1 Background and Motivation

1.1 Directional data and spherical outcomes

Many forecasting problems produce outcomes that are naturally expressed as directions rather than linear quantities. If an event location is summarized by a bearing, if a weather system is characterized by its propagation direction, or if an object’s attitude is recorded as an orientation, the relevant outcome can be treated as a point on a sphere. In such cases, evaluating a probabilistic forecast requires acknowledging that errors depend on angular separation, not on arithmetic distance in Euclidean space.

1.2 Proper scoring rules in forecast evaluation

A scoring rule maps a forecast and a realized outcome to a numerical score. It is called proper when the expected score is optimized by reporting the true data-generating probability model. Properness provides a theoretical guarantee that truthful probabilistic forecasts are incentivized rather than punished. This framework is widely used to compare probabilistic forecasts across different modeling choices while maintaining consistent incentives.

1.3 Intuition: rewarding angular accuracy

Spherical scores are designed so that forecasts aligning with the realized direction receive higher scores. When uncertainty is represented as a distribution over directions, the scoring rule rewards both “where” the probability mass is placed and “how tightly” it is concentrated near the realized outcome. Importantly, directions that are close on the sphere should be treated as more accurate than directions that are far away, with the scoring sensitivity governed by the selected spherical scoring construction.

2 Mathematical Foundations

2.1 Geometry of the sphere

2.1.1 Surface measures and angles

On the unit sphere \(S^2\), outcomes correspond to unit vectors. The sphere’s geometry is captured by its surface measure (used for integration over directions) and by angular separation, commonly expressed through the central angle between two directions. If \(u\) and \(v\) are unit vectors, their angular distance \(\theta\in[0,\pi]\) satisfies \(\cos(\theta)=u\cdot v\). This angle underlies many score definitions, especially those based on distance or dot products.

2.1.2 Converting between direction and coordinates

Although directions can be represented as unit vectors in \(\mathbb{R}^3\), practical modeling often uses coordinate systems such as longitude–latitude. Any such representation must respect the sphere’s geometry and the induced surface-area element when densities are defined. Coordinate transforms also affect numerical integration and the interpretation of “uniform” distributions, which depend on the surface measure rather than on coordinate grids alone.

2.2 Probability models on the sphere

2.2.1 Spherical distributions overview

Forecasts over directions are modeled using probability distributions supported on the sphere. Common families include distributions that depend on the angle to a mean direction, producing isotropic uncertainty around a preferred orientation. Depending on the application, models may be unimodal (a single concentration region) or multimodal (several plausible directions), and they may accommodate heavier tails to reflect occasional large directional deviations.

2.2.2 Forecast representations: density vs. point prediction

A forecast can be represented either as a full probability density over directions or as a point prediction with an associated uncertainty parameter. In scoring, density-based forecasts integrate naturally with log-likelihood–style constructions, while point forecasts can be evaluated using distance- or dot-product-based losses that effectively impose an implicit uncertainty model through their curvature. In general, spherical scores can compare both deterministic and probabilistic predictions when they are expressed in a compatible form.

2.3 Measurability and integrability requirements

For a spherical score to be mathematically well-defined for probabilistic forecasts, the score as a function of the predicted distribution and the realized direction must be measurable with respect to the relevant sigma-algebra on the sphere. Additionally, expected scores must exist, which typically requires integrability conditions such as finite moments under the forecast and reference distributions. These technical requirements ensure that propriety statements and statistical comparisons are valid rather than purely formal.

3 Definition of Spherical Scores

3.1 General form of a spherical score

A spherical score evaluates a probabilistic forecast for a realized direction \(y\in S^2\). Let \(P\) denote the forecast distribution on the sphere. A spherical score is a functional \(S(P,y)\) that returns a numerical value. The construction ensures dependence on the directional geometry, often through the angle or dot product between \(y\) and a summary of \(P\) such as its mean direction, or through integrals of functions of angular separation.

3.2 Score as a function of predicted vs. observed direction

When the forecast is represented by a predictive density \(p(\cdot)\), many score classes can be written in terms of \(p\) evaluated at the observation \(y\), or in terms of an integral involving the kernel of \(p\). When \(P\) is represented by a point direction \( \hat{y}\) (plus an optional uncertainty notion), scores often simplify to functions \(g(\theta(\hat{y},y))\) where \(\theta\) is the spherical angular distance. The key idea is that the score should be sensitive to angular proximity in a controlled, interpretable way.

3.3 Special cases and common parameterizations

3.3.1 Scores based on angular distance

Some spherical scoring rules are defined directly as decreasing functions of the angular distance \(\theta\) between predicted and observed directions. Examples include quadratic-like penalties that grow with \(\theta^2\) or other monotone transformations that emphasize different error regimes. Such scores can be interpreted as angular loss functions turned into scoring rules by sign changes or affine transformations.

