Portfolio Analysis & Risk Modeling
Saying yes or no to individual risks is only half the underwriting job. The other half is managing the portfolio -- the entire collection of policies the insurer has on its books. A portfolio loaded with coastal Florida homes might contain perfectly underwritten individual risks, but one hurricane could wipe out the entire book. This part covers how insurers analyze their portfolios, model catastrophe scenarios, use predictive analytics for pricing, and navigate the increasingly challenging personal lines environment.
Exam Alert!
Expect questions about geographic diversification, the three components of catastrophe models, the difference between GLMs and GBMs, the concept of proxy discrimination, and what happens when rate adequacy breaks down. Current market challenges (insurer withdrawals, FAIR Plans, climate change) are heavily tested.
Start Here: 5 Things You MUST Know
A portfolio of individually well-underwritten risks can still produce catastrophic losses if there is geographic concentration -- diversification across regions is essential.
Catastrophe models have three components: Hazard (what events could happen), Vulnerability (how much damage), and Financial (what it costs the insurer).
GLMs (Generalized Linear Models) are the industry standard for personal lines pricing. They produce relativities that show how each rating variable affects expected losses.
Proxy discrimination occurs when a facially neutral variable (like ZIP code) correlates with a protected class (like race) -- regulators are cracking down on this.
The actuarial indication is the recommended rate change. The selected change is what management actually files -- often less than the indication for competitive reasons.
1. Why Portfolio Management Matters
The Portfolio Mindset
An underwriter's job is not just saying yes or no to individual risks. The portfolio -- the entire collection of policies on the insurer's books -- is what determines overall profitability. A perfectly underwritten individual risk can still contribute to a bad portfolio if there is concentration risk (too many similar exposures in one area).
Individual Risk View
- -- Is this one home properly rated?
- -- Does this one driver meet guidelines?
- -- Is the premium adequate for this risk?
- -- Does this risk meet underwriting standards?
Limitation: Each risk looks fine on its own, but you cannot see the big picture.
Portfolio View
- -- How concentrated are we in hurricane zones?
- -- Which segments are profitable vs. losing money?
- -- What is our worst-case loss in a single event?
- -- Are we growing in the right places?
Advantage: Reveals concentration risk, identifies profitable growth opportunities.
Real-World Scenario: Concentration Risk Disaster
The Setup: Sunshine Insurance writes 50,000 homeowners policies in coastal Florida. Every single policy was properly underwritten -- good construction, up-to-date roofs, proper wind mitigation. Individually, each risk is solid.
What Happens: A Category 4 hurricane makes landfall, damaging 30,000 of those 50,000 homes. Average claim: $120,000.
The Result: Total losses: $3.6 billion from a single event. Sunshine's surplus is wiped out. The company becomes insolvent -- not because of bad underwriting on any one risk, but because the portfolio was dangerously concentrated in one catastrophe-prone region. This is exactly why portfolio-level thinking is essential.
2. Geographic Diversification
Geographic diversification means spreading risk across different regions so that no single catastrophe can devastate the entire portfolio. The goal: if a hurricane hits Florida, the insurer's Midwest and Northeast books keep the company profitable overall.
U.S. Catastrophe Exposure by Region
Hurricanes
FL, TX coast, LA, MS, AL, SC, NC. Wind + storm surge + flooding. Season: June-November.
Earthquakes
CA, OR, WA. San Andreas, Cascadia faults. Also New Madrid (MO/TN/AR). No seasonal pattern.
Wildfire
CA, CO, OR, WA, MT. Wildland-Urban Interface (WUI) expansion. Year-round in some areas.
Tornadoes
TX, OK, KS, NE, IA, IL, IN. "Tornado Alley" + expanding "Dixie Alley." Season: March-June.
Hail
TX, CO, NE, KS, SD. Convective storms produce baseball-sized hail. #1 cause of homeowners roof claims.
Flooding
All 50 states. Inland flooding from heavy rain now rivals coastal flooding. NFIP covers most flood risk.
