Note: Detailed model scenario descriptions can be found below the graph or on the Technical Notes tab.
Covid Act Now provides statewide and county-level estimates of R-effective, taking mortality and confirmed cases as inputs. Because of potential reporting delays and errors in the data, they perform smoothing, and require 10 preceding days of data. County-level estimates on the "Nowcasts" tab show average R-effective for the last seven (7) days.
Raw dataEpi Forecasts provides national and state-level estimates of R-effective, taking the number of cases as an input. It accounts for the delay from infection to onset of symptoms.
Raw dataCovidestim calculates state-level effective reproductive numbers, taking cases, deaths and test positivity rates as inputs. It has corrections to account for lags in diagnosis, disease duration and mortality risk.
Code repositoryLEMMA is an open-source SEIR model with compartments for hospitalization and symptom severity. The model is calibrated to hospitalization, ICU, and death data using Bayesian methods. LEMMA provides credibility intervals and scenarios for future hospitalization, ICU, deaths, and R_t. Users can download the LEMMA package and input their own data and priors for parameters using R or a simple Excel interface.
Raw dataSC-COSMO is an age-structured, multi-compartment SEIR model calibrated to reported case numbers using a Bayesian approach. The model incorporates contacts patterns by age, the effect of population density, and estimates of the case detection rate. SC-COSMO explicitly considers contacts and transmission in households as well as contacts in work, school, and other settings and the effects of non-pharmaceutical interventions like shelter-in-place and school closures that differentially reduce contacts by venue. Model outputs include number of infections, deaths, and hospitalizations.
Modeling methodsThe UCSF model estimates time-varying reproduction numbers (R_t, or R-effective), the average number of cases infected by a given case over the course of that individual’s disease progression, for select Bay Area and California counties/regions. Model inputs are publically available daily counts of COVID-19 cases, archived by the New York Times. The UCSF researchers use the Wallinga-Teunis technique of real-time estimation of reproduction numbers.
Raw dataTo obtain the Rt estimate, the Harvard Xihong Lin Group uses the EpiEstim method (Cori, A., et al., 2013; Thompson, R.N., et al., 2019) to estimate the daily Rt value, as implemented in the EpiEstim R package. The EpiEstim method requires the following inputs: daily positive increase in cases (source used is JHU-CSSE), the time window of daily positive increase in cases to be averaged (7-day window is used), and the serial interval (used a mean of 5.2 days and a SD of 5.1 days).
Raw dataThe CovidActNow model is a SEIR model with compartments for disease severity and medical intervention. Each county and state is calibrated separately, and R-effective is inferred using observed data.
Raw dataIHME is a multi-stage model, where the first stage fits an S-curve to historical daily deaths data, and the second stage is an SEIR compartment model. The SEIR model's R-effective is calibrated using the output of the first stage, but it also incorporates temperature data, population density, local testing capacity, and changes in mobility data. IHME provides projections of mortality, number of infections, and hospital utilization at the state and national level.
Raw dataThe Global Epidemic and Mobility Model (GLEAM) uses a individual-based, stochastic spatial epidemic model. The model uses mobility data and travel patterns to simulate spatial contact patterns. The likely ranges of basic parameters, such as R0 and IFR, are inferred from observed data.
Raw dataMIT DELPHI is a standard SEIR model with compartments for undetected cases and hospitalizations. R-effective is modeled as an S-curve to reflect government interventions and social distancing.
Raw dataA statistical machine learning extrapolation algorithm CLEP which forecasts deaths with MEPI prediction intervals with one week or two in advance by county. The forecast is calculated from an ensemble of linear and exponential predictors (CLEP), some of which pool data across counties or use demographic data.
Raw dataThe Reich Lab at the UMass-Amherst is an Influenza Forecasting Center of Excellence and the source for the official CDC COVID-19 Forecasting page. Taking other forecasts as the input, this is arithmetic average across eligible models of cumulative deaths forecasts. Forecasts are weekly out to four (4) weeks, at the state and national level.
Raw dataThe COVID Scenario Pipeline is a county-level metapopulation model that incorporates commuting patterns and stochastic SEIR disease dynamics. To produce near-term forecasts of deaths and hospitalizations in the population, county-specific transmission and county-specific risks of hospitalization and death were inferred using a novel Bayesian inference algorithm. This algorithm calibrates the model to weekly county-level incident cases and deaths, as reported by USAFacts. Fixed time delays from infection to case confirmation, hospitalization, and death were employed in projecting these outcomes. The estimates reported by this model incorporate uncertainty in baseline R0, the duration of the infectious period, the effectiveness of statewide intervention policies, and process stochasticity.
Code RepositoryLEMMA is an open-source SEIR model with compartments for hospitalization and symptom severity. The model is calibrated to hospitalization, ICU and death data using Bayesian methods. LEMMA provides credibility intervals and scenarios for future hospitalization, ICU, deaths and Rt. Users can download the LEMMA package and input their own data and priors for parameters using R or a simple Excel interface.
