Background: Recently the 2019-nCov pneumonia outbreak in China then the world. Search Query performs well in forecasting the epidemics. It is still a suspense whether search query can forecast the drift and the inflexion in 2019-nCov pneumonia. Based on the Baidu Search Index, we propose three prediction models: composite Index, composite Index with filtering (Fourier Transform) and suspected NCP(Novel Coronavirus Pneumonia) cases. With the trained models, we predict the new confirmed cases of 2019-nCov of forecast-period from Feb. 3rd to 9th. Attempting to identify the next peak period, we further estimate 10 day out-of-sample of the new confirmed cases from Feb. 10th to 19th.
Results: We select 16 search queries related to NCP and calculate the correlation coefficient. The maximum correlation coefficient of search queries is above 0.8. The composite Index performs 10 days ahead of the new confirmed cases. With the In-sample prediction, the result demonstrates that the predictive model of composite index with filtering performs the best with MAPE 24.98% and RMSE 192.71. By contract, the predictive model of the suspected NCP cases is calculated with the prediction error of MAPE 8.82% and RMSE 368.51(almost twice the best model). With the Out-of-sample prediction, we monitor that there might be a peak value in Feb. 16th to 17th in the next ten days.
Conclusion: With noise filtering, the predictive model can forecast the new confirmed NCP cases more accurately. However, the filtering eliminates the violent fluctuations of the series and cannot capture the rising and declining details of the predicted values. On contrast, the prediction accuracy based on search composite index is sensitive to prediction of peak and valley although its prediction error is larger. These two predictive models can be combined: monitoring the further volatility trend with filtering model while identifying the inflexion with composite model.