Tools for Managing Geographic Concentration
Geographic Mapping
Plot all policies on a map to visualize concentration. Cluster analysis reveals danger zones.
PML Analysis
Probable Maximum Loss -- the worst-case loss from a single event at a given probability level.
Aggregate Limits by ZIP Code
Cap total insured value in any single ZIP code. Once the cap is hit, stop writing new business there.
New Business Moratoriums
Temporarily stop writing in over-concentrated areas. State Farm and Allstate used this in California.
What Happens When Insurers Leave: FAIR Plans
When private insurers withdraw from high-cat areas, homeowners turn to state FAIR Plans (Fair Access to Insurance Requirements). These are residual market plans that provide basic coverage when the private market will not. FAIR Plans are growing rapidly in California, Florida, and Louisiana as private insurers reduce their exposure.
Real-World Scenario: Insurer Withdrawal
The Setup: State Farm is the largest homeowners insurer in California. After years of wildfire losses (Camp Fire 2018, multiple events 2020-2023) and difficulty getting rate increases approved by regulators, the company announces in May 2023 that it will stop accepting new homeowners applications in California.
What Happens: Allstate, USAA, and other insurers also reduce or exit the California market. Homeowners who cannot find private coverage turn to the California FAIR Plan, which grows from 270,000 policies to over 400,000.
The Result: The FAIR Plan now has over $300 billion in exposure -- a massive concentration of risk in the state's insurer of last resort. This is a textbook example of what happens when the private market cannot achieve rate adequacy due to regulatory constraints, forcing geographic concentration into the residual market.
3. Catastrophe Modeling
Catastrophe models are computer simulations that generate thousands of possible catastrophe scenarios to estimate potential losses. Unlike traditional actuarial analysis (which relies on historical loss data), cat models simulate events that haven't happened yet but could. This is critical because major catastrophes are rare -- you cannot wait for enough data points to price them historically.
The Three Components of a Cat Model
Hazard Module
Question: What events could happen?
- -- Hurricane category, path, and landfall location
- -- Earthquake magnitude, fault line, depth
- -- Wildfire ignition points, spread patterns
- -- Frequency: how often does each scenario occur?
Data sources: NOAA, USGS, historical records, climate science
Vulnerability Module
Question: How much damage would each event cause?
- -- Construction type (wood frame vs. masonry vs. steel)
- -- Age and condition of structures
- -- Building code compliance
- -- Elevation, proximity to coast/fault lines
Data sources: Policy records, building databases, inspection data
Financial Module
Question: What would each event cost the insurer?
- -- Policy terms (limits, deductibles, sub-limits)
- -- Reinsurance recoveries
- -- Demand surge (post-cat price inflation)
- -- Net cost to the insurer after all offsets
Output: Expected loss, PML at various return periods (100-yr, 250-yr)
What Cat Models Are Used For
Pricing
Setting adequate catastrophe loads in premiums. Without cat models, premiums would be guesses.
Reinsurance Purchasing
Determining how much reinsurance to buy and at what attachment point.
Portfolio Optimization
Deciding where to grow and where to limit new business based on concentration risk.
Regulatory Compliance
Demonstrating solvency to regulators. Many states require cat model results in rate filings.
Major Cat Modeling Firms
AIR Worldwide
(Verisk)
CoreLogic
(formerly RMS)
Moody's RMS
(Risk Management Solutions)
Climate Change is Breaking the Models
Cat models traditionally rely on historical data to predict future events. But climate change is making historical patterns unreliable. Wildfire seasons are longer. Convective storms (hail/wind) are more frequent. Hurricanes are intensifying faster. This means cat models may underestimate future losses if they rely too heavily on the past.
Real-World Scenario: Cat Model in Action
The Setup: Coastal Mutual runs its Florida homeowners portfolio through AIR Worldwide's hurricane model. The model simulates 100,000 possible hurricane seasons. The results show a 1-in-100-year event (the "100-year PML") would cost the company $800 million. Coastal Mutual has $500 million in surplus.
What Happens: The $800M PML exceeds surplus by $300M. Without reinsurance, a major hurricane could make Coastal Mutual insolvent. The company buys a catastrophe excess-of-loss reinsurance treaty: it retains the first $200M in losses and reinsurers cover the next $600M.
The Result: Now even the 100-year event only costs Coastal Mutual $200M out of its $500M surplus. The cat model directly informed both the reinsurance purchase decision and the maximum retention level. Without the model, Coastal Mutual would have been flying blind.
4. Loss Ratio Analysis by Segment
Looking at the overall portfolio loss ratio tells you if the book is profitable. But segment-level analysis tells you where the profit and loss is coming from. You break the portfolio into slices and analyze each one separately to find the winners and losers.
| Segment By | What It Reveals | Example Finding |
|---|---|---|
| Territory (ZIP/state/region) | Geographic hot spots of profit or loss | ZIP codes 33101-33199 (Miami) have 85% loss ratio vs. 55% statewide |
| Tier (preferred/standard/nonstandard) | Which risk quality segments perform | Preferred auto: 48% LR. Standard: 62%. Nonstandard: 91%. |
| Product (auto/home/umbrella) | Which lines are making vs. losing money | Auto: 68% LR. Home: 78% LR. Umbrella: 35% LR. |
| Vintage (time on books) | Are newer or older policies performing better? | Policies 5+ years: 52% LR. New business (year 1): 75% LR. |
Actions Based on Segment Analysis
Unprofitable Segments
- -- File for rate increases
- -- Tighten underwriting guidelines
- -- Increase deductibles
- -- Non-renew worst-performing risks
- -- Stop writing new business in that area
Profitable Segments
- -- Grow aggressively (more marketing)
- -- Competitive pricing to gain market share
- -- Relax underwriting slightly to capture more
- -- Offer multi-policy discounts to retain
- -- Invest in agent relationships in that territory
Real-World Scenario: Segment Analysis Reveals Hidden Problem
The Setup: Atlantic Auto's overall auto portfolio has a 65% loss ratio, which looks fine. But the underwriting manager runs a segment analysis by tier and territory.
What Happens: The analysis reveals that the preferred tier in suburban areas has a 42% loss ratio (very profitable), while the nonstandard tier in urban areas has a 105% loss ratio (losing money on every policy). The profitable segment is masking the unprofitable one in the overall average.
The Result: Atlantic Auto files for a 15% rate increase in the nonstandard urban tier, tightens underwriting (no more drivers with 3+ at-fault accidents), and redirects marketing spend to grow the profitable suburban preferred segment. Within 18 months, the overall loss ratio drops to 58%.
5. Predictive Modeling in Personal Lines
Predictive models use statistical and machine learning techniques to analyze the relationship between rating variables (age, location, credit score, vehicle type) and expected losses. The goal: price each risk as accurately as possible so that low-risk customers pay less and high-risk customers pay more.
Predictive Modeling Techniques Compared
| Technique | How It Works | Strengths | Limitations | Industry Use |
|---|---|---|---|---|
| GLMs (Generalized Linear Models) | Mathematical formula linking rating variables to expected losses. Produces "relativities" for each variable. | Transparent, explainable, accepted by regulators. Easy to audit. | Assumes linear relationships. May miss complex interactions between variables. | Industry standard for pricing. Used by nearly every personal lines insurer. |
| GBMs (Gradient Boosted Machines) | Machine learning algorithm that builds many decision trees, each correcting the errors of the previous one. | Captures complex, non-linear patterns. Often more accurate than GLMs. | "Black box" -- harder to explain to regulators. Risk of overfitting. | Used for risk segmentation, underwriting triage, claims fraud detection. |
| Neural Networks | Layers of interconnected nodes that learn patterns from data, inspired by the human brain. | Excellent at finding hidden patterns in massive datasets. Handles unstructured data (images, text). | Least transparent. Requires huge data. Regulators skeptical for pricing. | Emerging use: aerial imagery for roof condition, telematics driving scores. |
How GLM Relativities Work
A GLM produces a relativity (multiplier) for each level of each rating variable. A relativity of 1.00 means average risk. Above 1.00 means higher risk; below 1.00 means lower risk.