Raw dataSC-COSMO is an age-structured, multi-compartment SEIR model calibrated to reported case numbers using a Bayesian approach. The model incorporates contacts patterns by age, the effect of population density, and estimates of the case detection rate. SC-COSMO explicitly considers contacts and transmission in households as well as contacts in work, school, and other settings and the effects of non-pharmaceutical interventions like shelter-in-place and school closures that differentially reduce contacts by venue. Model outputs include number of infections, deaths, and hospitalizations.
Modeling methodsSC-COSMO-Lightweight is a new, short-term forecasting model. For each county, it starts with daily incident detected cases and uses General Additive Model smoothing to remove day-of-the-week reporting variation. It then estimates daily R-effective values based on the smoothed time-series of detected cases employing methods by described Wallinga and Teunis (2004). R-effective is then extrapolated for the desired forecasting period by considering its current level and how long it has been above 1.0 as well as considering the duration of time spent above 1.0 from previous outbreak periods. Forecasts of incident detected cases are then generated based on: 1) the smoothed incident detected case series; 2) the R-effective estimates/extrapolations; and 3) the Covid-19 serial interval. Finally, prevalent hospitalizations are generated by calibrating the time-varying probability of each incident case being hospitalized to county-level Covid-19 hospital census data, using our prior hospitalization module with length of stay distributions derived from data reported for California hospital systems (www.sc-cosmo.org). With the extrapolated incident case series and extrapolation of the probability of hospitalization for each case and length of stay distribution, prevalent Covid-19 hospitalization prevalence is then extrapolated.
The Columbia model projects nationwide, county-level estimates of R-effective, daily new confirmed case, daily new infection (both reported and unreported), cumulative demand of hospital beds, ICU, and ventilators, as well as daily mortality (2.5, 25, 50, 75 and 97.5 percentiles). County-level case and death data are compiled from Johns Hopkins University and USAFACTS.
Raw dataCovidNearTerm is a bootstrap-based method based on an autoregressive model to estimate at the county level the expected number of COVID-19 patients that will hospitalized 2-4 weeks into the future. It is based on the work of researchers at UCSF (Adam Olshen), Stanford (Kristopher Kapphahn, Ariadna Garcia, Isabel Wang and Manisha Desai) and Memorial Sloan-Kettering (Mithat Gonen).
Raw dataWe employ a purely data-driven model named ACTS to forecast COVID-19 related data, e.g. confirmed cases, hospitalizations and deaths, along time. We assume that the development of the pandemic in the current region will be highly similar to another region with similar patterns a few months ago. We use attention mechanism to compare and match such patterns and generate forecasts. We also leverage additional features such as demographic data and medical resources to more precisely measure the similarity between regions.
Raw dataModel forecasts are the result of utilizing the forecast package's automatic ARIMA forecasting model. Note that the form of the model may vary between counties and over subsequent published forecasts.
Modeling methodsGoogle Cloud’s COVID-19 Public Forecasts provide data on the estimated spread of COVID-19 throughout the United States. The forecasts predict projected death toll, confirmed case counts, hospitalizations, and other important values for tracking and projecting the spread of COVID-19. The forecasts predict data for 28 days, on a state and county level. The model expands upon the SEIR model, which assigns each individual in the population to a “compartment” based on their disease state. The model uses machine learning to estimate how quickly individuals transition between compartments based on historical data and taking into account other relevant factors (“covariates”) that influence the transition rates.
Raw state dataVaccination
All Scenarios assume 35% effectiveness after first dose, and 85% effectiveness after second doses of both Pfizer and Moderna vaccines. Johnson & Johnson are assumed to no longer be in use.
Low vaccination scenarios assume that coverage nationally reaches 70% by December 1, 2021.
High vaccination scenarios assume that coverage nationally reaches 80% by December 1, 2021.
Variants
All scenarios assume 70% prevalence of B.1.617+ (Delta) nationally by Dec 31, 2021
Low variant scenarios assume the B.1.617+ (Delta) variant is 20% more transmissible than the B.1.1.7. variant (first discovered in UK).
High variant scenarios assume the B.1.617+ (Delta) variant is 60% more transmissible than the B.1.1.7. variant (first discovered in UK).
LEMMA provides optimmistic and pessimistic scenarios to depict how emerging variants of concern might change disease trajectories.
LEMMA is an open-source SEIR model with compartments for hospitalization and symptom severity. The model is calibrated to hospitalization, ICU and death data using Bayesian methods. LEMMA provides credibility intervals and scenarios for future hospitalization, ICU, deaths and Rt. Users can download the LEMMA package and input their own data and priors for parameters using R or a simple Excel interface.
Raw data