Modelling election poll data using time series analysis

There is much interest in election forecasting in the UK. On election night, fore­casts are made and revised as the night progresses and seats declare results. We propose a new time series model which may be used in this context. Firstly, we have statistical models for the polls conducted in a run-up to the election; the model produces the distribution of voting amongst the parties. The key here is the use of modelling the probability of voting each poll as latent variables. Secondly, we use this information in the forecasting of the inevitable outcome, continually revising our forecasts as the actual declarations are made, until we can actually determine what we believe the final outcome to be, before it actually happens. We outline the nature and history of elections in the UK, and provide an account of time series analysis. These tools, as well as the theoretical basis of our method, the h-likelihood, are then applied to the creation of each of our models proposed. We study simulations of the models and then fit the models to actual data to assess forecasting accuracy, using existing models for comparison.