3.3.2 Scores based on dot products

Because \(u\cdot v=\cos(\theta)\) compresses directional information into a single scalar, dot-product parameterizations are often used. Scores based on \(u\cdot y\) allow compact expressions and can handle symmetries naturally. Depending on the chosen transformation, they may treat small-angle errors sharply while saturating for nearly opposite directions, or vice versa.

4 Properness and Consistency

4.1 What it means for a score to be proper

A spherical score \(S(P,y)\) is proper if, for any “true” distribution \(Q\), the forecast that maximizes the expected score under \(Q\) is \(P=Q\). Formally, \( \mathbb{E}_{Y\sim Q}[S(Q,Y)] \ge \mathbb{E}_{Y\sim Q}[S(P,Y)]\) for all candidate forecasts \(P\). This property ensures that probabilistic forecasting is aligned with reality in an incentive-compatible sense.

4.2 Conditions for strict propriety

Strict propriety strengthens the guarantee by making the maximizing forecast unique (up to equivalence classes where the score cannot distinguish them). In practice, strictness depends on how the score’s functional form separates distributions. Many commonly used spherical score families achieve strict propriety when the associated kernel or divergence measure is strictly convex and the underlying distribution is identifiable from the scoring functional.

Even when the forecast family cannot represent the true distribution, proper scores can still yield consistent estimates of the best approximation within the model class. This “best attainable” notion is tied to minimizing an expected scoring discrepancy rather than necessarily recovering the true distribution itself. Under misspecification, spherical scores can therefore support principled model selection and parameter fitting based on predictive accuracy.

4.4 Calibration implications for directional forecasts

Calibration in directional settings concerns whether predicted uncertainty matches observed frequency patterns on the sphere. Proper spherical scores encourage calibration because misaligned confidence (such as being overconfident in the wrong region) typically reduces expected score. While calibration can still fail due to model limitations or data shifts, properness provides a foundation for uncertainty-aware evaluation.

5 Examples and Variants

Quadratic-type spherical losses penalize angular deviations in a manner analogous to mean squared error in Euclidean problems. A typical construction uses a function of \(\theta\) such as \( \theta^2\) or a smooth equivalent in terms of the dot product. These losses are intuitive for applications where small directional errors dominate performance criteria, and they often connect to regression on the sphere via least-squares-like reasoning.

5.2 Log-likelihood–style spherical scoring

When the forecast density \(p\) is available, a natural scoring rule is based on \(\log p(y)\), rewarding forecasts that assign high probability density near the observed direction. This “log score” is a central example of a proper scoring rule under broad conditions. For spherical data, it leverages the geometry only through the density’s support and normalization over the sphere.

5.3 Bregman-divergence–inspired spherical scores

Some spherical scores can be related to divergences that act like generalized distances between probability distributions. In these constructions, the score often resembles a negative divergence evaluated at the observation or averaged with respect to an observation-dependent measure. Bregman-divergence connections provide a systematic way to design proper scoring rules using convex functions tied to the statistical model.

5.4 Scoring rules built from specific spherical kernels

Another variant class uses kernels defined on the sphere, creating scores that depend on how well the predicted distribution aligns with the realized direction through kernel evaluations. Such kernel-based designs can emphasize local accuracy (via narrow kernels) or broader alignment (via wider kernels). Their propriety and robustness properties depend on the kernel’s structure and how it induces an integral operator over spherical distributions.

6 Computation and Implementation

6.1 Numerical evaluation of spherical scores

Computing a spherical score requires evaluating either a density at the observed direction, an angular-distance function, or an integral involving a kernel. Numerical stability is important because spherical coordinate singularities (e.g., near poles) and floating-point limitations can distort distance calculations. Using vector-based representations and careful normalization helps avoid common implementation errors.

6.2 Monte Carlo estimation approaches

When analytic forms are unavailable—especially for integral-based scores—Monte Carlo methods approximate expected quantities by sampling directions from either the forecast distribution or from an auxiliary proposal. Variance reduction techniques such as importance sampling or quasi-Monte Carlo sampling can improve accuracy. For repeated evaluation across many observations, precomputing kernel terms or using vectorized operations may significantly reduce runtime.

6.3 Handling edge cases (antipodal directions, degeneracy)

Directions near antipodal points can produce special behavior, especially for scores using \(\cos(\theta)\) where \(\theta\approx\pi\) yields values near \(-1\). For models with concentration parameters approaching extremes (very sharp or nearly uniform), numerical underflow or division-by-zero risks may arise. Robust implementations typically include safeguards such as clipping arguments to inverse cosine operations, using stable log-density computations, and validating that predicted densities integrate to one.

6.4 Efficiency considerations for repeated scoring

In practice, scoring may occur over many time points or across competing model outputs. Efficient strategies include reusing angular distances, caching normalization constants for parametric distributions, and vectorizing across observations. When comparing multiple models with similar parameters, shared intermediate computations can reduce overhead without compromising correctness.

