to transformed daily series. Weekly forecasts are obtained by aggregating bootstrap daily forecast paths. | NA | NA | NA | NA | This is a seasonal ARIMA model with outlier detection fitted to transformed daily series. Weekly forecasts are obtained by aggregating bootstrap daily forecast paths. | NA | JHU CSSE |
| Department of Mathematics and Statistics Masaryk University Team | SEIAR | MUNI_DMS-SEIAR | Veronika Hajnova, Lenka Pribylova | https://share.uzis.cz/s/cmFHjc4jbqPBAER/6_Metodicka_dokumentace_model_ZSEIAR.pdf | cc-by-4.0 | primary | SEIAR model with A compartment of absent unobserved infected estimated from hospital data with incorporated mobility data dependence; optimized to the compartment of all exposed (unobserved included) | NA | MUNI/11/02202001/2020 | NA | NA | This is a model based on a mechanistic compartmental approach, where some parameters are taken from literature, some parameters are estimated from an anonymized dataset of confirmed subjects. It estimates moving ascertainment rate using data of hospitalized subjects (using a proportion of cases not caught before admission to hospital), so it estimates not only the observed part of the epidemic (compartment I) but also the undetected absent infected (compartment A). The model incorporates transmission rate estimate dependence on mobility data and immunization after vaccination and optimizes affected clusters' size to estimate all exposed individuals using the moving ascertainment rate estimate. To model deaths, we incorporated fixed time delay from positivity report to death estimated from data. Currently, we assume around 0.44% IFR, a simple division of the age structure 0-20/20-65/65+ is used. The model estimates continuously from the first outbreak (spring 2020), the prevalence corresponds to the Czech prevalence study from May 2020 as well as to December community test screenings. We incorporated dominance of alpha variant at the beginning of the year 2021 by multiplying the transmission rate by 1.5 (the estimate is based on data that reveal evidence on its dependence on the number of risk contacts - study PAQ research https://zivotbehempandemie.cz/kontakty or mobility data). With the dominance of alpha variant we increased the probability of hospitalization by factor 1.3 and the conditional probability of death in case of hospitalization by factor 1.1. There is a significant decline of death rate of 65+ cohort in hospitals from the second half of March. We started to adjust its value to estimated 0.36% daily decrease based on fit to data of hospitalized positive subjects (starting form submission 2021-05-10). Thanks to time series of vaccinated we could compare two different estimates of the probability of hospitalization and we deduced that we underestimated the prob. of hospitalization and overestimated the size of the affected clusters, and that 1/100 (starting from 2021-03-15) now seems more appropriate than 1/160 (starting from submission 2021-06-07). We incorporated dominance of delta variant. We assume the transmisibility of the delta varinant is 1.4 times higher than the alpha variant (submissions 2021-07-12, 2021-07-19). Starting from submission 2021-07-26 we work again without incresed transmissibility of the delta variant. We submit a single calibration to forecast deaths and cases while the prediction intervals are estimated using time series decomposition. Especially, we subtract a deterministic trend given by the SEIAR model. | NA | UZIS data for predictive models (anonymized set modely_05_datumy.csv of confirmed subjects https://onemocneni-aktualne.mzcr.cz/api/account/dokumentace), google mobility reports (https://github.com/ActiveConclusion/COVID19_mobility/tree/master/google_reports) |
| Grzegorz Redlarski | Sum of districts forecasts | PL_GRedlarski-DistrictsSum | Grzegorz Redlarski (GUMed) gred@gumed.edu.pl | https://docs.google.com/spreadsheets/d/e/2PACX-1vRpH4yhKRts7Co5tydhZhojIPTcTTybms1PqJ9j1tmSBzzPLoU2U9XjUWDwiKYxnE6gMLayl71rpGC8/pubhtml?gid=493251550&single=true | cc-by-4.0 | primary | Modified SIR method, applied to all districts. Forecasts for districts are summed up. | NA | NA | @Grzegorz_Red | NA | Generating forecasts for all districts including historical data analysis. Then the numbers of cases from all districts are summed up. | NA | data provided by the Ministry of Health https://wojewodztwa-rcb-gis.hub.arcgis.com/pages/dane-do-pobrania |
