On reliability analysis of multi-categorical forecasts

Reliability analysis of probabilistic forecasts, in particular through the rank histogram or Talagrand diagram, is revisited. Two shortcomings are pointed out: Firstly, a uniform rank histogram is but a necessary condition for reliability. Secondly, if the forecast is assumed to be reliable, an indication is needed how far a histogram is expected to deviate from uniformity merely due to randomness. Concerning the first shortcoming, it is suggested that forecasts be grouped or stratified along suitable criteria, and that reliability is analyzed individually for each forecast stratum. A reliable forecast should have uniform histograms for all individual forecast strata, not only for all forecasts as a whole. As to the second shortcoming, instead of the observed frequencies, the probability of the observed frequency is plotted, providing and indication of the likelihood of the result under the hypothesis that the forecast is reliable. Furthermore, a Goodness-Of-Fit statistic is discussed which is essentially the reliability term of the Ignorance score. The discussed tools are applied to medium range forecasts for 2 m-temperature anomalies at several locations and lead times. The forecasts are stratified along the expected ranked probability score. Those forecasts which feature a high expected score turn out to be particularly unreliable.

Reliability, sufficiency, and the decomposition of proper scores

Scoring rules are an important tool for evaluating the performance of probabilistic forecasting schemes. A scoring rule is called strictly proper if its expectation is optimal if and only if the forecast probability represents the true distribution of the target. In the binary case, strictly proper scoring rules allow for a decomposition into terms related to the resolution and the reliability of a forecast. This fact is particularly well known for the Brier Score. In this article, this result is extended to forecasts for finite-valued targets. Both resolution and reliability are shown to have a positive effect on the score. It is demonstrated that resolution and reliability are directly related to forecast attributes that are desirable on grounds independent of the notion of scores. This finding can be considered an epistemological justification of measuring forecast quality by proper scoring rules. A link is provided to the original work of DeGroot and Fienberg, extending their concepts of sufficiency and refinement. The relation to the conjectured sharpness principle of Gneiting, et al., is elucidated.

Regularized logistic models for probabilistic forecasting and diagnostics

Logistic models are studied as a tool to convert dynamical forecast information (deterministic and ensemble) into probability forecasts. A logistic model is obtained by setting the logarithmic odds ratio equal to a linear combination of the inputs. As with any statistical model, logistic models will suffer from overfitting if the number of inputs is comparable to the number of forecast instances. Computational approaches to avoid overfitting by regularization are discussed, and efficient techniques for model assessment and selection are presented. A logit version of the lasso (originally a linear regression technique), is discussed. In lasso models, less important inputs are identified and the corresponding coefficient is set to zero, providing an efficient and automatic model reduction procedure. For the same reason, lasso models are particularly appealing for diagnostic purposes.

Probabilistic evaluation of time series models : a comparison of several approaches

Several methods are examined which allow to produce forecasts for time series in the form of probability assignments. The necessary concepts are presented, addressing questions such as how to assess the performance of a probabilistic forecast. A particular class of models, cluster weighted models (CWMs), is given particular attention. CWMs, originally proposed for deterministic forecasts, can be employed for probabilistic forecasting with little modification. Two examples are presented. The first involves estimating the state of (numerically simulated) dynamical systems from noise corrupted measurements, a problem also known as filtering. There is an optimal solution to this problem, called the optimal filter, to which the considered time series models are compared. (The optimal filter requires the dynamical equations to be known.) In the second example, we aim at forecasting the chaotic oscillations of an experimental bronze spring system. Both examples demonstrate that the considered time series models, and especially the CWMs, provide useful probabilistic information about the underlying dynamical relations. In particular, they provide more than just an approximation to the conditional mean.

Predicting outliers in ensemble forecasts

An ensemble forecast is a collection of runs of a numerical dynamical model, initialized with perturbed initial conditions. In modern weather prediction for example, ensembles are used to retrieve probabilistic information about future weather conditions. In this contribution, we are concerned with ensemble forecasts of a scalar quantity (say, the temperature at a specific location). We consider the event that the verification is smaller than the smallest, or larger than the largest ensemble member. We call these events outliers. If a K-member ensemble accurately reflected the variability of the verification, outliers should occur with a base rate of 2/(K + 1). In operational forecast ensembles though, this frequency is often found to be higher. We study the predictability of outliers and find that, exploiting information available from the ensemble, forecast probabilities for outlier events can be calculated which are more skilful than the unconditional base rate. We prove this analytically for statistically consistent forecast ensembles. Further, the analytical results are compared to the predictability of outliers in an operational forecast ensemble by means of model output statistics. We find the analytical and empirical results to agree both qualitatively and quantitatively.