7 Statistical Properties

7.1 Robustness to noisy directional observations

Directional observations may be corrupted by measurement noise, small unmodeled rotations, or systematic sensor biases. Proper spherical scores often remain informative under noise because they evaluate the full predictive distribution rather than only a deterministic estimate. Robustness depends on how sharply the score penalizes large angular errors, which is governed by the score’s curvature.

7.2 Variance and sensitivity analysis

The variability of a score across samples depends on both the data-generating process and the score’s functional form. Scores with heavy tails in their loss landscape can yield higher variance estimators, affecting confidence in model comparisons. Sensitivity analysis evaluates how much score values change under small perturbations of forecast parameters or observations, which helps anticipate instability in optimization and selection routines.

7.3 Bias considerations in finite samples

When scores are estimated from finite data, empirical averages can be biased relative to their population expectations, particularly for non-linear integral-based scores. Bias can also arise from Monte Carlo approximation error if sampling is used inside scoring. Standard remedies include increasing sample sizes, employing unbiased estimators where possible, and separating approximation error from observational sampling error.

8 Practical Use Cases

8.1 Evaluating forecasts for orientations and headings

In navigation and tracking, forecasts often include a direction representing a heading, drift direction, or target bearing. Spherical scoring provides an evaluation framework that treats slight heading errors as less severe than large misalignments, while still rewarding uncertainty estimates that appropriately spread probability mass. This supports comparisons among filters and predictors that differ in calibration and concentration.

8.2 Directional uncertainty in forecasting workflows

Forecast workflows may require propagating uncertainty through multiple steps, such as when estimating current attitude from sensor readings and projecting it forward. Spherical scores fit naturally into such workflows by measuring the quality of probabilistic outputs at each stage. The scores help distinguish models that are accurate but overconfident from models that are diffuse but realistically cover uncertainty.

8.3 Comparing competing predictive models

Spherical scores enable principled model comparison by providing a scalar summary of predictive quality. Because proper scores reward correct distributional specification, comparisons are less arbitrary than those based solely on point errors. In applications with multiple candidate models—such as different directional distribution families or kernel widths—scores can guide selection using an incentive-compatible criterion.

9 Scoring and Interpretation

9.1 Reading score values and comparing models

Score magnitudes depend on the specific scoring rule and parameterization. For log-like scores, larger values correspond to higher assigned density near observations, while for loss-derived scores, lower values correspond to better performance. Interpretation therefore requires knowing the sign convention and the normalization implicit in the chosen rule.

9.2 Normalization and interpretation across horizons

In time series settings with forecasts at different lead times, raw scores can vary simply because uncertainty grows over the horizon. Normalizing or reporting relative improvements can make comparisons more meaningful. For example, comparing each model against a baseline using differences in expected score can isolate genuine predictive skill from horizon-induced scaling.

9.3 Reporting conventions and visualization ideas

Common reporting practices include averaging scores over test sets and providing uncertainty intervals via bootstrapping. Visualization may include plotting score differences across models, or mapping angular accuracy and calibration diagnostics onto spherical heatmaps when probabilistic forecasts are available. Such approaches help translate numerical scores into interpretable patterns about where models succeed or fail on the sphere.

10.1 Angular loss functions vs. spherical scores

Angular loss functions evaluate prediction quality through penalties based on angular distance but may lack the incentive-compatibility guarantees of proper scoring rules. Spherical scores generalize evaluation by incorporating full probabilistic information in a way that, under properness, encourages truthful uncertainty reporting. The two frameworks are related but not interchangeable.

10.2 Scoring rules for manifolds and generalized spaces

Spherical outcomes are a specific instance of prediction on a manifold. Broader scoring rule theory extends ideas of properness and divergence-based evaluation to other geometries and constraint sets. Understanding these connections clarifies what is “spherical” versus what is general: geometry-specific scoring sensitivity versus universal properties tied to properness.

10.3 Connections to information geometry high-level

At a high level, proper scoring rules are related to divergence measures between probability distributions, which can be studied through information geometry. In this perspective, the curvature of the scoring functional corresponds to how distributions separate under the chosen evaluation criterion. While details vary by construction, the shared theme is that proper scoring induces an information-like notion of discrepancy relevant to estimation and learning.

11 Summary and Further Reading

11.1 Key takeaways

Spherical scores provide an incentive-compatible way to evaluate probabilistic forecasts when outcomes are directional and lie on a sphere. They generalize proper scoring-rule ideas to angular domains by using spherical geometry, angle-based or kernel-based constructions, and distributional models defined over directions. Properness supports calibrated uncertainty reporting and principled model comparison.

11.2 Suggested references and learning paths

A productive reading path typically starts with proper scoring rules in general, then focuses on their extension to spaces like spheres and manifolds. From there, studying examples—such as log scores, distance-based spherical losses, and kernel or divergence constructions—helps connect theory to implementation. Finally, reviewing application notes in navigation, geoscience, or meteorology provides context for how predictive distributions on \(S^2\) are represented and scored in practice.