14.1 Characteristics of volatility in time series

Understanding Volatility in Time Series

Definition and Importance of Volatility

Volatility measures how much a time series fluctuates over time. More precisely, it quantifies the degree of variation in a series, capturing the uncertainty or risk tied to how large those changes can be. Think of stock prices or exchange rates: some periods are calm, others are wild. Volatility puts a number on that difference.

Why does this matter? Volatility is central to:

• Risk management — understanding how much a portfolio's value could swing
• Portfolio optimization — balancing expected returns against potential losses
• Pricing financial derivatives — options and futures prices depend directly on expected volatility

Beyond finance, volatility shows up in economics, environmental science, and anywhere that uncertainty in a time series affects decisions.

Key Characteristics of Financial Volatility

Two patterns show up again and again in real financial data, and both are reasons why simple models fall short.

Volatility clustering means that large changes tend to follow large changes, and small changes tend to follow small changes. If you look at daily stock returns, you'll notice stretches of high turbulence bunched together, then stretches of relative calm. This isn't random — it implies that volatility is persistent and autocorrelated. A volatile day makes tomorrow more likely to be volatile too.

Asymmetry (the leverage effect) means that negative shocks move volatility more than positive shocks of the same size. A 3% drop in a stock index typically increases future volatility more than a 3% gain decreases it. This is why market crashes produce outsized spikes in volatility compared to equivalent rallies. The term "leverage effect" comes from the idea that falling prices raise a firm's debt-to-equity ratio, making it riskier.

Modeling Volatility in Time Series

Concept of Heteroskedasticity

Traditional time series models like ARMA and ARIMA assume homoskedasticity, meaning the variance of the error terms stays constant over time: $\sigma^2_t = \sigma^2$ for all $t$.

Heteroskedasticity is when that assumption breaks down — the variance of the error terms changes over time: $\sigma^2_t \neq \sigma^2$. In financial data, this is the norm, not the exception. Calm periods have small error variance; turbulent periods have large error variance.

Why you can't just ignore it:

• Parameter estimates become inefficient (not as precise as they could be)
• Confidence intervals and hypothesis tests become unreliable
• Forecasts miss the changing risk level entirely

This is exactly why specialized models like ARCH and GARCH were developed — they let the variance itself be a function of past data, rather than treating it as fixed.

Limitations of Traditional Volatility Models

ARMA and ARIMA models are powerful tools for modeling the level of a time series, but they have real blind spots when it comes to volatility:

• Constant variance assumption — They treat variance as fixed, so they can't represent periods of high vs. low volatility.
• No volatility clustering — They have no mechanism to capture the persistence where turbulent periods stick together and calm periods stick together.
• Symmetric treatment of shocks — A positive surprise and a negative surprise of the same size produce identical effects, so the leverage effect goes completely unmodeled.

These gaps motivated the development of ARCH and GARCH models, which explicitly model time-varying variance. ARCH lets today's variance depend on past squared errors, and GARCH extends that by also including past variance values. Together, they capture clustering and persistence. Extensions like EGARCH and GJR-GARCH go further by incorporating asymmetry.