Estimating reliability and resolution of probability forecasts through decomposition of the empirical score

References (20)Cited By (1)Export CitationAboutAbstract Proper scoring rules provide a useful means to evaluate probabilistic forecasts. Independent from scoring rules, it has been argued that reliability and resolution are desirable forecast attributes. The mathematical expectation value of the score allows for a decomposition into reliability and resolution related terms, demonstrating a relationship between scoring rules and reliability/resolution. A similar decomposition holds for the empirical (i.e. sample average) score over an archive of forecast–observation pairs. This empirical decomposition though provides a too optimistic estimate of the potential score (i.e. the optimum score which could be obtained through recalibration), showing that a forecast assessment based solely on the empirical resolution and reliability terms will be misleading. The differences between the theoretical and empirical decomposition are investigated, and specific recommendations are given how to obtain better estimators of reliability and resolution in the case of the Brier and Ignorance scoring rule.

Anthropogenic greenhouse gas contribution to flood risk in England and Wales in autumn 2000

Interest in attributing the risk of damaging weather-related events to anthropogenic climate change is increasing1. Yet climate models used to study the attribution problem typically do not resolve the weather systems associated with damaging events2 such as the UK floods of October and November 2000. Occurring during the wettest autumn in England and Wales since records began in 17663, 4, these floods damaged nearly 10,000 properties across that region, disrupted services severely, and caused insured losses estimated at £1.3 billion (refs 5, 6). Although the flooding was deemed a ‘wake-up call’ to the impacts of climate change at the time7, such claims are typically supported only by general thermodynamic arguments that suggest increased extreme precipitation under global warming, but fail8, 9 to account fully for the complex hydrometeorology4, 10 associated with flooding. Here we present a multi-step, physically based ‘probabilistic event attribution’ framework showing that it is very likely that global anthropogenic greenhouse gas emissions substantially increased the risk of flood occurrence in England and Wales in autumn 2000. Using publicly volunteered distributed computing11, 12, we generate several thousand seasonal-forecast-resolution climate model simulations of autumn 2000 weather, both under realistic conditions, and under conditions as they might have been had these greenhouse gas emissions and the resulting large-scale warming never occurred. Results are fed into a precipitation-runoff model that is used to simulate severe daily river runoff events in England and Wales (proxy indicators of flood events). The precise magnitude of the anthropogenic contribution remains uncertain, but in nine out of ten cases our model results indicate that twentieth-century anthropogenic greenhouse gas emissions increased the risk of floods occurring in England and Wales in autumn 2000 by more than 20%, and in two out of three cases by more than 90%.

Rank histograms of stratified Monte-Carlo ensembles

The application of forecast ensembles to probabilistic weather prediction has spurred considerable interest in their evaluation. Such ensembles are commonly interpreted as Monte Carlo ensembles meaning that the ensemble members are perceived as random draws from a distribution. Under this interpretation, a reasonable property to ask for is statistical consistency, which demands that the ensemble members and the verification behave like draws from the same distribution. A widely used technique to assess statistical consistency of a historical dataset is the rank histogram, which uses as a criterion the number of times that the verification falls between pairs of members of the ordered ensemble. Ensemble evaluation is rendered more specific by stratification, which means that ensembles that satisfy a certain condition (e.g., a certain meteorological regime) are evaluated separately. Fundamental relationships between Monte Carlo ensembles, their rank histograms, and random sampling from the probability simplex according to the Dirichlet distribution are pointed out. Furthermore, the possible benefits and complications of ensemble stratification are discussed. The main conclusion is that a stratified Monte Carlo ensemble might appear inconsistent with the verification even though the original (unstratified) ensemble is consistent. The apparent inconsistency is merely a result of stratification. Stratified rank histograms are thus not necessarily flat. This result is demonstrated by perfect ensemble simulations and supplemented by mathematical arguments. Possible methods to avoid or remove artifacts that stratification induces in the rank histogram are suggested.

