arXiv subjects
Publications and source records attributed to Stephen Jewson.
Uncertainty around multimodel ensemble forecasts of changes in future climate reduces the accuracy of those forecasts. For very uncertain forecasts this effect may mean that the forecasts should not be used. We investigate the use of the well-known Bayesian Information Criterion (BIC) to make the decision as to whether a forecast should be used or ignored.
Objective probabilistic forecasts of future climate that include parameter uncertainty can be made by using the Bayesian prediction integral with the prior set to Jeffreys' Prior. The calculations involved in determining the prior can then be simplified by making parametric assumptions about the distribution of the output from the climate model. The most obvious assumption to make is that the climate model output is normally distributed, in which case evaluating the prior becomes a question of evaluating gradients in the parameters of the normal distribution. In previous work we have considered the special cases of diagonal (but not constant) covariance matrix, and constant (but not diagonal) covariance matrix. We now derive expressions for the general multivariate normal distribution, with non-constant non-diagonal covariance matrix. The algebraic manipulation required is more complex than for the special cases, and involves some slightly esoteric matrix operations including taking the expectation of a vector quadratic form and differentiating the determinants, traces and inverses of matrices.
To include parameter uncertainty into probabilistic climate forecasts one must first specify a prior. We advocate the use of objective priors, and, in particular, the Jeffreys' Prior. In previous work we have derived expressions for the Jeffreys' Prior for the case in which the observations are independent and normally distributed. These expressions make the calculation of the prior much simpler than evaluation directly from the definition. In this paper, we now relax the independence assumption and derive expressions for the Jeffreys' Prior for the case in which the observations are distributed with a multivariate normal distribution with constant covariances. Again, these expressions simplify the calculation of the prior: in this case they reduce it to the calculation of the differences between the ensemble means of climate model ensembles based on different parameter settings. These calculations are simple enough to be applied to even the most complex climate models.
When climate forecasts are highly uncertain, the optimal mean squared error strategy is to ignore them. When climate forecasts are highly certain, the optimal mean squared error strategy is to use them as is. In between these two extremes there are climate forecasts with an intermediate level of uncertainty for which the optimal mean squared error strategy is to make a compromise forecast. We present two new methods for making such compromise forecasts, and show, using simulations, that they improve on previously published methods.
We describe a simple method that utilises the standard idea of bias-variance trade-off to improve the expected accuracy of numerical model forecasts of future climate. The method can be thought of as an optimal multi-model combination between the forecast from a numerical model multi-model ensemble, on one hand, and a simple statistical forecast, on the other. We apply the method to predictions for UK temperature and precipitation for the period 2010 to 2100. The temperature predictions hardly change, while the precipitation predictions show large changes.
The CMIP3 multi-model ensemble spread most likely underestimates the real model uncertainty in future climate predictions because of the similarity, and shared defects, of the models in the ensemble. To generate an appropriate level of uncertainty, the spread needs inflating. We derive the mathematical connection between an assumed level of correlation between the model output and the necessary inflation of the spread, and illustrate the connection by making temperature predictions for the UK for the 21st century using four different correlation scenarios.
We argue that it would be desirable to use Jeffreys' priors in the construction of numerical model based probabilistic climate forecasts, in order that those forecasts could be argued to be objective. Hitherto, this has been considered computationally unfeasible. We propose an approximation that we believe makes it feasible, and derive closed-form expressions for various simple cases.
We employ a statistical model of North Atlantic tropical cyclone (TC) tracks to investigate the relationship between sea-surface temperature (SST) and North American TC landfall rates. The track model is conditioned on summer SST in the tropical North Atlantic being in either the 19 hottest or the 19 coldest years in the period 1950-2005. For each conditioning many synthetic TCs are generated and landfall rates computed. Compared to direct analysis of historical landfall, the track model reduces the sampling error by projecting information from the entire basin onto the coast. There are 46% more TCs in hot years than cold in the model, which is highly significant compared to random sampling and corroborates well documented trends in North Atlantic TC number in recent decades. In the absence of other effects, this difference results in a significant increase in model landfall rates in hot years, uniform along the coast. Hot-cold differences in the geographic distribution of genesis and in TC propagation do not significantly alter the overall landfall-rate difference in the model, and the net landfall rate is 1/4.7 yr in hot years and 1/3.1 yr in cold years. SST influence on genesis site and propagation does modify the geographic distribution of landfall, however. The Yucatan suffers 3 times greater landfall rate in hot years than cold, while the U.S. mid-Atlantic coast exhibits no significant change. Landfall probabilities increase in hot years compared to all years in Florida, the U.S Gulf coast, the Mexican Gulf coast, and Yucatan with at least 95% confidence.
