The Art of Term Structure Models: Volatility and Distribution

Welcome to the Art of Term Structure Models!

Hello there! If you’ve made it to FRM Part II, you already know that modeling interest rates is a bit like trying to predict the weather—it’s complex, constantly changing, and involves a lot of moving parts. In this chapter, "The Art of Term Structure Models: Volatility and Distribution," we dive into the "how" and "why" behind different interest rate models. We focus on how these models handle volatility (the "wiggle" in the rates) and the distribution (the shape of potential future rates).

Don't worry if this seems a bit abstract at first. We aren't just looking at math for math's sake; we are trying to find the best tool to price bonds and manage risk. Let's break it down step-by-step!

1. The Foundation: Normal vs. Lognormal Distributions

When we model how interest rates move, we have to decide what "shape" their future possibilities take. This is the difference between Normal and Lognormal models.

Normal Models (Arithmetic Models)

In a Normal model (like the Ho-Lee or Vasicek models), the change in the interest rate is expressed in absolute terms (e.g., a 1% increase).
The Pros: They are mathematically simple to work with.
The Cons: Interest rates can technically become negative. While we have seen negative rates in the real world (like in Europe or Japan), many traditional models weren't designed for this.

Lognormal Models (Geometric Models)

In a Lognormal model (like the Black-Derman-Toy (BDT) model), we model the logarithm of the interest rate.
The Pros: Interest rates can never be negative. As rates get closer to zero, their volatility (in absolute terms) decreases.
The Cons: They can be mathematically more "expensive" or difficult to solve, and they might overstate the probability of very high interest rates.

Quick Review: Normal vs. Lognormal

• Normal: Rates can go negative. Constant volatility in basis point terms.
• Lognormal: Rates stay positive. Volatility is proportional to the level of the rate.

2. The Concept of Mean Reversion

Think of Mean Reversion like a rubber band. If you pull the interest rate too far away from its "natural" average level, the model pulls it back. This is a crucial feature because, unlike stock prices, interest rates don't usually head toward infinity or stay at zero forever; they tend to hang around a long-term average driven by the economy.

The speed at which the rate returns to the average is called the reversion speed (usually denoted as \(k\)).
- If \(k\) is high, the "rubber band" is very tight, and rates snap back quickly.
- If \(k\) is zero, the model has no mean reversion (like the Ho-Lee model).

Analogies for your memory:
Imagine a marble in a bowl. No matter which way you flick it, it eventually rolls back to the bottom. That "bottom" is the long-term mean.

3. Meet the "Family" of Models

The curriculum focuses on a few key models. Let's look at them based on their "personality" (their features):

A. The Ho-Lee Model

This is the simplest "No-Arbitrage" model. It assumes the short rate follows a normal distribution and has no mean reversion.
\( dr_t = \theta(t)dt + \sigma dW_t \)
- Key Insight: It uses a "drift" term \(\theta(t)\) to make sure the model perfectly matches today's term structure (the current yield curve).

B. The Vasicek Model

This model introduces Mean Reversion but keeps the Normal distribution.
\( dr_t = k(\theta - r_t)dt + \sigma dW_t \)
- The Catch: Because it's a Normal model, rates can still go negative.

C. The Cox-Ingersoll-Ross (CIR) Model

CIR is like the Vasicek model's smarter sibling. It has mean reversion, but it adds a "square root" term to the volatility.
\( dr_t = k(\theta - r_t)dt + \sigma \sqrt{r_t} dW_t \)
- Why it matters: As the interest rate \(r_t\) gets close to zero, the volatility (\(\sigma \sqrt{r_t}\)) also goes to zero. This helps prevent rates from becoming negative.

D. The Black-Derman-Toy (BDT) Model

This is a Lognormal model. It is very popular in industry for valuing options on bonds.
- Key Insight: It can match both the current yield curve and a provided volatility smile.
- Remember: BDT does not have mean reversion in its original form, and it assumes volatility is a function of time.

Key Takeaway Table

Model | Distribution | Mean Reversion? | Can rates be negative?
Ho-Lee | Normal | No | Yes
Vasicek | Normal | Yes | Yes
CIR | Non-Central Chi-Sq | Yes | No
BDT | Lognormal | No | No

4. Volatility Structures

In "The Art" of modeling, we have to decide how volatility (\(\sigma\)) behaves over time. There are two main ways to look at this:

1. Constant Volatility: We assume the "wiggle" in rates is the same today as it will be in 10 years. (Simple, but often unrealistic).
2. Time-Dependent Volatility: We allow \(\sigma\) to change over time (\(\sigma(t)\)). This allows the model to match the prices of market-traded options (like swaptions or caplets).

Common Mistake to Avoid:
Don't confuse Rate Volatility with Price Volatility. As a bond's maturity increases, its price becomes more sensitive to interest rate changes (higher duration), even if the interest rate volatility stays the same!

5. The Drift Term: Fitting the Yield Curve

You will often see the symbol \(\theta\) or \(\theta(t)\). In these models, this is the Drift.
In "Equilibrium" models, drift is often constant.
In "No-Arbitrage" models, we "force" the drift to change over time so that the model's output perfectly matches the current market yield curve.
Think of it like adjusting the settings on your GPS so that your "current location" exactly matches where you are standing on the map.

6. Path Dependency and Trees

To calculate prices, we often use Binomial or Trinomial Trees.
- Path-Independent: It doesn't matter how you got to a specific rate; the value is the same. Most of the models we discussed (like Ho-Lee or Vasicek) are path-independent when mapped on a recombining tree.
- Path-Dependent: The value depends on the specific "path" the interest rates took (e.g., if the rate hit 5% at any point in the past). These are much harder to value and usually require Monte Carlo simulations instead of trees.

Quick Tip: Recombining Trees

A tree "recombines" if an Up-Down move leads to the same rate as a Down-Up move. This makes the math much faster because there are fewer nodes to calculate! Ho-Lee and Vasicek recombine easily. BDT also recombines because it uses a log-scale.

Summary and Key Takeaways

- The Distribution Choice: Normal models are easier but allow negative rates. Lognormal (BDT) and Square-root (CIR) models prevent negative rates.
- Mean Reversion: This is the "gravity" that pulls rates back to a long-term average. It is a key feature of Vasicek and CIR.
- Calibration: We adjust the "drift" (\(\theta\)) to make the model match today's market prices.
- The Goal: There is no "perfect" model. The "Art" is choosing the model whose assumptions (distribution, volatility, mean reversion) best fit the specific risk management task at hand.

Keep practicing those tree calculations and remember the "personality" of each model. You've got this!