7.4 Gini Index, Lorenz Curves & ROC Curves

Discrimination vs. Calibration in Actuarial Predictive Modeling

When evaluating a property and casualty (P&C) predictive model—whether a frequency-severity Generalized Linear Model (GLM), a Tweedie compound Poisson-Gamma regression, or a logistic large-loss model—actuaries must assess model performance across two distinct, complementary dimensions:

1. Discrimination (Risk Separation / Rank Ordering): The model's ability to differentiate low-risk policyholders from high-risk policyholders. A highly discriminatory model successfully concentrates actual incurred losses into the highest predicted risk percentiles.
2. Calibration (Scale / Absolute Dollar Adequacy): The model's ability to predict the true expected dollar loss cost. A well-calibrated model produces fitted values whose aggregate sum matches observed losses both across the entire portfolio and within specific rating segments.

In competitive insurance markets, discrimination is the primary weapon against adverse selection. If a competitor segments risks more granularly than your company, they will discount low-risk insureds and leave you with overpriced, unprofitable risks. However, evaluating discrimination requires specialized statistical metrics that are sensitive to policy exposure and severe claim skewness.

---

The Lorenz Curve for Insurance Pure Premium

The Lorenz curve, originally introduced by Max Lorenz in 1905 to depict economic income inequality, was adapted to casualty actuarial science by Frees, Meyers, and Cummings (2011, 2014). In economics, the curve plots the cumulative share of population against the cumulative share of income. In P&C ratemaking, the Lorenz curve assesses how effectively a predictive model concentrates actual claim dollars by ordering policies according to predicted pure premium.

                                  LORENZ CURVE CONSTRUCTION

    Cumulative % of
    Actual Losses (y)
        100% ┼───────────────────────────────────────────────────────────────┐
             │                                                            . ╱│
             │                                                         . ╱   │
             │                                                      . ╱      │
             │                                                   . ╱         │
             │                                                . ╱            │
             │                                             . ╱    Area A     │
             │                       45° Line of        . ╱  (Between Line   │
             │                     Perfect Equality  . ╱        & Curve)     │
             │                         (Gini = 0) . ╱                        │
             │                                 . ╱                           │
             │                              . ╱     Empirical Lorenz Curve   │
             │                           . ╱       L(u) [Sorted by mu_hat]   │
             │                        . ╱          ───────────────────────── │
             │                     . ╱             ╱                         │
             │                  . ╱               ╱     Area B               │
             │               . ╱                 ╱    (Under Empirical       │
             │            . ╱                   ╱          Curve)            │
             │         . ╱                     ╱                             │
             │      . ╱                       ╱                              │
             │   . ╱                         ╱                               │
          0% ┼──┴───────────────────────────┴────────────────────────────────┤
             0%                                                            100%
                           Cumulative % of Earned Exposure (u)

The 3-Step Construction Protocol

To construct an insurance Lorenz curve from a validation dataset containing $n$ policies, follow this three-step protocol:

Step 1: Sort Policies by Model-Predicted Pure Premium

Order all $n$ policies in strictly ascending order of their model-predicted pure premium (or predicted loss cost per unit of exposure), $\hat{\mu}_i$:

$$\hat{\mu}_{(1)} \le \hat{\mu}_{(2)} \le \dots \le \hat{\mu}_{(n)}$$

Where $(i)$ represents the index of the policy in the ordered sequence. Let $e_{(i)}$ denote the earned exposure (e.g., car-years, house-years) and $y_{(i)}$ denote the actual observed aggregate loss dollars incurred by policy $(i)$.

Step 2: Plot Cumulative Percentage of Earned Exposure on the X-Axis

For any policy cutoff $t \in \{1, 2, \dots, n\}$, calculate the cumulative fraction of total portfolio earned exposure:

$$\nu(t) = \frac{\sum_{i=1}^t e_{(i)}}{\sum_{i=1}^n e_{(i)}} = \frac{\sum_{i=1}^t e_{(i)}}{E_{\text{total}}}, \quad u(t) \in [0, 1]$$

The horizontal axis represents $u$, spanning from $0\%$ to $100\%$ of earned exposure.

