While the statsmodels implementation of this model also accepts exogenous variables (SARIMAX), we decided to stick with modeling how previous cases and deaths can predict future ones. Instead we turned our attentions back to the Seasonal Auto-Regressive Integrated Moving Average model (SARIMA). Much more evidence would need to be gathered to support such an outrageous finding! The only other explanation is that stricter mandates actually cause rising case rates, and this seems unlikely. This, too, led to the same problem of the model learning the reversed causitive relationship between exogenous and endogenous variables. Instead of feeding the model the current NPIs (no pharmeceutical interventions, things like school clostures, mask mandates, and travel restrictions) we gave it what they were two weeks before, hoping perhaps that the model would find the lagged relationships between the interventions and the cases. We also tried introducing a lag on the exogenous variables. While governments increase the stringency of the restrictions in response to increased cases, the model seemed to predict that increased stringency would increase infection rates. The accuracy on the test range improved when proprosed intervention changes were provided, but we suspect that the model reversed the direction of the causation of the interventions. We introduced government interventions to attempt to improve the accuracy. We split the data between case totals before December first as a training set and after December 1st to validate the models. We achieved 84% accuracy with Facebook Prophet and 99% accuracy with tuned SARIMAX model. We prepared the data and used both SARIMAX and Facebook Prophet models to make predictions. We were inspired by the COVID-19 Xprize competition to create a model of daily cases to predict future infection rates.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |