Welcome to the World of Distributions!
Welcome, future actuary! In this chapter, we are going to explore the "DNA" of actuarial science: univariate distributions. Think of a distribution as a pattern or a "blueprint" that tells us how likely different outcomes are. Whether you are predicting how many people will make an insurance claim today or how large those claims might be, these distributions are your most important tools.
Don't worry if statistics has felt a bit dry in the past. We’re going to break these down into real-world scenarios so you can see exactly how they work.
1. Discrete Distributions: The "Counting" Tools
Discrete distributions are used when we are counting things (1, 2, 3...). You can't have 2.5 insurance claims!
The Bernoulli Distribution: The "Yes/No" Trial
This is the simplest distribution. It models a single trial with only two possible outcomes: Success (1) or Failure (0).
Example: Flipping a coin once, or checking if a single policyholder renews their contract.
Key Property: If the probability of success is \(p\), the Mean is \(p\) and the Variance is \(p(1-p)\).
The Binomial Distribution: Multiple Trials
If you take a bunch of Bernoulli trials and add them together, you get a Binomial distribution. It tells us the number of successes in \(n\) independent trials.
Example: If you have 100 policyholders, how many will make a claim this year? Each policyholder is a "trial."
Quick Review: For \(X \sim Bin(n, p)\):
\(E[X] = np\)
\(Var(X) = np(1-p)\)
The Poisson Distribution: Events in Time
The Poisson distribution models the number of events happening in a fixed interval of time or space.
Analogy: Imagine standing on a bridge counting cars passing under you. If you know that, on average, 10 cars pass per hour, the Poisson distribution helps you calculate the chance that exactly 15 pass in the next hour.
Did you know? In a Poisson distribution, the Mean and Variance are equal (\(\lambda\)). This is a classic exam "shortcut" to identify a Poisson process!
Geometric & Negative Binomial: Waiting for Success
These distributions focus on when a success happens.
- Geometric: How many trials until the first success? (e.g., How many houses do I inspect before I find one with fire damage?)
- Negative Binomial: How many trials until the k-th success? (e.g., How many calls does a salesperson make before getting 3 sales?)
Common Mistake: Always check your Formula Tables! Some versions of these distributions count the "number of trials," while others count the "number of failures." In CS1, the IFoA usually defines the Geometric as the number of failures before the first success.
Key Takeaway: Use Discrete distributions for counts. Remember that Binomial is for a fixed number of trials, while Poisson is for a fixed window of time.
2. Continuous Distributions: The "Measuring" Tools
Continuous distributions are used for measurements like time, weight, or money. These can take any value in a range (e.g., a claim could be £1,250.55).
The Exponential Distribution: Waiting Time
The Exponential distribution is the "partner" to the Poisson. While Poisson counts how many events happen, Exponential measures the time between those events.
Crucial Feature: Memorylessness. This is a favorite exam topic! It means that the probability of an event happening in the next 10 minutes is the same regardless of whether you’ve already been waiting for an hour or just arrived. The distribution "forgets" the past.
The Gamma Distribution: Adding Up Wait Times
If you add up several Exponential variables, you get a Gamma distribution. It’s often used to model the total time taken for several independent events to occur, or to model claim sizes because it is flexible and always positive.
The Normal Distribution: The Famous Bell Curve
The Normal distribution \(N(\mu, \sigma^2)\) is symmetric and describes many natural phenomena. In actuarial work, we often assume the sum of many independent risks follows a Normal distribution (thanks to the Central Limit Theorem).
The Lognormal and Pareto: Modeling Big Risks
Insurance claims aren't always symmetric. Often, we have many small claims and a few massive ones.
- Lognormal: If \(\ln(X)\) is Normal, then \(X\) is Lognormal. It is "skewed" to the right (has a long right tail).
- Pareto: This is a "heavy-tailed" distribution. It is used for catastrophic risks (like earthquakes or giant industrial fires) where the chance of an extreme "outlier" event is much higher than the Normal distribution would suggest.
Key Takeaway: Continuous distributions model "how much" or "how long." If your data has extreme values, look toward Lognormal or Pareto rather than the Normal distribution.
3. Generating Samples: The Probability Integral Transform (PIT)
How do computers simulate random numbers for a specific distribution? They use a clever trick called the Probability Integral Transform (PIT).
The Concept
Most computers can easily generate a random number \(U\) between 0 and 1 (a Uniform(0,1) distribution). The PIT tells us that we can turn that \(U\) into a value from any distribution we want, as long as we know its Cumulative Distribution Function (CDF), \(F(x)\).
Step-by-Step: How to Generate a Sample
- Generate a random number \(u\) from a \(Uniform(0,1)\) distribution.
- Set \(u = F(x)\), where \(F(x)\) is the CDF of the distribution you want to sample from.
- Solve for \(x\). This is often written as \(x = F^{-1}(u)\).
Example: Generating an Exponential Sample
The CDF of an Exponential distribution is \(F(x) = 1 - e^{-\lambda x}\).
1. Set \(u = 1 - e^{-\lambda x}\)
2. Rearrange: \(1 - u = e^{-\lambda x}\)
3. Take the natural log: \(\ln(1-u) = -\lambda x\)
4. Solve for \(x\): \(x = -\frac{1}{\lambda} \ln(1-u)\)
Since \(1-u\) is also Uniform(0,1), you will often see this simplified as \(x = -\frac{1}{\lambda} \ln(u)\).
Mnemonic: Think of the CDF as a "translator." You give it a standardized value (0 to 1), and it translates it into the "language" of your specific distribution.
Key Takeaway: To generate a sample, set the random number \(u\) equal to the CDF and solve for \(x\). This is the foundation of Monte Carlo simulation.
Quick Review Box
Checklist for Exams:
• Do you know which distributions are Discrete vs Continuous?
• Can you identify the Memoryless distribution? (Exponential)
• Do you know which distribution has Mean = Variance? (Poisson)
• Can you perform the PIT steps to solve for \(x\)?
• Remember: Skewed distributions (Lognormal/Pareto) are usually better for claim sizes than the Normal distribution.
Don't worry if this seems tricky at first! Actuarial statistics is about patterns. Once you recognize the "personality" of each distribution, choosing the right one becomes second nature. Keep practicing those CDF rearrangements!