Step 3: Plot Cumulative Percentage of Actual Losses on the Y-Axis

For the same policy cutoff $t$, calculate the cumulative fraction of total actual observed losses:

$$L(u(t)) = \frac{\sum_{i=1}^t y_{(i)}}{\sum_{i=1}^n y_{(i)}} = \frac{\sum_{i=1}^t y_{(i)}}{Y_{\text{total}}}, \quad L(u(t)) \in [0, 1]$$

The vertical axis represents $L(u)$, spanning from $0\%$ to $100\%$ of actual incurred loss dollars.

Interpreting the Geometry of the Lorenz Curve

• The 45-Degree Diagonal Line of Perfect Equality ($L(u) = u$): If a pricing model has zero discriminatory capability—such as assigning the exact same flat average loss cost to every policyholder or predicting random numbers—the policies are ordered randomly with respect to risk. Under this uninformative baseline, the bottom $20\%$ of exposure accounts for $20\%$ of losses, the bottom $50\%$ of exposure accounts for $50\%$ of losses, and $L(u) = u$ across the entire domain.
• The Empirical Lorenz Curve ($L(u)$): Because the policies are sorted from lowest predicted risk to highest, an effective model places policyholders with near-zero claim likelihood at the beginning. Consequently, the empirical curve initially hugs the horizontal axis (e.g., the bottom $40\%$ of exposure might incur only $6\%$ of actual losses). As the curve accumulates higher-risk policies, its slope increases monotonically, rising steeply toward $(1, 1)$.
• Curvature and Model Strength: The greater the downward sag (bowing) of the Lorenz curve away from the 45-degree diagonal, the superior the model's ability to separate low-risk from high-risk insureds.

---

Derivation of the Gini Coefficient

The Gini coefficient ($G$) summarizes the discriminatory strength of the Lorenz curve into a single scalar statistic between $0.0$ and $1.0$. Geometrically, it is defined as the ratio of the area between the 45-degree diagonal line of equality and the empirical Lorenz curve (Area $A$) to the total area under the 45-degree diagonal:

$$G = \frac{\text{Area}(A)}{\text{Area under Diagonal}}$$

Because the diagonal bisects the unit square with vertices $(0,0), (1,0), (1,1), (0,1)$, the total area under the diagonal is:

$$\text{Area under Diagonal} = \frac{1}{2} \times 1 \times 1 = 0.5$$

Substituting this into the definition:

$$G = \frac{\text{Area}(A)}{0.5} = 2 \times \text{Area}(A)$$

Let $\text{Area}(B)$ denote the area strictly underneath the empirical Lorenz curve $L(u)$ from $u = 0$ to $u = 1$:

$$\text{Area}(B) = \int_0^1 L(u) \, du$$

Because the sum of Area $A$ and Area $B$ equals the area of the lower triangle:

$$\text{Area}(A) + \text{Area}(B) = 0.5 \implies \text{Area}(A) = 0.5 - \text{Area}(B)$$

Substituting $\text{Area}(A)$ into the Gini formula gives the classical actuarial derivation:

$$G = 2 \left( 0.5 - \text{Area}(B) \right) = 1 - 2 \times \text{Area}(B) = 1 - 2 \int_0^1 L(u) \, du$$

Discrete Calculation via Trapezoidal Rule

In practical actuarial applications, policyholder data is evaluated across $m$ discrete ordered observations (or sorted quantiles, such as percentiles or ventiles). Let $u_0 = 0, u_1, u_2, \dots, u_m = 1$ be the cumulative exposure points, with corresponding cumulative actual loss fractions $L(u_0) = 0, L(u_1), \dots, L(u_m) = 1$.

Using the trapezoidal approximation for numerical integration:

$$\text{Area}(B) \approx \sum_{k=1}^m \frac{L(u_{k-1}) + L(u_k)}{2} \times (u_k - u_{k-1})$$

Substituting $\text{Area}(B)$ into the Gini equation:

$$G = 1 - \sum_{k=1}^m \left[ L(u_{k-1}) + L(u_k) \right] (u_k - u_{k-1})$$

When evaluating over $m$ equal-exposure bins (such as 10 deciles where $u_k - u_{k-1} = \frac{1}{m} = 0.10$):

$$G = 1 - \frac{1}{m} \sum_{k=1}^m \left[ L(u_{k-1}) + L(u_k) \right]$$

Benchmark Gini Ranges in P&C Insurance

---

The Normalized Gini: Benchmarking Against Inherent Process Risk

A critical conceptual trap on the CAS PCPA exam is assuming that an ideal insurance pricing model should achieve a Gini coefficient near $1.0$.

In economics, if one individual earned all income in an economy, the Lorenz curve would follow the bottom axis and right vertical edge, yielding $G = 1.0$. However, in property and casualty insurance, claims are stochastic realizations governed by severe process risk (pure claim randomness).

The Problem of Inherent Process Risk

Consider an omniscient actuary who knows the exact hypothetical true loss cost $\mathbb{E}[Y_i \mid \mathbf{X}_i]$ for every policyholder in the portfolio (the theoretical "Bayes optimal" or "Oracle" model):

• A high-risk commercial driver with a $30\%$ annual claim probability still has a $70\%$ probability of experiencing zero claims in any single policy year.
• A pristine driver with a $1.5\%$ annual claim probability can still suffer a catastrophic multi-car collision due to severe icy weather.

Because actual observed losses $y_i$ contain irreducible random process noise, even the true generating model cannot order individual realized claims perfectly. Realized losses will never align monotonically with expected risk, which bounds the empirical Gini well below $1.0$.

Constructing the Oracle (Maximum) Lorenz Curve

To establish the realistic upper ceiling of discrimination for a specific portfolio, actuaries calculate the theoretical maximum Gini, denoted $G_{\max}$:

1. Sort the validation policies in ascending order of actual observed losses, $y_i$ (or actual empirical pure premium $y_i / e_i$).
2. All policyholders with zero claims ($y_i = 0$) are placed first. These zero-claim policies account for a massive fraction of exposure (often $85\%$ to $95\%$ in personal auto).
3. Policyholders with claims are then sorted from smallest dollar severity to largest dollar severity.
4. Calculate the resulting empirical Lorenz curve, $L_{\max}(u)$, and its associated Gini coefficient, $G_{\max}$.

The curve $L_{\max}(u)$ remains flat along the horizontal axis ($L_{\max}(u) = 0$) until all zero-claim exposures are exhausted, after which it rises steeply to $1.0$ as the shock losses are accumulated.

                               MAXIMUM GINI (G_max) VS. MODEL GINI

    Cumulative % of
    Actual Losses (y)
        100% ┼───────────────────────────────────────────────────────────────┐
             │                                                         . ╱  ││
             │                                                      . ╱     ││
             │                                                   . ╱        ││
             │                                                . ╱          .││
             │                                             . ╱           .  ││
             │                                          . ╱            .    ││
             │                                       . ╱    Model L(u)      ││
             │                                    . ╱     ...........       ││
             │                                 . ╱     ..                   ││
             │                              . ╱     ..                      ││
             │                           . ╱     ..                         ││
             │                        . ╱     ..                            ││
             │                     . ╱     ..                               ││
             │                  . ╱    ..                                   ││
             │               . ╱   ..                                       ││
             │            . ╱  ..                                           ││
             │         . ╱..                                                ││
             │      . ╱..                                                   ││
             │   . ╱.                                                       ││
          0% ┼──┴───────────────────────────────────────────────────────────┴┤
             0%                                          88%               100%
                           Cumulative % of Earned Exposure (u)
             ├─────────────────── Zero-Loss Policies ───────────────────────┤▲ Loss-Incurring
             │                    (L_max hugs x-axis = 0)                   │  Policies (L_max
                                                                               rockets to 1)

The Normalized Gini Formula

The Normalized Gini coefficient (also known as the Gini index ratio) scales the model's achieved discrimination relative to this empirical maximum:

$$G_{\text{norm}} = \frac{G_{\text{model}}}{G_{\max}}$$

Actuarial Interpretation

• If a commercial auto collision model achieves $G_{\text{model}} = 0.28$, an executive might mistakenly judge the model as ineffective because $0.28 \ll 1.0$.
• However, if the portfolio's maximum achievable Gini given claim frequency and process variance is $G_{\max} = 0.56$, the normalized Gini is: $$G_{\text{norm}} = \frac{0.28}{0.56} = 0.50 \quad (50\%)$$
• This demonstrates that the predictive model successfully captures $50\%$ of all theoretically extractable risk discrimination present in the empirical data.