| Robert Walraven | EmpericalSkewedGaussian | RobertWalraven-ESG | Robert Walraven walraven@multiwareinc.com | http://rwalraven.com/COVID19 | cc-by-4.0 | primary | Multiple skewed gaussian distribution peaks fit to raw data | NA | None (I'm a retired Experimental Physicist working at home) | NA | NA | The incremental cases and deaths of a single outbreak predicted by a simple SEIR model can be approximated closely with a particular skewed gaussian distribution that we developed that has four empirical parameters: height, position, left growth rate, and right decay rate. The ESG model assumes the JHU data consists of independent outbreaks at different times, and fits multiple peaks to the data, keeping the number of peaks to a minimum in order to fit with high accuracy the data after Loess smoothing. The most recent peaks are then used to generate a daily forecast forward three months. Hospitalization data and the pattern of older peaks is used to visually confirm the forecasts look correct and if not, manual adjustments are made to the forecast. The ESG model is a emperical mathematical model that makes no epidemiological assumptions and has no epidemiological parameters, so can serve as a baseline forecast against which to compare SEIR-based models. | NA | Existing case and death raw JHU data plus hospitalization data from HealthData.gov. |
| Swiss Data Science Center / University of Geneva | Trend Model | SDSC_ISG-TrendModel | Ekaterina Krymova (Swiss Data Science Center), Dorina Thanou (Center for Intelligent Systems, EPFL), Benjamin Bejar Haro (Swiss Data Science Center), Tao Sun (Swiss Data Science Center) tao.sun@datascience.ch, Gavin Lee (Swiss Data Science Center), Elisa Manetti (University of Geneva), Christine Choirat (Swiss Data Science Center), Antoine Flahault (University of Geneva), Guillaume Obozinski (Swiss Data Science Center) | https://renkulab.shinyapps.io/COVID-19-Epidemic-Forecasting/ | cc-by-4.0 | primary | The Trend Model predicts daily cases and deaths using linear extrapolation on the linear or log scale of the underlying trend estimated by a robust LOESS seasonal-trend decomposition model. | NA | NA | NA | https://renkulab.shinyapps.io/COVID-19-Epidemic-Forecasting/ | Our forecasts are based on the reported numbers of cases and deaths at the country or regional level. Our modeling substantially relies on estimation of the underlying trend by a robust LOESS seasonal-trend decomposition model, which allows to account for non-stationary weekly seasonality, outliers, missing data and delayed reports. To further predict daily cases and deaths we use linear extrapolation of the estimated smooth trend either on the original or on the logarithmic scale. | https://renkulab.io/gitlab/covid-19/covid-19-forecast | JHU CSSE (confirmed cases; reported fatalities) |
| UB | BSLCoV v0.1 | UB-BSLCoV | David Moriña (UB) dmorina@ub.edu | https://dmorina.shinyapps.io/UB-CoV/ | cc-by-sa-4.0 | primary | Bayesian synthetic likelihood estimation for underreported non-stationary time series | Universitat de Barcelona -- UB (www.ub.edu) | UB, Instituto de Salud Carlos III, Gobierno de España | dmorinya | https://arxiv.org/abs/2003.09202 | NA | NA | JHU (https://github.com/epiforecasts/covid19-forecast-hub-europe/wiki/Targets-and-horizons#truth-data) |
| UMass-Amherst | MechBayes | UMass-MechBayes | Dan Sheldonsheldon@cs.umass.edu, Graham Gibsongcgibson@umass.edu, Nick Reich | https://github.com/dsheldon/covid | cc-by-4.0 | secondary | Bayesian compartmental model with observations on cumulative case counts and cumulative deaths. Model is fit independently to each state. Model includes observation noise and a case detection rate. | NA | This model development is supported by NIGMS grant R35GM119582. The content is solely the responsibility of the authors and does not necessarily represent the official views of NIGMS or the National Institutes of Health. | NA | NA | See GitHub for full description | NA | JHU -CSSEGISandData |
