The study, published in Harvard Data Science Review, focuses on the spread of epidemics and specifically the current COVID-19 coronavirus pandemic, but the models are applicable to many other disciplines as well.
Models that seek to estimate epidemic trajectories fall into two classes: structural models that seek to capture the causal mechanisms underlying disease transmission, and time series models that are oblivious to the precise mechanism and extract changing trends from historical data. When there is uncertainty about many clinical and transmission aspects of disease, as in the current pandemic, “structural models are forced to rely on many assumptions and unknowns,” the study says.
The study’s new class of time series models therefore provides a “simple and transparent” route to predict an epidemic’s trajectory – and can be adapted to include additional factors including seasonal or day-of-the-week effects.
Specifically, the paper formulates a class of models in which the growth rate of the cumulative series – for example the total number of COVID-19 infections to date – changes over time. “Allowing the trend to be time-varying introduces further flexibility and enables the effects of changes in policy to be tracked and evaluated,” the study says.
The paper – entitled “Time series models based on growth curves with applications to forecasting coronavirus” – is co-authored by Andrew Harvey, Emeritus Professor of Econometrics at the Faculty of Economics of the University of Cambridge, and by Dr Paul Kattuman, Reader in Economics at Cambridge Judge Business School.
“The significance of the time series models is that good forecasts of an epidemic’s trajectory can be developed before the epidemic reaches a peak as measured by new cases, deaths, or both those factors,” says Paul Kattuman.
“Structural models can be very useful in illuminating causal mechanisms and counterfactual analysis. But when there are so many uncertainties and changing external factors, as in the current pandemic, these models must make a lot of assumptions about structures and mechanics that we simply don’t know much about with any assurance.”
The authors conclude by suggesting that the possibility of successive waves can be monitored “by tracking the filtered estimates of new cases or deaths given by our model” – but caution that these methods will be useful in practice only if there is “reliable up-to-date observations on new cases.”
The method has been successfully applied to provide weekly forecasts of COVID-19 case rates and hospital bed occupancy for the NHS in the East of England as well as for Kerala state in India.