Receiver Operating Characteristic (ROC) & AUC for Claim Occurrence

While the Lorenz curve and Gini coefficient evaluate continuous pure premium or aggregate loss models, casualty actuaries frequently build classification models to predict binary risk events. The primary application is claim occurrence:

$$Y_i = \begin{cases} 1 & \text{if policy } i \text{ incurs } \ge 1 \text{ claim during the policy term} \\ 0 & \text{if policy } i \text{ incurs zero claims} \end{cases}$$

Binary claim probability models (such as logistic regression or probit GLMs) output a predicted claim occurrence probability $\hat{p}_i = P(Y_i = 1 \mid \mathbf{x}_i) \in [0, 1]$.

Confusion Matrix Across Decision Thresholds

Converting a continuous probability prediction $\hat{p}_i$ into a binary prediction requires selecting a decision threshold $c \in [0, 1]$: predict $\hat{Y}_i = 1$ if $\hat{p}_i \ge c$, and $\hat{Y}_i = 0$ if $\hat{p}_i < c$.

From this 2x2 matrix, two core operating rates are defined as functions of the cutoff $c$:

1. True Positive Rate (TPR / Sensitivity / Hit Rate): $$\text{TPR}(c) = \frac{\text{TP}(c)}{\text{TP}(c) + \text{FN}(c)} = \frac{\text{TP}(c)}{P} = P(\hat{p}_i \ge c \mid Y_i = 1)$$
2. False Positive Rate (FPR / $1 - \text{Specificity}$ / False Alarm Rate): $$\text{FPR}(c) = \frac{\text{FP}(c)}{\text{FP}(c) + \text{TN}(c)} = \frac{\text{FP}(c)}{N} = P(\hat{p}_i \ge c \mid Y_i = 0)$$

Constructing the ROC Curve

The Receiver Operating Characteristic (ROC) curve plots $\text{TPR}(c)$ on the y-axis against $\text{FPR}(c)$ on the x-axis for all possible classification thresholds $c \in [0, 1]$:

• When $c = 1.0$: $\text{TPR} = 0.0$ and $\text{FPR} = 0.0$ (point at $(0, 0)$).
• When $c = 0.0$: $\text{TPR} = 1.0$ and $\text{FPR} = 1.0$ (point at $(1, 1)$).
• As $c$ sweeps continuously from $1$ down to $0$, the ROC curve traces a path from $(0,0)$ to $(1,1)$.

                             RECEIVER OPERATING CHARACTERISTIC (ROC) CURVE

    True Positive Rate (TPR)
    [Sensitivity]
        1.0 ┼───────────────────────────────────────────────────────────────┐
            │                                                         . ╱   │
            │                                                 . . . ╱       │
            │                                           . . ╱               │
            │                                       . ╱                     │
            │                                   . ╱                         │
            │                                . ╱                            │
            │                            . ╱                                │
            │                         . ╱    Empirical Model ROC            │
            │                     . ╱        (AUC = 0.76)                   │
            │                 . ╱                                           │
            │             . ╱                                               │
            │         . ╱                                                   │
            │     . ╱              45° Random Guess Baseline                │
            │ . ╱                  (AUC = 0.50)                             │
        0.0 ┼───────────────────────────────────────────────────────────────┤
            0.0                                                           1.0
                              False Positive Rate (FPR) [1 - Specificity]

Area Under the ROC Curve (AUC) & Concordance Statistic Equivalence

The overall discriminatory power across all possible cutoffs is quantified by the Area Under the ROC Curve (AUC):

$$\text{AUC} = \int_0^1 \text{TPR}(\text{FPR}) \, d(\text{FPR})$$

• $\text{AUC} = 0.50$: Represents a completely uninformative model (the diagonal line). Predictions are no better than flipping an unbiased coin.
• $\text{AUC} = 1.00$: Perfect separation; all claim policies are assigned higher predicted probabilities than all non-claim policies.

The Concordance Statistic ($C$-statistic)

A fundamental theorem of mathematical statistics establishes that the AUC is identically equal to the Concordance statistic ($C$):

$$\text{AUC} = C = P\left( \hat{p}_i > \hat{p}_j \;\middle|\; Y_i = 1, Y_j = 0 \right) + \frac{1}{2} P\left( \hat{p}_i = \hat{p}_j \;\middle|\; Y_i = 1, Y_j = 0 \right)$$

In words: If an actuary randomly selects one policyholder who experienced a claim ($Y_i = 1$) and one policyholder who experienced zero claims ($Y_j = 0$), the AUC is the exact mathematical probability that the model assigns a higher predicted probability of loss to the claim-incurring policyholder.

