Automatic Forecasting Systems

David Reilly had the idea to try and automate the Box-Jenkins modeling process. George Box told him that "you can't do that". You don't want to tell a stubborn Irishman that so he went off and like SAS and BMDP licensed the code Box-Jenkins created and modified it and launched a product called "Autobj" which was available on Time Sharing Mainframes in the 1970's on IDC, Compuserve, dial up platforms. It evolved and evolves to this day! Autobox was the first-to-market forecasting software and picked as the best dedicated forecasting software in the "Principles of Forecasting" textbook and now website. Autobox was the best performing automated software in the 2008 Neural Network Forecasting Competition! Instead of using a pick best approach Autobox automatically builds the model (univariate and multivariate) all while identifying and incorporating any interventions. It selects the best lead/lag structures for each input series and corrects for omitted variables (holidays or price changes that have affected the historical data, but that the system has no knowledge of) by identifying pulses, seasonal pulses, level shifts and local time trends, and then adding the needed structure through surrogate variables. Handles daily data specific issues like day of the week, week of the year, and day of the month effects.