Driver Age 19-24
1.85
85% more than average
Driver Age 45-54
0.78
22% less than average
Sports Car
1.40
40% more than average
Minivan
0.85
15% less than average
Combined effect: A 21-year-old with a sports car = 1.85 x 1.40 = 2.59 relativity. That driver's expected loss is 2.59x the average -- their premium should reflect this.
Regulatory Concern: Proxy Discrimination
A rating variable may appear neutral but actually correlate with a protected class. This is proxy discrimination -- the variable is a proxy for race, ethnicity, or income even though it does not mention those categories directly.
The Classic Example: ZIP Code
ZIP code is a powerful predictor of loss. But ZIP codes also correlate with race and income due to historical housing segregation. Using ZIP code in pricing may result in systematically higher rates for minority neighborhoods -- even if the insurer never intended discrimination.
Regulatory Response
- -- Colorado SB 21-169: Requires insurers to test algorithms for proxy discrimination
- -- NAIC Model Bulletin on AI (Dec 2023): Guidelines for responsible use of AI/ML in insurance
- -- Some states limit or ban credit score use in auto/home pricing
Real-World Scenario: GLM Reveals Hidden Risk Factor
The Setup: Pinnacle Insurance builds a GLM for its personal auto book. Traditional rating used just age, gender, territory, and driving record. The new GLM adds credit-based insurance score, vehicle use (commute distance), and years of continuous coverage.
What Happens: The model reveals that years of continuous coverage is one of the strongest predictors of future loss. Drivers who have maintained coverage for 5+ years without a lapse have 35% fewer claims than drivers who let coverage lapse -- regardless of age or driving record. The relativity for "lapse in last 3 years" is 1.52 (52% higher expected loss).
The Result: Pinnacle adds "continuous coverage" as a rating factor. Loyal, continuously insured customers get lower rates, while lapsed-coverage applicants are surcharged appropriately. The loss ratio improves by 4 points in the first year. However, Pinnacle must also test whether "coverage lapse" is a proxy for income, since lower-income individuals may be more likely to let coverage lapse.
6. Rate Adequacy & the Pricing Balancing Act
Personal lines pricing must balance three competing demands: rates must be adequate (enough to pay claims and expenses), competitive (not so high that customers leave), and compliant (not excessive, inadequate, or unfairly discriminatory under state law).
The Rate Adequacy Process
Analyze Loss Trends
Are losses increasing or decreasing? At what rate? Loss trend projects current losses into the future to ensure premiums match where losses are going, not where they have been.
Analyze Expense Trends
Is it getting more or less expensive to acquire, service, and adjust claims? Commissions, technology costs, staffing -- all factor in.
Calculate Indication
The actuarial indication is the mathematically recommended rate change. Example: "The indication is +8%" means rates need to go up 8% to be adequate.
Management Selects
The selected change is what management actually decides to file. Often less than the indication for competitive reasons. "Indication is +8%, but we will file +5% to avoid losing customers."
File & Get Approval
Must go through the state's rate regulation process. Prior approval states require DOI sign-off before rates take effect. File-and-use states allow implementation while review is pending.
Indication vs. Selected Change
Indication = what the math says is needed. Selected = what management files. These are often different because management considers competitive pressure, customer retention, and political optics. Filing less than the indication means the insurer is knowingly underpricing to maintain market position.
Loss Trend vs. Expense Trend
Loss trend = how fast claim costs are growing (driven by inflation, medical costs, litigation). Expense trend = how fast operating costs change (technology may reduce costs; staffing may increase them). Both must be projected forward to set adequate future rates.