Mathematical Equivalence: Binary Gini vs. AUC

In binary classification settings, the Gini coefficient derived from the ROC curve (often termed the Accuracy Ratio) maps directly to the AUC through a linear transformation:

$$G = 2 \times \text{AUC} - 1$$ $$\text{AUC} = \frac{G + 1}{2}$$

Verification of Boundary Values:

• If $\text{AUC} = 0.50$ (random guessing): $G = 2(0.50) - 1 = 0.0$ (zero discrimination).
• If $\text{AUC} = 1.00$ (perfect separation): $G = 2(1.00) - 1 = 1.0$ (perfect discrimination).
• If a personal auto frequency model achieves $\text{AUC} = 0.72$, its equivalent binary Gini coefficient is $G = 2(0.72) - 1 = 0.44$.

---

The Fundamental Distinction: Discrimination vs. Calibration

The single most critical actuarial insight regarding Gini and AUC is that they evaluate relative risk ordering, NOT rate level calibration, dollar scale, or profit adequacy.

Mathematical Proof of Scale Invariance

Let $\hat{y} = [\hat{y}_1, \hat{y}_2, \dots, \hat{y}_n]^T$ denote a vector of model loss cost predictions. Consider any strictly increasing monotonic transformation $g(t)$ such that $t_1 > t_2 \implies g(t_1) > g(t_2)$. The most obvious example is scalar multiplication by a positive constant $k > 0$:

$$g(\hat{y}_i) = k \cdot \hat{y}_i, \quad k > 0$$

Because $g(t)$ is strictly monotonic:

$$\hat{y}_{(1)} \le \hat{y}_{(2)} \le \dots \le \hat{y}_{(n)} \iff g(\hat{y}_{(1)}) \le g(\hat{y}_{(2)}) \le \dots \le g(\hat{y}_{(n)})$$

The policy ordering $(1), (2), \dots, (n)$ is 100% identical under both sets of predictions. Consequently:

1. The cumulative exposure coordinates $u(t)$ are unchanged.
2. The cumulative actual loss coordinates $L(u(t))$ are unchanged.
3. The empirical Lorenz curve is mathematically identical.
4. The Gini coefficient is mathematically identical: $G(k \cdot \hat{\mathbf{y}}) \equiv G(\hat{\mathbf{y}})$.
5. For classification, ROC thresholds simply scale by $k$, leaving the ROC curve and AUC identical: $\text{AUC}(k \cdot \hat{\mathbf{p}}) \equiv \text{AUC}(\hat{\mathbf{p}})$.

                    THE ILLUSION OF HIGH DISCRIMINATION WITHOUT CALIBRATION

    Policy Index (Sorted) :    Policy 1     Policy 2     Policy 3     Policy 4
    Actual Loss Cost      :     $100         $200         $500        $1,000
    ──────────────────────────────────────────────────────────────────────────
    Model A (Well-Calibrated):  $105         $195         $510         $990   --> Gini = 0.38, A/E = 1.00
    Model B (10x Scaled)    :  $1,050       $1,950       $5,100       $9,900  --> Gini = 0.38, A/E = 0.10
    Model C (0.1x Scaled)   :   $10.50       $19.50       $51.00       $99.00 --> Gini = 0.38, A/E = 10.0
    ──────────────────────────────────────────────────────────────────────────
    Conclusion: All three models produce IDENTICAL Lorenz curves and IDENTICAL Gini coefficients,
    yet Model B creates catastrophic commercial overpricing and Model C guarantees corporate insolvency!

Real-World Actuarial Consequences

An actuarial team that relies exclusively on Gini or AUC when reviewing a rate filing risks severe financial damage:

• The Overpricing Blunder (Under-calibration): If an algorithm multiplies all loss costs by $2.0$, its Gini coefficient remains stellar. However, when deployed, the company charges double the required premium. Close ratios plummet, high-retention policies lapse, and the company is driven out of the market.
• The Insolvency Blunder (Over-calibration): If an algorithm divides all loss costs by $2.0$, Gini remains equally high. The company experiences explosive sales growth because its rates are half the market level. However, actual losses will be double the collected premium, producing a catastrophic combined ratio exceeding $180\%$ and rapid surplus exhaustion.

Discrimination metrics confirm that you have correctly identified who is riskier than whom. Calibration diagnostics (Section 7.1) confirm that you are charging the right amount of money.

---

Comparison of Discrimination Metrics

---

Actuarial Exam Traps & Pitfalls