Accuracy of video frame size forecasting

We experimentally explore the forecast accuracy of three forecast model types, namely (i) Auto-Regressive (AR), (ii) Moving Average (MA), and (iii) Exponentially Weighted Moving Average (EWMA). Specifically, we use a set of five MPEG-4 video traces and measure the Normalized Mean Squared Error (NMSE) for each of the model types with varying model parameters. We find that increasing model order is not beneficial for I and P frame size forecasting, but is beneficial to B frame size forecasting. Most importantly, we find that the forecast accuracy has the strongest dependence on the statistical properties of the video content itself. It appears that the forecast accuracy is not dependent on the forecast model properties but rather strictly dependent on the logical and statistical properties of the content.