| UMass-Amherst | SemiMech | UMass-SemiMech | Graham Gibsongcgibson@umass.edu, Evan Ray, Dan Sheldon Nick Reich | https://github.com/dsheldon/covid | cc-by-4.0 | primary | Bayesian semi-compartmental model with observations on incident case counts and incident deaths. Model is fit independently to each state. Model includes observation noise and a case detection rate. | NA | This model development is supported by NIGMS grant R35GM119582. The content is solely the responsibility of the authors and does not necessarily represent the official views of NIGMS or the National Institutes of Health. | NA | NA | See GitHub for full description | NA | JHU -CSSEGISandData |
| UNED | PreCoV2 v0.1 | UNED-PreCoV2 | José L. Aznarte (UNED) jlaznarte@dia.uned.es, César Pérez (Inverence) cesar.perez@inverence.com, José Almagro (Inverence) jose.almagro@inverence.com, Pedro Álvarez - Roxu pag@xsto.info, Álvaro Ortiz (POPULATE) alvaro@populate.tools, Fernando Blat (POPULATE) fernando@populate.tools | https://precov2.org | cc-by-sa-4.0 | primary | Bayesian time series models with ARIMA noise and fixed transfer functions for each input. | Universidad Nacional de Educacion a Distancia -- UNED (www.uned.es), Inverence -- Bayesian Data Modelling to drive Business Strategy (www.inverence.com), Populate (www.populate.tools) | UNED, Inverence, POPULATE, Instituto de Salud Carlos III, Gobierno de España | jlaznarte, inverence | https://precov2.org/metodologia | NA | NA | EsCovid19Data (https://github.com/montera34/escovid19data) and Google. |
| UNIPV Periscope Working Group | Bayesian Time-varying INGARCH with NPI covariates | UNIPV-BayesINGARCHX | Paolo Giudici (University of Pavia) giudici@unipv.it, Barbara Tarantino (University of Pavia) barbara.tarantino@unipv.it. | https://periscopeproject.eu/home | cc-by-4.0 | primary | Bayesian estimation of time-dependent models with time-varying coefficients to predict COVID-19 positive counts. | University of Pavia, Department of Economics and Management | Pan European Response to the ImpactS of COvid-19 and future Pandemics and Epidemics (PERISCOPE) | NA | Giudici, P., Tarantino, B., A Bayesian time-dependent framework to assess the effectiveness of policy measures on COVID-19 counts. Working paper 2021. | Our model accounts for uncertainty via a Bayesian framework, for time-dependence on past COVID-19 counts via an INGARCH structure and non-linearity via time-varying coefficients. In addition, time-lagged NPI covariates have been coded and incorporated into the Bayesian framework to assess whether policy measures can effectively reduce positive counts. | NA | JHU (confirmed cases), OxCGRT (NPI covariates) |
| University of Southern California | SIkJalpha | USC-SIkJalpha | Ajitesh Srivastava ajiteshs@usc.edu, Frost Tianjian Xu | https://scc-usc.github.io/ReCOVER-COVID-19 | mit | primary | A heterogeneous infection rate model with human mobility for epidemic modeling. Our model adapts to changing trends and provide predictions of confirmed cases and deaths. | NA | Supported by US National Science Foundation Award No. 2027007 (RAPID) [May 2020 - April 2021] | NA | https://arxiv.org/abs/2007.05180 | We use our own epidemic model called SI-kJalpha, preliminary version of which we have successfully used during DARPA Grand Challenge 2014. Our model can consider the effect of many complexities of the epidemic process and yet be simplified to a few parameters that are learned using fast linear regressions. Therefore, our approach can learn and generate forecasts extremely quickly. On a 2 core desktop machine, our approach takes only 3.18s to tune hyper-parameters, learn parameters and generate 100 days of forecasts of reported cases and deaths for all the states in the US. The total execution time for 184 countries is 11.83s and for more than 3000 US counties is 30s. Despite being fast, the accuracy of our forecasts is on par with the state-of-the-art as demonstrated by our evaluation and benchmarking page at https://scc-usc.github.io/covid19-forecast-bench Our model is able to quickly adapt to changing trends, and the variations in parameters during different times/policies allow us to forecast different scenarios such as what would happen if we were to disregard social distancing suggestions. | https://github.com/scc-usc/ReCOVER-COVID-19 | JHU (cases, deaths), Our World in Data (vaccinations) |