Real-World Scenario: Rate Inadequacy Spiral
The Setup: Bayview Home Insurance writes homeowners in a prior-approval state. Their actuaries calculate an indication of +18% due to rising wildfire losses and construction cost inflation. Management files +18%.
What Happens: The state DOI approves only +7%, citing consumer affordability concerns. The remaining 11% gap means Bayview is now charging less than it costs to provide coverage. The next year, losses increase further. The new indication is +25%. The DOI approves +10%. The gap widens.
The Result: After three years of rate inadequacy, Bayview announces it will stop writing new homeowners policies in the state and begins non-renewing existing policies in the highest-risk areas. Thousands of homeowners are forced into the state FAIR Plan. The "consumer protection" of limited rate increases actually harmed consumers by driving insurers out of the market entirely.
7. Current Challenges in Personal Lines
Exam Alert!
The exam increasingly tests current industry challenges. You need to understand not just what is happening, but why and how these challenges connect to each other. Climate change drives insurer withdrawals, which overload FAIR Plans, which creates systemic risk.
Climate Change
Increasing frequency and severity of catastrophes -- especially wildfire, convective storms (hail/wind), and flooding. Historical data no longer reliably predicts future losses.
Impact: Cat model results are rising. Reinsurers are charging more. Premiums must increase -- but regulatory approval is slow.
Insurer Withdrawals
State Farm, Allstate, and others have reduced or stopped writing new home policies in California, Florida, and Louisiana due to catastrophe losses and rate inadequacy.
Impact: Fewer choices for consumers. Remaining insurers face adverse selection. FAIR Plans absorb the overflow.
FAIR Plan Growth
State residual market plans are growing as private insurers withdraw. California FAIR Plan exposure exceeds $300 billion. If a major wildfire hits, the FAIR Plan may not have enough reserves.
Impact: FAIR Plans can levy assessments on all insurers in the state to cover shortfalls -- a systemic risk.
Social Inflation
Nuclear verdicts ($10M+ jury awards) are increasingly common in personal auto liability. Litigation funding and anti-corporate jury sentiment drive claim costs beyond what traditional trends predict.
Impact: Auto liability loss trends are accelerating. BI severity is outpacing medical inflation.
Technology Disruption
Telematics (usage-based insurance) rewards good drivers. Autonomous vehicles may reduce frequency but increase severity. Ride-sharing creates coverage gaps between personal and commercial use.
Impact: New products needed. Traditional rating factors may become obsolete. First-movers gain competitive advantage.
Rate Inadequacy & Reinsurance Costs
Many states have lengthy approval processes, causing rates to lag behind loss trends. Meanwhile, catastrophe reinsurance pricing has increased significantly since 2020, squeezing insurer margins from both sides.
Impact: Insurers caught between rising costs and capped revenue. Withdrawal becomes the only option for some.
How These Challenges Connect (The Vicious Cycle)
Climate Change
More cat losses
Reinsurance Costs Rise
Reinsurers charge more
Rates Need to Rise
Indication goes up
Regulators Limit Increases
Approved < needed
Insurers Withdraw
Can't operate at a loss
FAIR Plans Overloaded
Systemic risk grows
Cheat Sheet
Print this page for quick referencePortfolio Management
- -- Individual risk OK does NOT mean portfolio OK
- -- Geographic concentration = catastrophe risk
- -- Tools: mapping, PML, ZIP code caps, moratoriums
- -- FAIR Plans = residual market when private exits
Cat Model Components
- -- Hazard: what events could happen?
- -- Vulnerability: how much damage?
- -- Financial: what does it cost the insurer?
