Discretizing continuous distributions: rounding and local moment matching

Welcome to the World of Discretization!

Hello there! If you are preparing for Exam ASTAM, you already know that actuarial science often involves complex continuous distributions like the Gamma or Pareto. But here is a secret: computers and certain recursive formulas (like Panjer’s recursion) are much happier working with discrete numbers—integers like 0, 1, 2, and so on.

In this chapter, we are going to learn how to take a smooth, continuous distribution and "chop it up" into a discrete version. We call this discretization. Think of it like taking a smooth ramp and turning it into a set of stairs. It makes calculations for Aggregate Loss Models much more manageable. Let’s dive in!

Why Do We Discretize?

In the "Aggregate Models" section of your studies, you often need to calculate the total loss \(S = X_1 + X_2 + ... + X_N\). While the individual losses (\(X\)) are often continuous, the math for combining them is much faster if we treat \(X\) as a discrete variable. Discretization is the bridge that gets us there.

Did you know? Actuaries use these techniques to simplify "Collective Risk Models" so they can provide faster pricing for insurance policies without losing too much accuracy!

Method 1: The Method of Rounding

The Method of Rounding (sometimes called the Mass Dispersal Method) is the most intuitive way to discretize. We simply pick a "step size" (called the span, denoted by h) and round every continuous value to the nearest multiple of \(h\).

How it Works:

Imagine we have a span of \(h = 1\). If a loss is 1.2, we round it to 1. If it is 1.8, we round it to 2. To find the probability of a specific discrete point \(x_j = jh\), we collect all the probability from the continuous distribution that is "closest" to that point.

The formulas you need to know are:

\( f_0 = P(X < h/2) = F(h/2) \)

\( f_j = P(jh - h/2 \le X < jh + h/2) = F(jh + h/2) - F(jh - h/2) \) for \(j = 1, 2, ...\)

An Everyday Analogy:

Think of a row of buckets placed 1 meter apart. You have a hose (the continuous distribution) spraying water along the line. The Method of Rounding simply says that all the water falling within half a meter of a bucket drops into that specific bucket.

Quick Review: Rounding Basics

1. h is your span (the distance between discrete points).
2. \(f_0\) is special; it only looks "forward" from zero to half the span.
3. \(f_j\) looks half a span back and half a span forward.

Method 2: Local Moment Matching

Don't worry if this seems a bit more technical! The Method of Local Moment Matching is just a more sophisticated way of discretizing. Instead of just rounding, we want the discrete distribution to "act" like the continuous one by keeping the mean (and sometimes higher moments) the same within each interval.

Matching the First Moment (The Mean)

Usually, in ASTAM, we focus on matching the first moment. This means that for any interval \([jh, (j+1)h]\), the average value of the discrete mass we assign to the endpoints must equal the average value of the continuous distribution in that same interval.

For each interval \([jh, (j+1)h]\), we distribute the probability mass to the two endpoints (\(jh\) and \((j+1)h\)).

Step-by-Step Process:
1. Calculate the probability mass in the interval: \(m_0^j = \int_{jh}^{(j+1)h} f(x) dx = F((j+1)h) - F(jh)\).
2. Calculate the local mean in that interval: \(m_1^j = \int_{jh}^{(j+1)h} x f(x) dx\).
3. Assign the mass to the endpoints using these weights:
The mass added to point \(jh\) is: \( \frac{(j+1)h \cdot m_0^j - m_1^j}{h} \)
The mass added to point \((j+1)h\) is: \( \frac{m_1^j - jh \cdot m_0^j}{h} \)

The "Balance Scale" Analogy:

Imagine a seesaw between two points, \(jh\) and \((j+1)h\). The continuous mass is sitting somewhere on that seesaw. To keep the seesaw balanced exactly the same way, you have to split that mass between the two ends. If the "weight" is closer to the right side, the right endpoint gets more mass!

Key Point to Remember:

In this method, each discrete point \(jh\) (except for zero) actually receives "contributions" from two adjacent intervals: the interval to its left and the interval to its right.

Common Mistakes to Avoid

1. Forgetting the Span (h): Students often forget to include \(h\) in their calculations. If the problem asks for a span of 2, your points are 0, 2, 4, 6... not 0, 1, 2, 3!

2. Miscalculating \(m_1^j\): Remember that \(m_1^j\) is the integral of \(x \cdot f(x)\). For many distributions (like the Exponential), you might need to use integration by parts or use the formula for the Limited Expected Value to find this easily.

3. The Endpoints: In the Rounding Method, \(f_0\) only covers half an interval, while in Moment Matching, the points are built up from two side-by-side intervals. Always double-check your boundaries.

Summary and Key Takeaways

Discretization is essential for calculating aggregate losses in practice. You have two main tools in your belt:

Method of Rounding:
- Simple and fast.
- Uses the CDF (\(F(x)\)) at the midpoints between your discrete values.
- Best for a quick approximation.

Local Moment Matching:
- More accurate because it preserves the mean of the distribution.
- Requires calculating the mass (\(m_0\)) and the local mean (\(m_1\)) for each interval.
- Remember that each discrete point \(f_j\) is the sum of two pieces: one piece from the interval ending at \(j\) and one from the interval starting at \(j\).

Keep practicing these calculations! Once you get the rhythm of the integrals, these are "guaranteed points" on the exam. You've got this!