| University of Sydney Forecast Lab | One Model by Manifold Embedding | USyd-OneModelMan | Pablo Montero Manso (University of Sydney) pablo.monteromanso@sydney.edu.au | https://github.com/pmontman/covid19forec | cc-by-4.0 | primary | A single autoregressive model fit jointly to all European time series, adding time series from the top regions across the world. A high-dimensional manifold embedding is used capture the process. | NA | NA | NA | https://arxiv.org/abs/2008.00444 | The information of multiple time series can be shared in a single model via a large dimensional manifold embedding. In addition to Europe death series, the regions with the largest average daily deaths are added to reduce the variance of the model estimation and share information (the regions more advanced in the pandemic can help forecast the others). Each time series is time-delay embedded and stacked together before for fitting a single linear autoregressive model. The dimension of the embedding is tuned by temporal validation, the best dimension of the last 4 weeks. This methodology has been successfully applied in the ensemble forecast efforts of Spain and Australia. See citation for detailed description and statistical properties. | NA | JHU (reported fatalities) |
| University of Virginia, Biocomplexity COVID-19 Response Team | Ensemble model | UVA-Ensemble | Aniruddha Adiga (UVA)aniruddha@virginia.edu, Lijing Wang(UVA), Srinivasan Venkatramanan (UVA), Akhil Sai Peddireddy (UVA), Benjamin Hurt (UVA), Przemyslaw Porebski, Bryan Lewis (UVA), Madhav Marathe (UVA), Jiangzhou Chen, Anil Vullikanti. | https://biocomplexity.virginia.edu/ | cc-by-4.0 | primary | An ensemble of multiple methods such as auto-regressive (AR)models with exogenous variables, Long short-term memory (lSTM) models,Kalman filter and PatchSim (an SEIR model). | NA | NA | NA | NA | This is a national-level (specific to DE, PL, GB in this round) multi-method model forecasting the new confirmed cases. Multiple methods include AR methods, an LSTM model(not included in this round),Kalman filters, ARIMA(not included in this round), and PatchSim(not included in this round), variant of SEIR(not included in this round). Multimethod forecasts are combined using Bayesian model averaging. | NA | NA |
| UNIPG_UNIMIB_USI_UNINSUBRIA | MULTINOMIAL_BAYESIAN | UpgUmibUsi-MultiBayes | Francesco Bartolucci (Università di Perugia) francesco.bartolucci@unipg.it, Fulvia Pennoni (Università di Milano Bicocca) fulvia.pennoni@unimib.it, Antonietta Mira (Università della Svizzera italiana and Università dell’Insubria) antonietta.mira@usi.ch | https://github.com/francescobartolucci/ARMultinomial | cc-by-4.0 | primary | Bayesian Dirichlet-Multinomial models for counts of patients in mutually exclusive and exhaustive categories such as hospitalized in regular wards and in intensive care units, deceased and recovered | NA | NA | NA | NA | We us a Bayesian Dirichlet-Multinomial autoregressive models for time-series of counts of patients in mutually exclusive and exhaustive observational categories, defined according to the severity of the patient status and the required treatment. Categories include hospitalized in regular wards (H) and in intensive care units (ICU), together with deceased (D) and recovered (R). These models explicitly formulate assumptions on the transition probabilities between these categories across time, thanks to a flexible formulation based on parameters that a-priori follow Normal distributions, possibly truncated to incorporate specific hypotheses having an epidemiological interpretation. The posterior distribution of model parameters and the transition matrices are estimated by an efficient Markov Chain Monte Carlo algorithm that also provides predictions and allows us to compute the reproduction number Rt. All estimates and predictions are endowed with an accuracy measure obtained thanks to the Bayesian approach. | NA | NA |
| Centre for Modelling of Biological and Social Processes | SeirFilter | bisop-seirfilter | Martin Šmíd, Jan Trnka, Vít Tuček, Milan Zajíček | https://www.medrxiv.org/content/10.1101/2021.02.16.21251834v1 | mit | primary | please see https://www.medrxiv.org/content/10.1101/2021.02.16.21251834v1 | NA | no funding | NA | NA | please see https://www.medrxiv.org/content/10.1101/2021.02.16.21251834v1 | NA | please see https://www.medrxiv.org/content/10.1101/2021.02.16.21251834v1 |
