Exponential smoothing decomposes time series into interpretable components (such as level, trend, and seasonality), and is used to understand and to forecast time series is weather prediction, financial markets, energy demands, and indicators of economic stability. Outliers, missing data, and nonstationary series make it difficult to reliably forecast trends and associated uncertainty.
We develop a new time series model, by linking together exponential smoothing cells in a dynamic framework. The resulting approach uses robust losses and regularization to handle outliers and high levels of noise, detect evolving trends, interpolate missing data, and improve forecasting. The new model is fit by solving a single convex optimization problem.