Seasonal time series data imputation: Comparison between feed forward neural networks and parametric approaches

Missing data in a time series may be an obstacle that may prevents further analysis of the available series either for control, explanations or forecasting. This paper addresses the problem of "‘filling in"' missing data is a segment of seasonal univariate time series in order for further analysis to be possible. We focus on extrapolation from models fitted to available segments using both parametric and non parametric methods. Specifically we examine how recursive and direct estimates from forward and backward learning Artificial Neural Networks (ANN) compares with seasonal ARIMA estimates and interpolation estimates of Additive outliers in seasonal ARIMA models. A comparison statistics is also proposed. Keywords: Time Series; Artificial Neural Network (ANN); Direct estimates; Recursive estimates; Outliers; Proximity > East African Journal of Statistics Vol. 1 (1) 2005: pp. 68-8