| Centre for Modelling of Biological and Social Processes | SeirFilterLite | bisop-seirfilterlite | Martin Šmíd, Jan Trnka, Vít Tuček, Milan Zajíček | www.bisop.eu | mit | secondary | A simple stochastic SEIR state space model | NA | no funding | NA | NA | A variant of SeirFilter (https://www.medrxiv.org/content/10.1101/2021.02.16.21251834v1) with compartnents S-E-Ia-Ip-Is-R-D where Ia are the asymptomatic, Ip the presymptomatic and Is the symptomatic. Only past data and the Google mobility indices used as inputs, the variability of R0 (virus variant, vaccination, seasonality) is handled via its estimation from past. In particular, R0 is estimated by the weighted time-average of reproduction number (computed as R(t)=(C(t) - C(t-7))/(C(t-5)-C(t-12))) divided by the estimated number of contacts (obtained from the Google mobilty indices). | NA | John Hopkins data, Google Mobility Recors |
| epiMOX | SUIHTER | epiMOX-SUIHTER | Giovanni Ardenghi (Politecnico di Milano) giovanni.ardenghi@polimi.it, Luca Dede' (Politecnico di Milano) luca.dede@polimi.it, Nicola Parolini (Politecnico di Milano) nicola.parolini@polimi.it, Alfio Quarteroni (Politecnico di Milano, EPFL) alfio.quarteroni@polimi.it | https://www.epimox.polimi.it | mit | primary | Compartmental model SUIHTER | NA | NA | NA | https://arxiv.org/abs/2101.03369 | The results are obtained using the compartmental model SUIHTER which simulates the evolution of COVID-19 epidemic in Italy. SUIHTER comprises the following compartments: Susceptibles(S), Undetected(U), Isolated(I), Hospitalized(H), Threatened(T), Extinct(E), Recovered(R). The model parameters are calibrated thanks to a combination of the least squares method with the Markov Chain Monte Carlo (MCMC) method. The current version of the model accounts for two additional compartments collecting individuals who have received the first and the second dose of vaccine, respectively. Different NPIs and vaccination scenarios can be accounted for in the model forecast. | NA | NA |
| Epiforecasts / London School of Hygiene and Tropical Medicine | EpiExpert (epiforecasts) | epiforecasts-EpiExpert | Nikos Bosse nikos.bosse@lshtm.ac.uk, Sam Abbott sam.abbott@lshtm.ac.uk, Sebastian Funk | https://epiforecasts.io/ | mit | primary | Mean ensemble of human predictions | NA | Funding by the Health Protection Research Unit (grant code NIHR200908) | NA | https://github.com/epiforecasts/covid-german-forecasts | Forecasts from experts and non-experts are elicited using a shinyApp. Two variants of this shiynApp exist. In one, forecasters are asked to predict cases and deaths directly. In the other, forecasters are asked to predict Rt. The Rt predictions are then mapped to death and cases using a renewal equation and the R package EpiNow2. Individuals currently make forecasts by choosing a distribution and specifying the median and width of that predictive distribution for evey forecast horizon. Forecasts from both apps are collected and a mean ensemble is formed. | NA | ECDC deaths and cases |
| Epiforecasts / London School of Hygiene and Tropical Medicine | EpiExpert Rt (epiforecasts) | epiforecasts-EpiExpert_Rt | Nikos Bosse nikos.bosse@lshtm.ac.uk, Sam Abbott sam.abbott@lshtm.ac.uk, Sebastian Funk | https://epiforecasts.io/ | mit | other | Mean ensemble of human predictions of Rt which get mapped to cases and deaths using a renewal equation | NA | Funding by the Health Protection Research Unit (grant code NIHR200908) | NA | https://github.com/epiforecasts/covid-german-forecasts | Forecasts of Rt from experts and non-experts are elicited using a shinyApp. Individuals choose a distribution (default is normal) for Rt and change the median and width of that predictive distribution for evey forecast horizon. Based on individual forecasts, samples are drawn from the individual forecast distributions. These samples are used to simulate cases and deaths based on a renewal equation using the R package EpiNow2. | NA | ECDC deaths and cases |
