Understanding Filtrations

Hello! Let's begin our next lesson.

Introduction

In our last session, we established a rigorous definition for conditional expectation. We saw that conditioning on a $\sigma$-algebra $\mathcal{G}$ means finding the best estimate of a random variable given the "information" encoded in $\mathcal{G}$. This provided a powerful, unified view of a concept you've used extensively in econometrics and statistics.

However, our model of information, the $\sigma$-algebra $\mathcal{G}$, was static. In nearly all applications you've encountered—from time series analysis to corporate finance—information is not static; it arrives sequentially over time. Today, we will build the formal machinery to model this dynamic flow of information.

This lesson addresses the learning outcome: Define a filtration as an increasing sequence of σ-algebras that formally models the flow of information over time. We will see how this simple but profound concept provides the foundation for analyzing stochastic processes.

1. From Static to Evolving Information

In the previous lesson, we interpreted a $\sigma$-algebra $\mathcal{G}$ as representing a fixed set of distinguishable events. For example, knowing only whether a dice roll is "even" or "odd" corresponds to the $\sigma$-algebra ${\mathcal{G}}_{\text{parity}}=\{\varnothing ,\{1,3,5\},\{2,4,6\},\Omega \}$.

Now, let's imagine a process that unfolds over time, like a series of coin flips.

• At time $t=0$, before any flips, we have no information. Our knowledge is represented by the trivial $\sigma$-algebra, ${\mathcal{F}}_0=\{\varnothing ,\Omega \}$. We only know something will happen.
• At time $t=1$, after the first flip, we learn its outcome. Our information set expands. We can now distinguish between the events {First flip is H} and {First flip is T}. Our new $\sigma$-algebra, ${\mathcal{F}}_1$, contains these sets (and their unions and complements).
• At time $t=2$, after the second flip, our information set grows again. We can now distinguish between {HH}, {HT}, {TH}, and {TT}. Our $\sigma$-algebra ${\mathcal{F}}_2$ is larger than ${\mathcal{F}}_1$.

Notice the crucial property: anything we knew at time $t=1$ is still known at time $t=2$. This means that every event in ${\mathcal{F}}_1$ must also be in ${\mathcal{F}}_2$, so ${\mathcal{F}}_1\subseteq {\mathcal{F}}_2$. We are modeling a world where information accumulates and is never forgotten. This sequence of nested, or "increasing," $\sigma$-algebras is the core idea of a filtration.

The following video provides a clear introduction to this concept and walks through the coin-flip example in detail.

2. The Formal Definition of a Filtration

The intuition from the coin-flip example leads directly to the formal definition. We are simply formalizing the idea of an "increasing sequence of $\sigma$-algebras."

As the text states, a filtration on a measurable space $(\Omega ,\mathcal{F})$ is a family of sub-$\sigma$-algebras of $\mathcal{F}$, indexed by time $t\in T$, denoted $\{{\mathcal{F}}_t{\}}_{t\in T}$, with the property that:
$\text{For any }s,t\in T\text{ with }s\le t,\text{ we have }{\mathcal{F}}_s\subseteq {\mathcal{F}}_t.$

• The index set $T$ can be discrete, like $T=\{0,1,2,\dots \}$ for our coin-flip example, or continuous, like $T=[0,\infty )$ for modeling stock prices.
• The condition ${\mathcal{F}}_s\subseteq {\mathcal{F}}_t$ is the mathematical statement for "information is never forgotten."
• A probability space equipped with a filtration, $(\Omega ,\mathcal{F},\{{\mathcal{F}}_t{\}}_{t\in T},P)$, is called a filtered probability space. This is the standard setting for the modern theory of stochastic processes.

3. The Natural Filtration and Adapted Processes

This raises a practical question: where do these filtrations come from? In most applications, the information we have is generated by observing some random process unfolding over time. This gives rise to the most important type of filtration.

A stochastic process is simply a collection of random variables $\{X_t{\}}_{t\in T}$ indexed by time. You are deeply familiar with these from your work in time series and finance (e.g., a stock price, an interest rate series, etc.).

The natural filtration generated by a process $X$ is the sequence of $\sigma$-algebras where ${\mathcal{F}}_t^X$ is the information generated by observing the process $X$ up to and including time $t$. Formally:
${\mathcal{F}}_t^X=\sigma (\{X_s:s\in T,s\le t\})$
This is the smallest $\sigma$-algebra that makes all the random variables $X_s$ for $s\le t$ measurable. It represents precisely the information obtained by watching the history of $X$, and nothing more.

This leads to another key definition: a process $\{Y_t{\}}_{t\in T}$ is said to be adapted to a filtration $\{{\mathcal{F}}_t{\}}_{t\in T}$ if, for every $t\in T$, the random variable $Y_t$ is ${\mathcal{F}}_t$-measurable.

Intuitively, this means that the value of the process at time $t$, $Y_t$, can be determined from the information available at time $t$, ${\mathcal{F}}_t$. A process is, by definition, always adapted to its own natural filtration.

Let's connect this to your background. Consider the Black-Scholes model in finance. The stock price process $\{S_t{\}}_{t\ge 0}$ generates a natural filtration $\{{\mathcal{F}}_t^S{\}}_{t\ge 0}$. An option written on this stock has a price at time $t$, let's call it $C_t$. For the model to be sensible, the option's price $C_t$ must be determined by the history of the stock price up to time $t$. In other words, the option price process $\{C_t\}$ must be adapted to the filtration $\{{\mathcal{F}}_t^S\}$. You cannot know the price of the option if it depends on future, unknown stock movements.

The concept of an adapted process formalizes this fundamental requirement of non-anticipation.

Conclusion

Let's recap the central ideas of this lesson.

• Filtration: A filtration, $\{{\mathcal{F}}_t{\}}_{t\in T}$, is an increasing sequence of $\sigma$-algebras (${\mathcal{F}}_s\subseteq {\mathcal{F}}_t$ for $s\le t$). It is the formal mathematical structure for modeling the accumulation of information over time.
• Filtered Probability Space: This is the complete structure $(\Omega ,\mathcal{F},\{{\mathcal{F}}_t\},P)$ that serves as the foundation for studying processes that evolve in time.
• Natural Filtration: A stochastic process $\{X_t\}$ generates its own filtration, ${\mathcal{F}}_t^X=\sigma (\{X_s:s\le t\})$, which represents the history of the process itself.
• Adapted Process: A process $\{Y_t\}$ is adapted to a filtration $\{{\mathcal{F}}_t\}$ if $Y_t$ is ${\mathcal{F}}_t$-measurable for all $t$. This formalizes the idea that the state of the process at time $t$ is knowable given the information available at time $t$.

Preview of the Next Lesson:
We have now modeled the flow of information. The next question is: how does a process behave relative to this information flow? We will introduce the concept of a martingale, a process whose expected future value, given all the information we have today, is simply its value today. This formalizes the notion of a "fair game" and has deep connections to the efficient market hypothesis, a cornerstone of financial economics.