Understand the core concepts underlying time series models

Welcome to the World of Time Series!

Hello! If you’ve ever looked at a chart of stock prices, tracked your heart rate over an hour, or wondered if it will rain tomorrow based on the last few days, you’ve already encountered Time Series. In CS2, we move from looking at data as independent "snapshots" to looking at data as a sequence where the order matters.

Don't worry if this seems a bit abstract at first. We are going to break down the "DNA" of time series models—the core concepts that allow us to predict the future by understanding the past. Let's dive in!

1. What exactly is a Time Series?

A time series is simply a collection of observations \(X_t\) made sequentially over time. Usually, these observations are taken at equally spaced time intervals (like every day, every month, or every year).

In CS2, we mostly deal with discrete-time series. This means we look at specific points in time, \(t = 1, 2, 3, ...\). Our goal is usually two-fold:
1. Explanation: Understanding why the data behaves the way it does.
2. Prediction (Forecasting): Using the history of the data to guess what happens next.

Key Terms to Know:

Deterministic: If a series follows a perfect rule (like \(X_t = 2t\)), it is deterministic. We can predict it perfectly. Actuaries rarely deal with these because life is messy!
Stochastic: This is what we care about. A stochastic process involves randomness. We can’t predict the exact value, but we can model the probabilities of different outcomes.

2. The Concept of Stationarity (The "Stability" Rule)

Stationarity is perhaps the most important concept in this chapter. Imagine trying to learn how to drive a car, but every five minutes, the steering wheel and pedals change their functions. It would be impossible! Stationarity means the "rules" governing the series don't change over time.

Weak Stationarity (Second-Order Stationarity)

For a time series to be weakly stationary, it must satisfy three strict conditions:
1. Constant Mean: The average value \(E[X_t]\) must be a constant \(\mu\) and not depend on time \(t\).
2. Constant Variance: The spread of the data \(\text{Var}(X_t)\) must be finite and constant.
3. Consistent Covariance: The way two points (\(X_t\) and \(X_{t+k}\)) relate to each other should only depend on the lag (\(k\)) between them, not the actual time \(t\).

Analogy: Think of a fountain. The water splashes around (randomness), but the overall height and shape of the spray stay roughly the same over time. That’s stationary. Now think of a rocket launch. The position is constantly changing and moving further away. That is non-stationary.

Quick Review: If you see a graph where the data is "trending" upwards or the swings are getting wider and wider, it is not stationary!

3. Measuring Relationships: ACVF and ACF

Since time series data is sequential, today's value is often related to yesterday's value. We measure this "memory" using two tools:

The Autocovariance Function (ACVF)

The ACVF, denoted by \(\gamma_k\), measures the covariance between \(X_t\) and \(X_{t+k}\).
\(\gamma_k = \text{Cov}(X_t, X_{t+k}) = E[(X_t - \mu)(X_{t+k} - \mu)]\)
Note that \(\gamma_0\) (lag 0) is just the variance of the series!

The Autocorrelation Function (ACF)

The ACF, denoted by \(\rho_k\), is a "standardized" version of the ACVF. It tells us the correlation between the series and itself at a later time. It always stays between -1 and 1.
Formula: \(\rho_k = \frac{\gamma_k}{\gamma_0}\)

Did you know? The ACF is a powerful diagnostic tool. If \(\rho_1\) is high, it means the value today is very similar to the value yesterday. If all \(\rho_k\) values are near zero, the data has no memory!

4. The Building Block: White Noise

White Noise is the simplest time series model. It is the "pure randomness" left over after we have explained everything else. A process \(\epsilon_t\) is White Noise if:
1. It has a constant mean of zero: \(E[\epsilon_t] = 0\).
2. It has a constant variance: \(\text{Var}(\epsilon_t) = \sigma^2\).
3. There is zero correlation between different time points: \(\text{Cov}(\epsilon_s, \epsilon_t) = 0\) for \(s \neq t\).

Memory Aid: Think of "static" on an old television. It’s just random jumps with no pattern and no way to predict the next "pixel" from the current one.

5. The Actuary's Toolbox: Operators

To handle time series math easily, we use two "shortcuts" or operators. Mastering these will make your life much easier when we get to AR and MA models later!

The Backward Shift Operator (\(B\))

The \(B\) operator simply moves the clock back one step.
\(B X_t = X_{t-1}\)
If you apply it twice, you go back two steps: \(B^2 X_t = X_{t-2}\).

The Difference Operator (\(\nabla\))

This is used to remove trends and make a series stationary. It measures the change between steps.
\(\nabla X_t = X_t - X_{t-1}\)
Notice that \(\nabla = 1 - B\). This relationship is very important for algebraic manipulations in exams!

Common Mistake: Students often confuse \(B\) and \(\nabla\). Remember: \(B\) is a time machine (it just changes the label), while \(\nabla\) is a subtractor (it calculates a new value).

6. Key Takeaways Summary

Summary Checklist:
• A time series is a sequence of random variables indexed by time.
• Weak Stationarity requires constant mean, constant variance, and covariance that depends only on the lag \(k\).
• ACF (\(\rho_k\)) tells us how strongly the past influences the present.
• White Noise is the ultimate stationary process with zero "memory".
• Use the Backward Shift Operator (\(B\)) to simplify equations.

Don't worry if the math for ACFs feels heavy right now. The most important thing is to understand the concept of stationarity. If you can identify why a series is stationary or not, you've conquered the hardest part of this chapter!