Combining multiple forecasts in order to generate a single, more accurate one is a well-known approach. A simple average of forecasts has been found to be robust despite theoretically better approaches, increasing availability in the number of expert forecasts, and improved computational capabilities. The dominance of a simple average is related to the small sample sizes and to the estimation errors associated with more complex methods. We study the role that expert correlation, multiple experts, and their relative forecasting accuracy have on the weight estimation error distribution. The distributions we find are used to identify the conditions when a decision maker can confidently estimate weights versus using a simple average. We also propose an improved expert weighting approach that is less sensitive to covariance estimation error while providing much of the benefit from a covariance optimal weight. These two improvements create a new heuristic for better forecast aggregation that is simple to use. This heuristic appears new to the literature and is shown to perform better than a simple average in a simulation study and by application to economic forecast data.

Economic agents often face situations, where there are multiple competing fore- casts available. Despite five decades of research on forecast combinations, most of the methods introduced so far fail to outperform the equal weights forecast combination in empirical applications. In this study, we gather a wide spectrum of forecast combination methods and reexamine these findings in two different classical economic times series forecasting applications. These include out-of- sample combining forecasts from the ECB Survey of Professional Forecasters and forecasts of the realized volatility of the U.S. Treasury futures log-returns. We asses the performance of artificial predictions markets, a class of machine learning methods, which has not yet been applied to the problem of combin- ing economic times series forecasts. Furthermore, we propose a new simple method called Market for Kernels, which is designed specifically for combining time series forecasts. We found that equal weights can be significantly out- performed by several forecast combinations, including Bates-Granger methods and artificial prediction markets in the ECB Survey of Professional Forecasters application and by almost all examined forecast combinations in the financial application. We also found that the Market for Kernels forecast...

Combining forecasts have been proven as one of the most successful methods to improve predictive performance. However, while there often is a focus on theoretically optimal methods, this is an ill-posed issue in practice where the problem of robustness is of more empirical relevance. This thesis focuses on the latter issue, where the risk associated with different combination methods is examined. The problem is addressed using Monte Carlo experiments and an application to automated forecasting with data from the M4 competition. Overall, our results indicate that the choice of combining methodology could constitute an important source of risk. While equal weighting of forecasts generally works well in the application, there are also cases where estimating weights improve upon this benchmark. In these cases, many robust and simple alternatives perform the best. While estimating weights can be beneficial, it is important to acknowledge the role of estimation uncertainty as it could outweigh the benefits of combining. For this reason, it could be advantageous to consider methods that effectively acknowledge this source of risk. By doing so, a forecaster can effectively utilize the benefits of combining forecasts while avoiding the risk associated with uncertainty in weights.

The objective of this paper is to retrospectively evaluate forecast models for polling data, to be used prospectively for the Swedish general election in 2022. One of the simplest ways of forecasting an election result is through opinion polls, and using the latest observation as the forecast. This paper considers five different forecasting models on polling data which are evaluated based on different error measures and the results are compared to previous research done on the same topic. The data in this paper consists of time series data of party-preference polls from Statistics Sweden. When forecasting polling data, the naive forecasting model was the most accurate, but forecasting the election in 2018 resulted in the forecast combinations model being the most accurate. Finally, the models are used to make forecasts on the Swedish general election taking place in September of 2022.