- -- Firms: AIR, CoreLogic, Moody's RMS
- -- Uses: pricing, reinsurance, optimization, regulation
Predictive Modeling
- -- GLMs: industry standard, transparent, relativities
- -- GBMs: more accurate, "black box" risk
- -- Neural networks: emerging, least transparent
- -- Proxy discrimination: neutral var + protected class correlation
- -- CO SB 21-169 + NAIC AI Bulletin (Dec 2023)
Rate Adequacy
- -- Indication = actuarial recommendation
- -- Selected = what management files (often less)
- -- Must be: adequate + competitive + compliant
- -- Prior approval states slow the process
- -- Rate lag drives insurer withdrawals
Segment Analysis
- -- Segment by: territory, tier, product, vintage
- -- Profitable segments: grow aggressively
- -- Unprofitable: rate increases, tighten UW
- -- Overall averages hide segment problems
Current Challenges
- -- Climate change: wildfire, convective storms, flooding
- -- Insurer exits: CA, FL, LA markets shrinking
- -- Social inflation: nuclear verdicts in auto BI
- -- Reinsurance costs up sharply since 2020
- -- Telematics + autonomous vehicles = disruption
Exam Trap Alerts
1. Individual Risk vs. Portfolio Risk -- They Are NOT the Same
The exam may describe a perfectly underwritten risk and ask if the insurer should write it. The answer depends on the portfolio context. If the insurer already has 40,000 homes in that hurricane zone, adding one more "good" risk still increases concentration. Always ask: "What does the portfolio need?"
2. Cat Models Simulate -- They Do NOT Predict
A cat model does not predict that a Category 4 hurricane will hit Miami next year. It simulates thousands of possible scenarios and estimates the probability and cost of each. The output is a distribution of possible losses, not a forecast. Do not confuse simulation with prediction.
3. GLMs vs. GBMs -- Know the Regulatory Difference
GLMs are transparent and explainable -- regulators accept them for pricing. GBMs and neural networks are more accurate but harder to explain (the "black box" problem). The exam may ask which model type is most appropriate for a rate filing -- the answer is GLM because regulators require explainability.
4. Proxy Discrimination is NOT Intentional Discrimination
An insurer using ZIP code in its model may have no intent to discriminate by race. But if the ZIP code variable produces systematically higher rates for minority neighborhoods, it is still proxy discrimination. Intent does not matter -- the effect does. The exam tests this distinction.
5. Indication vs. Selected -- Management Often Files LESS
Do not assume the filed rate change equals the actuarial indication. Management routinely selects a lower number than the indication for competitive reasons. If the exam says "the indication is +12% and the company files +7%," the company is knowingly underpricing by 5 points. This is a business decision, not an actuarial error.
6. FAIR Plans Are NOT a Long-Term Solution
FAIR Plans provide coverage when the private market will not, but they are designed as a temporary safety net, not a permanent replacement. FAIR Plans typically offer less coverage, higher premiums, and higher deductibles than private market policies. Their rapid growth signals a market in distress, not a working system.
7. Social Inflation is NOT Regular Inflation
Economic inflation raises the cost of materials and labor. Social inflation increases claim costs through changing jury attitudes, litigation funding, and expanded liability theories. A $50M nuclear verdict for an auto accident is social inflation -- the medical bills did not cost $50M. These are fundamentally different forces on loss trends.
Quick Reference Summary
Portfolio vs. Individual
Good individual risks can create bad portfolios through concentration risk
Geographic Diversification
Spread risk across regions; use mapping, PML, ZIP caps, moratoriums
Cat Model: Hazard
What events could happen? (type, path, frequency, intensity)
Cat Model: Vulnerability
How much damage? (construction, age, codes, elevation)
Cat Model: Financial
What does it cost? (policy terms, deductibles, reinsurance recoveries)
GLMs
Industry standard pricing models -- transparent, produce relativities
GBMs / Neural Networks
More accurate but "black box" -- used for segmentation, not rate filings
Proxy Discrimination
Neutral variable correlating with protected class -- intent doesn't matter
Indication vs. Selected
Actuarial recommendation vs. what management files (often less)
FAIR Plans
Residual market when private insurers withdraw -- growing rapidly in CA, FL, LA
Social Inflation
Nuclear verdicts + litigation funding -- NOT the same as economic inflation
The Vicious Cycle
Climate change -> costs rise -> rates lag -> insurers exit -> FAIR Plans overloaded