| Epiforecasts / London School of Hygiene and Tropical Medicine | EpiExpert Direct (epiforecasts) | epiforecasts-EpiExpert_direct | Nikos Bosse nikos.bosse@lshtm.ac.uk, Sam Abbott sam.abbott@lshtm.ac.uk, Sebastian Funk | https://epiforecasts.io/ | mit | other | Mean ensemble of human predictions | NA | Funding by the Health Protection Research Unit (grant code NIHR200908) | NA | https://github.com/epiforecasts/covid-german-forecasts | Forecasts from experts and non-experts are elicited using a shinyApp. Individuals currently make forecasts by choosing a distribution and specifying the median and width of that predictive distribution for evey forecast horizon. Forecasts are collected and a mean ensemble is formed. | NA | ECDC deaths and cases |
| Epiforecasts / London School of Hygiene and Tropical Medicine | EpiNow2 (epiforecasts) | epiforecasts-EpiNow2 | Nikos Bosse nikos.bosse@lshtm.ac.uk, Sam Abbott sam.abbott@lshtm.ac.uk, Sebastian Funk | https://epiforecasts.io/EpiNow2 | mit | secondary | Semi-mechanistic estimation of the time-varying reproduction number for latent infections mapped to reported cases/deaths. | London School of Hygiene and Tropical Medicine | Wellcome Trust via a Senior Research Fellowship to Sebastian Funk (210758/Z/18/Z) and the Health Protection Research Unit (grant code NIHR200908) | ftargument, seabbs, sbfnk | https://doi.org/10.5281/zenodo.3957489 | EpiNow2 implements a Bayesian latent variable approach using the probabilistic programming language Stan, which works as follows. For an initial, unobserved, seeding time infections were estimated using an exponential model with priors based on observed growth. For each subsequent time step, previous imputed infections were summed, weighted by an uncertain generation time probability mass function, and combined with an estimate of Rt to give the incidence at that time. The infection trajectories were then mapped to mean reported case counts by convolving over an uncertain incubation period and report delay distribution. Observed reported case counts were then assumed to be generated from a negative binomial observation model with overdispersion, multiplied by a day of the week effect with an independent parameter for each day of the week. Temporal variation was controlled using an approximate Gaussian process with a Matern 3/2 kernel. Rt was assumed to be constant over the forecast horizon although a correction was applied to adjust for the time-varying proportion of the population that was susceptible. Deaths were then modelled as a convolution of forecast cases combined with some scaling factor, a day of the week effect and a negative binomial observation model. | https://github.com/epiforecasts/europe-covid-forecast | ECDC deaths and cases |
| Fraunhofer Institute for Industrial Mathematics ITWM | ITWM | itwm-dSEIR | Jan Mohring jan.mohring@itwm.fraunhofer.de, Neele Leithäuser, Michael Helmling | https://www.itwm.fraunhofer.de/ | mit | primary | cohort based, integral equation | NA | Fraunhofer Gesellschaft, Land Rheinland-Pfalz | NA | https://www.itwm.fraunhofer.de/content/dam/itwm/de/documents/PressemitteilungenPDF/2020/20200429_Bericht_Prognosemodelle-f%C3%BCr-die-Coronapandemie.pdf | Fraunhofer-ITWM's predictions are based on a cohort model that groups people according to four age groups and according to the status infected, detected and successfully vaccinated. The dynamics of the epidemic are described by integral equations. In particular, we assume an infectious period with fixed onset, end and infectivity. The most important parameters are contact rates between age groups, detection rates and times, and death rates and times. They are adjusted to the historical data of the RKI. For forecasts, the simulation is continued with the parameters determined for the last week. In principle, the forecast quality could still be improved by anticipating the effects of events such as the end of public holidays on contact and detection rates. However, this is not yet done in the automatic submissions. All calculations use automatic differentiation. This speeds up parameter adjustment and allows for error estimates. The latter are determined by comparing counted and simulated cases and by matching the empirical standard deviations with the standard deviations predicted by the calculated sensitivities. The model is described in detail in https://www.itwm.fraunhofer.de/de/presse-publikationen/presseinformationen/2021/2021-06-22_Dritte_Welle_Starker-Effekt-von-Schnelltests-an-Schulen.html | NA | RKI |
kathsherratt
sbfnk
@kathsherratt, my understanding is that you started something like this?