One possible method for predicting landfalling hurricane numbers is to first predict the number of hurricanes in the basin and then convert that prediction to a prediction of landfalling hurricane numbers using an estimated proportion. Should this work better than just predicting landfalling hurricane numbers directly? We perform a basic statistical analysis of this question in the context of a simple abstract model, and convert some previous predictions of basin numbers into landfalling numbers.
We are building a hurricane number prediction scheme based on first predicting main development region sea surface temperature (SST), then predicting the number of hurricanes in the Atlantic basin given the SST prediction, and finally predicting the number of US landfalling hurricanes based on the prediction of the number of basin hurricanes. We have described a number of SST prediction methods in previous work. We now investigate the empirical relationship between SST and basin hurricane numbers, and put this together with the SST predictions to make predictions of both basin and landfalling hurricane numbers.
We consider two ways that one might convert a prediction of sea surface temperature (SST) into a prediction of landfalling hurricane numbers. First, one might regress historical numbers of landfalling hurricanes onto historical SSTs, and use the fitted regression relation to predict future landfalling hurricane numbers given predicted SSTs. We call this the direct approach. Second, one might regress \emph{basin} hurricane numbers onto historical SSTs, estimate the proportion of basin hurricanes that make landfall, and use the fitted regression relation and estimated proportion to predict future landfalling hurricane numbers. We call this the \emph{indirect} approach. Which of these two methods is likely to work better? We answer this question for two simple models. The first model is reasonably realistic, but we have to resort to using simulations to answer the question in the context of this model. The second model is less realistic, but allows us to derive a general analytical result.
We are developing schemes that predict future hurricane numbers by first predicting future sea surface temperatures (SSTs), and then apply the observed statistical relationship between SST and hurricane numbers. As part of this overall goal, in this study we compare the historical performance of three simple statistical methods for making five-year SST forecasts. We also present SST forecasts for 2006-2010 using these methods and compare them to forecasts made from two structural time series models.
There is a clear positive correlation between boreal summer tropical Atlantic sea-surface temperature and annual hurricane numbers. This motivates the idea of trying to predict the sea-surface temperature in order to be able to predict future hurricane activity. In previous work we have used simple statistical methods to make 5 year predictions of tropical Atlantic sea surface temperatures for this purpose. We now compare these statistical SST predictions with SST predictions made by an ensemble mean of IPCC climate models.
One way to predict hurricane numbers would be to predict sea surface temperature, and then predict hurricane numbers as a function of the predicted sea surface temperature. For certain parametric models for sea surface temperature and the relationship between sea surface temperature and hurricane numbers, closed-form solutions exist for the mean and the variance of the number of predicted hurricanes, and for the standard error on the mean. We derive a number of such expressions.
We are building a hurricane number prediction scheme that relies, in part, on statistical modelling of the empirical relationship between Atlantic sea surface temperatures and landfalling hurricane numbers. We test out a number of simple statistical models for that relationship, using data from 1900 to 2005 and data from 1950 to 2005, and for both all hurricane numbers and intense hurricane numbers. The results are very different from the corresponding analysis for basin hurricane numbers.
There is significant correlation between main development region sea surface temperature and the number of hurricanes that form in the Atlantic basin. The correlation between the same sea surface temperatures and the number of \emph{landfalling} hurricanes is much lower, however. Why is this? Do we need to consider complex physical hypotheses, or is there a simple statistical explanation?
We compare two methods for making predictions of the climatological distribution of the number of hurricanes making landfall along short sections of the North American coastline. The first method uses local data, and the second method uses a basin-wide track model. Using cross-validation we show that the basin-wide track model gives better predictions for almost all parts of the coastline. This is the first time such a comparison has been made, and is the first rigourous justification for the use of basin-wide track models for predicting hurricane landfall rates and hurricane risk.
The time series of the number of hurricanes per year in the Atlantic basin shows a clear change of level between 1994 and 1995. The time series of the number of hurricanes that make landfall in the US, however, does not show the same obvious change of level. Prima-facie this seems rather surprising, given that the landfalling hurricanes are a subset of the basin hurricanes. We investigate whether it really should be considered surprising or whether there is a simple statistical explanation for the disappearance of this change-point at landfall.