A comparison between bootstrap and dropout for uncertainty estimates of time series forecast using a convolutional neural network
Abstract: Forecasting the future is important in many applications, including forecasting future sales volumes. However, point forecasts without any estimates of the uncertainty are not as useful as if the uncertainty in the forecast is included. Uncertainty estimates allow a range of possible outcomes to be considered, not only the most likely as in the case of point forecasts.In this study the aim is to produce uncertainty estimations for forecasts made 120 days into the future for three different products. The three products are usually used for several years by the consumers and all three products belong to the same category. For the purpose of estimating the uncertainty in forecasts made with a neural network, two different methods were used, bootstrapping residuals and dropout, respectively. The methods were applied to a convolutional neural network, CNN, with dilated causal convolutions. The two methods were compared based one their coverage probability when constructing an interval which aims for 95 % coverage probability. In brief, bootstrapping residual performed better on all three products with a mean coverage probability of 76 % compared to 43 % for dropout. The study shows that it is possible to estimate the uncertainty in forecasts made with a CNN using bootstrap and dropout. Bootstrap seems more suitable than dropout, however the study was only made on three products in the same product category which makes it hard to draw any general conclusions.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)