Online Prediction under Model Uncertainty Via Dynamic Model Averaging: Application to a Cold Rolling Mill

Abstract

Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it does not display a currently valid OMB control number. We found that when only a small number of physically motivated models were considered and one was clearly best, the method quickly converged to the best model, and the cost of model uncertainty was small; indeed DMA performed slightly better than the best physical model. When model uncertainty and the number of models considered were large, our method ensured that the penalty for model uncertainty was small. At the beginning of the process, when control is most difficult, we found that DMA over a large model space led to better predictions than the single best performing physically motivated model. AbstractWe consider the problem of online prediction when it is uncertain what the best prediction model to use is. We develop a method called Dynamic Model Averaging (DMA) in which a state space model for the parameters of each model is combined with a Markov chain model for the correct model. This allows the "correct" model to vary over time. The state space and Markov chain models are both specified in terms of forgetting, leading to a highly parsimonious representation. The method is applied to the problem of predicting the output strip thickness for a cold rolling mill, where the output is measured with a time delay. We found that when only a small number of physically motivated models were considered and one was clearly best, the method quickly converged to the best model, and the cost of model uncertainty was small; indeed DMA performed slightly better than the best physical model. When model uncertainty and the number of models considered were large, our method ensured that the penalty for model uncertainty was small. At the beginning of the process, when control is most difficult, we found that DMA over a large model space led to better predictions than the single best performing physically motivated model.

Dynamic Model Averaging and CPI Inflation Forecasts: A Comparison between the Euro Area and the United States

“…Further, if different forecasting models may apply at each point in time, then the number of model combinations that must be estimated in order to forecast at time τ amounts to 2 mτ . To counter this limit, Raftery et al (2007Raftery et al ( , 2010 propose to use approximations based on state-space methods with the Kalman filter. These approximations involve two parameters λ and α, which they refer to as forgetting factors.…”

“…To simplify the modelling approach, Raftery et al (2007Raftery et al ( , 2010) approximate equation 7 by 5…”

“…Recently, Koop and Korobilis (2012) applied the DMA/DMS approach to forecast inflation in the USA. Dynamic model averaging /dynamic model selection (DMA/DMS), put forward by Raftery et al (2007Raftery et al ( , 2010, allows not solely for coefficients to change over time but also for changes in the entire forecasting model over time. This forecasting approach has been previously applied for spot and future prices in the carbon markets (Koop and Tole, 2012), exchange rates (Buncic and Frey, 2012;De Bruyn et al, 2013) and GDP growth dynamics (Nicoletti and Passaro, 2012).…”

Dynamic Latent Class Model Averaging for Online Prediction

“…We have presented a probabilistic model and a statistical method for online prediction. The method builds on the dynamic model averaging of Raftery et al (2010) developed for the cold rolling mill problem, and extends it by accounting for spatial correlation and time dependence through the proposed polychotomous regression model for the weights of the mixture models. The proposed DLCMA method combines a state-space model for the parameters of the time series or regression models used for forecasting with a Markov chain describing switches in the model governing the system.…”

“…Much work has been done to combine models and procedures based on the same data. Recent methods include convex combinations of parametric and kernel estimates for density estimation (Olkin and Spiegelman, 1987;Long et al, 2011), cross-validation based 'stacking' (Wolpert, 1992;Dzeroski and Zenko, 2004), a bootstrap-based method (LeBlanc and Tibshirani, 1996), a stochastic approximation-based method (Juditsky and Nemirovski, 2000), information-theoretic methods to combine density estimators (Catoni, 1997;Kapetanios et al, 2008) and Bayesian model averaging (Hoeting et al, 1999;Raftery et al, 2010). This article is motivated by the problem of forecasting IBM revenue based on historical revenues for different regions over time.…”