The determination of total nitrogen and total phosphorus concentrations in freshwaters from land use, stock headage and population data: testing of a model for use in conservation and water quality management

1. Nutrient concentrations (particularly N and P) determine the extent to which water bodies are or may become eutrophic. Direct determination of nutrient content on a wide scale is labour intensive but the main sources of N and P are well known. This paper describes and tests an export coefficient model for prediction of total N and total P from: (i) land use, stock headage and human population; (ii) the export rates of N and P from these sources; and (iii) the river discharge. Such a model might be used to forecast the effects of changes in land use in the future and to hindcast past water quality to establish comparative or baseline states for the monitoring of change. 2. The model has been calibrated against observed data for 1988 and validated against sets of observed data for a sequence of earlier years in ten British catchments varying from uplands through rolling, fertile lowlands to the flat topography of East Anglia. 3. The model predicted total N and total P concentrations with high precision (95% of the variance in observed data explained). It has been used in two forms: the first on a specific catchment basis; the second for a larger natural region which contains the catchment with the assumption that all catchments within that region will be similar. Both models gave similar results with little loss of precision in the latter case. This implies that it will be possible to describe the overall pattern of nutrient export in the UK with only a fraction of the effort needed to carry out the calculations for each individual water body. 4. Comparison between land use, stock headage, population numbers and nutrient export for the ten catchments in the pre-war year of 1931, and for 1970 and 1988 show that there has been a substantial loss of rough grazing to fertilized temporary and permanent grasslands, an increase in the hectarage devoted to arable, consistent increases in the stocking of cattle and sheep and a marked movement of humans to these rural catchments. 5. All of these trends have increased the flows of nutrients with more than a doubling of both total N and total P loads during the period. On average in these rural catchments, stock wastes have been the greatest contributors to both N and P exports, with cultivation the next most important source of N and people of P. Ratios of N to P were high in 1931 and remain little changed so that, in these catchments, phosphorus continues to be the nutrient most likely to control algal crops in standing waters supplied by the rivers studied.

Modelling the impact of land use change on water quality in agricultural catchments

Export coefficient modelling was used to model the impact of agriculture on nitrogen and phosphorus loading on the surface waters of two contrasting agricultural catchments. The model was originally developed for the Windrush catchment where the highly reactive Jurassic limestone aquifer underlying the catchment is well connected to the surface drainage network, allowing the system to be modelled using uniform export coefficients for each nutrient source in the catchment, regardless of proximity to the surface drainage network. In the Slapton catchment, the hydrological path-ways are dominated by surface and lateral shallow subsurface flow, requiring modification of the export coefficient model to incorporate a distance-decay component in the export coefficients. The modified model was calibrated against observed total nitrogen and total phosphorus loads delivered to Slapton Ley from inflowing streams in its catchment. Sensitivity analysis was conducted to isolate the key controls on nutrient export in the modified model. The model was validated against long-term records of water quality, and was found to be accurate in its predictions and sensitive to both temporal and spatial changes in agricultural practice in the catchment. The model was then used to forecast the potential reduction in nutrient loading on Slapton Ley associated with a range of catchment management strategies. The best practicable environmental option (BPEO) was found to be spatial redistribution of high nutrient export risk sources to areas of the catchment with the greatest intrinsic nutrient retention capacity.

Incorporating technology buying behaviour into UK-based long term domestic stock energy models to provide improved policy analysis

The UK has a target for an 80% reduction in CO2 emissions by 2050 from a 1990 base. Domestic energy use accounts for around 30% of total emissions. This paper presents a comprehensive review of existing models and modelling techniques and indicates how they might be improved by considering individual buying behaviour. Macro (top-down) and micro (bottom-up) models have been reviewed and analysed. It is found that bottom-up models can project technology diffusion due to their higher resolution. The weakness of existing bottom-up models at capturing individual green technology buying behaviour has been identified. Consequently, Markov chains, neural networks and agent-based modelling are proposed as possible methods to incorporate buying behaviour within a domestic energy forecast model. Among the three methods, agent-based models are found to be the most promising, although a successful agent approach requires large amounts of input data. A prototype agent-based model has been developed and tested, which demonstrates the feasibility of an agent approach. This model shows that an agent-based approach is promising as a means to predict the effectiveness of various policy measures.

