Artificial Neural Network for Combining Forecasts

ＡｒｔｉｆｉｃｉａｌＮｅｕｒａｌＮｅｔｗｏｒｋｆｏｒＣｏｍｂｉｎｉｎｇＦｏｒｅｃａｓｔｓ￥ＳｈａｎｍｉｎｇＳｈｉ，ＬｉＤ．Ｘｕ＆ＢａｏＬｉｕ（ＤｅｐａｒｔｍｅｎｔｏｆＣｏｍｐｕｔｅｒＳｃｉｅｎｃｅ，ＵｎｉｖｅｒｓｉｔｙｏｆＣｏｌｏｒａｄｏａｔＢｏｕｌ...     

Combining multiple forecasts given multiple objectives

This paper is concerned with expanding the decision support capabilities of computerized forecasting systems. The expansion allows for the systematic combination of multiple forecasts and the explicit consideration of multiple objectives in the forecast selection process. The methodology used is multiple objective linear programming. Selecting an individual forecast based upon a single objective may not make the best use of available information for a variety of reasons. Combined forecasts may provide a better fit with respect to a single objective than any individual forecast. Even if an individual forecast does provide a good fit with respect to a single objective, a combined forecast may provide a better fit with respect to multiple objectives. An example is used to illustrate the expanded decision support system, its outputs and their properties.     

To combine or not to combine? issues of combining forecasts

This paper addresses issues such as: Does it always pay to combine individual forecasts of a variable? Should one combine an unbiased forecast with one that is heavily biased? Should one use optimal weights as suggested by Bates and Granger over twenty years ago? A simple model which accounts for the main features of individual forecasts is put forward. Bayesian analysis of the model using noninformative and informative prior probability densities is provided which extends and generalizes results obtained by Winkler (1981) and compared with non-Bayesian methods of combining forecasts relying explicitly on a statistical model for the individual forecasts. It is shown that in some instances it is sensible to use a simple average of individual forecasts instead of using Bates and Granger type weights. Finally, model uncertainty is considered and the issue of combining different models for individual forecasts is addressed.     

A note on diagrams for turning-point diagnostics

A new diagram for discriminating a forecasting technique's history with respect to the two types of turning-point errors, missed turns and false signals is presented. This diagram extends Theil's original prediction-realization (P-R) presentation so that each of the four quadrants and the axes locates a distinct type of forecasting error or correct forecast, thus reducing much of the former ambiguity. This improvement is accomplished without the loss of any information.     

Forecast comparisons of four models of U.S. interest rates

In this paper we compare the out of sample forecasts from four alternative interest rate models based on expanding information sets. The random walk model is the most restrictive. The univariate time series model allows for a richer dynamic pattern and more conditioning information on own rates. The multivariate time series model permits a flexible dynamic pattern with own- and cross-series information. Finally, the forecasts from the MPS econometric model depend on the full model structure and information set. In theory, more information is preferred to less. In practice, complicated misspecified models can perform much worse than simple (also probably misspecified) models. For forecasts evaluated over the volatile 1970s the multivariate time series model forecasts are considerably better than those from simpler models which use less conditioning information, as well as forecasts from the MPS model which uses substantially more conditioning information but also imposes ‘structural’ economic restrictions.     

Linear regression forecasting in the presence of ar(1) disturbances

This paper is concerned with time-series forecasting based on the linear regression model in the presence of AR(1) disturbances. The standard approach is to estimate the AR(1) parameter, ρ, and then construct forecasts assuming the estimated value is the true value. We introduce a new approach which can be viewed as a weighted average of predictions assuming different values of ρ. The weights are proportional to the marginal likelihood of ρ. A Monte Carlo experiment was conducted to compare the new method with five more conventional predictors. Its results suggest that the new approach has a distinct edge over existing procedures.