Real-time multi-model decadal climate predictions

We present the first climate prediction of the coming decade made with multiple models, initialized with prior observations. This prediction accrues from an international activity to exchange decadal predictions in near real-time, in order to assess differences and similarities, provide a consensus view to prevent over-confidence in forecasts from any single model, and establish current collective capability. We stress that the forecast is experimental, since the skill of the multi-model system is as yet unknown. Nevertheless, the forecast systems used here are based on models that have undergone rigorous evaluation and individually have been evaluated for forecast skill. Moreover, it is important to publish forecasts to enable open evaluation, and to provide a focus on climate change in the coming decade. Initialized forecasts of the year 2011 agree well with observations, with a pattern correlation of 0.62 compared to 0.31 for uninitialized projections. In particular, the forecast correctly predicted La Niña in the Pacific, and warm conditions in the north Atlantic and USA. A similar pattern is predicted for 2012 but with a weaker La Niña. Indices of Atlantic multi-decadal variability and Pacific decadal variability show no signal beyond climatology after 2015, while temperature in the Niño3 region is predicted to warm slightly by about 0.5 °C over the coming decade. However, uncertainties are large for individual years and initialization has little impact beyond the first 4 years in most regions. Relative to uninitialized forecasts, initialized forecasts are significantly warmer in the north Atlantic sub-polar gyre and cooler in the north Pacific throughout the decade. They are also significantly cooler in the global average and over most land and ocean regions out to several years ahead. However, in the absence of volcanic eruptions, global temperature is predicted to continue to rise, with each year from 2013 onwards having a 50 % chance of exceeding the current observed record. Verification of these forecasts will provide an important opportunity to test the performance of models and our understanding and knowledge of the drivers of climate change.

Exploring the role of reanalysis data in simulating regional wind generation variability over Northern Ireland

As wind generation increases, system impact studies rely on predictions of future generation and effective representation of wind variability. A well-established approach to investigate the impact of wind variability is to simulate generation using observations from 10 m meteorological mast-data. However, there are problems with relying purely on historical wind-speed records or generation histories: mast-data is often incomplete, not sited at a relevant wind generation sites, and recorded at the wrong altitude above ground (usually 10 m), each of which may distort the generation profile. A possible complimentary approach is to use reanalysis data, where data assimilation techniques are combined with state-of-the-art weather forecast models to produce complete gridded wind time-series over an area. Previous investigations of reanalysis datasets have placed an emphasis on comparing reanalysis to meteorological site records whereas this paper compares wind generation simulated using reanalysis data directly against historic wind generation records. Importantly, this comparison is conducted using raw reanalysis data (typical resolution ∼50 km), without relying on a computationally expensive “dynamical downscaling” for a particular target region. Although the raw reanalysis data cannot, by nature of its construction, represent the site-specific effects of sub-gridscale topography, it is nevertheless shown to be comparable to or better than the mast-based simulation in the region considered and it is therefore argued that raw reanalysis data may offer a number of significant advantages as a data source.

Characteristics of high-resolution versions of the Met Office unified model for forecasting convection over the United Kingdom

With many operational centers moving toward order 1-km-gridlength models for routine weather forecasting, this paper presents a systematic investigation of the properties of high-resolution versions of the Met Office Unified Model for short-range forecasting of convective rainfall events. The authors describe a suite of configurations of the Met Office Unified Model running with grid lengths of 12, 4, and 1 km and analyze results from these models for a number of convective cases from the summers of 2003, 2004, and 2005. The analysis includes subjective evaluation of the rainfall fields and comparisons of rainfall amounts, initiation, cell statistics, and a scale-selective verification technique. It is shown that the 4- and 1-km-gridlength models often give more realistic-looking precipitation fields because convection is represented explicitly rather than parameterized. However, the 4-km model representation suffers from large convective cells and delayed initiation because the grid length is too long to correctly reproduce the convection explicitly. These problems are not as evident in the 1-km model, although it does suffer from too numerous small cells in some situations. Both the 4- and 1-km models suffer from poor representation at the start of the forecast in the period when the high-resolution detail is spinning up from the lower-resolution (12 km) starting data used. A scale-selective precipitation verification technique implies that for later times in the forecasts (after the spinup period) the 1-km model performs better than the 12- and 4-km models for lower rainfall thresholds. For higher thresholds the 4-km model scores almost as well as the 1-km model, and both do better than the 12-km model.