Why ARCH and GARCH? A Gap in Classical Modeling

Financial data rarely behaves calmly. Prices spike, plunge, pause, and surge again—often without warning. Traditional time-series models like ARIMA or linear regression help capture patterns and trends, but they assume constant variance. Real markets don’t behave that politely. Their volatility changes constantly, and this “changing variance” itself becomes part of the story.
To navigate these turbulent seas of volatile data, statisticians developed a class of models specifically designed to capture changing variance over time: ARCH and GARCH.

Why ARCH and GARCH? A Gap in Classical Modeling
Classical time-series techniques aim to model the level or trend in a series—sales, exchange rates, traffic, etc. But analysts kept confronting a practical frustration:
These models could capture the mean, but they could not explain the changing variability around the mean.
In real business settings—new markets, product launches, unstable economic periods—volatility spikes. During calm phases, the turbulence settles. If we ignore this changing volatility, we risk underestimating uncertainty and making poor forecasts.
Hence, the need for models that explicitly describe how variance behaves with time. This is where ARCH (Autoregressive Conditional Heteroskedasticity) models enter.

ARCH: The Foundation of Volatility Modeling
ARCH models treat volatility as a dynamic quantity. The variance of the error term at time t depends on past error terms.
In simpler words:
Big shocks in the past → high volatility now
Calm periods in the past → low volatility now
ARCH assumes error terms follow an autoregressive model. However, if the residuals have both autoregressive (AR) and moving average (MA) structure—i.e., they follow ARMA—you need something more powerful: GARCH.

GARCH: A More Flexible Version of ARCH
GARCH (Generalized ARCH) models allow the variance to depend not just on past errors, but also on past variances. This makes them extremely effective in modeling financial and economic time series—stock markets, exchange rates, commodities, and other instruments that exhibit volatility clustering.
Before we dig deeper, let’s understand the key phenomenon these models capture.

Volatility Clustering: The Heart of GARCH
Financial data often shows a striking pattern:
Periods of high volatility tend to be followed by more high volatility.
Periods of low volatility tend to stay quiet.
This “clustering” is why markets may stay turbulent for weeks after a shock—news, policy changes, geopolitical events—before gradually calming.
GARCH models excel at capturing these patterns.
But you must remember:
GARCH captures volatility patterns – it does not explain the underlying causes.
GARCH is great for predicting when volatility will rise or fall, but not how much the actual price will move.
Reliable modeling usually needs thousands of observations.

Understanding GARCH(p, q)
GARCH models use the notation GARCH(p, q):
p = number of GARCH terms (lagged variances)
q = number of ARCH terms (lagged squared residuals)
The most widely used model is GARCH(1,1) because it captures most real-world volatility dynamics with minimal parameters.
The parameters (often called alpha1 and beta1) give deep insights.
Volatility Persistence
Volatility persistence is measured by the sum:
alpha1 + beta1

If > 1 : volatility explodes (rare in real markets)
If = 1 : shocks decay exponentially → half-life becomes infinite
If < 1 : volatility eventually stabilizes (most common scenario)
Half-Life of Volatility Shock
A useful metric:
half-life = log(0.5) / log(alpha1 + beta1)

This tells you how long it takes for a volatility shock to fall by half.
Because persistence and half-life rely on training data patterns, you need large samples—often 10,000+ points—to avoid misleading estimates.

Hands-On Implementation in R
Let’s now move from theory to practice using R.
We will use:
Ecdat package for exchange rate dataset
fGarch package for fitting a GARCH model
The dataset contains 1867 daily observations of USD exchange rates (1980–1987).

Step 1: Load the Data
install.packages("Ecdat")
library(Ecdat)
mydata = Garch
str(mydata)

We then convert date and day into proper formats:
mydata$date = as.Date(mydata$date, origin = "01-02-1980")
mydata$day = as.factor(mydata$day)

Step 2: Load Supporting Packages
install.packages(c("tseries","urca","fUnitRoots","forecast","fGarch"))

library(fGarch)
library(tseries)
library(urca)
library(fUnitRoots)
library(forecast)

Step 3: Create the Time Series
exchange_rate_dollar_deutsch_mark <- ts(
mydata$dm, start=c(1980, 1), end=c(1987, 5), frequency=266
)
plot.ts(exchange_rate_dollar_deutsch_mark,
main="exchange_rate_dollar_deutsch_mark")

The plot reveals small, persistent fluctuations—perfect for volatility analysis.

Step 4: Compute Log Differences (Inflation)
inflation_series <- diff(log(exchange_rate_dollar_deutsch_mark)) * 100
plot.ts(inflation_series, main="Inflation of exchange rate")
summary(inflation_series)

The summary reveals heavy variability (range: −2.8 to +5.5), indicating suitability for GARCH.

Step 5: Identify ARIMA Structure
We examine ACF and PACF:
acf(inflation_series)
pacf(inflation_series)

The spikes suggest an ARIMA(5,0,0) structure.
Fit the model:
Arima_5_0_0 <- arima(inflation_series[1:499], order=c(5,0,0))
residual <- Arima_5_0_0$resid
acf(residual); pacf(residual)

Test residual autocorrelation:
Box.test(residual, lag=20, type="Ljung-Box")

A high p-value confirms a good ARIMA fit.

Step 6: Fit GARCH(1,1)
garch.fit <- garchFit(
formula = ~ arma(5,0) + garch(1,1),
data = inflation_series[1:500]
)

Check the model:
summary(garch.fit)

The output will show parameter estimates, residual diagnostics, and persistence.

Step 7: Visualize GARCH Diagnostic Plots
plot(garch.fit)

The fGarch package offers 13 diagnostic plots, including:
Time series
Conditional standard deviations
Standardized residuals
Volatility clustering view
QQ-plots
ACF of squared residuals
For example:
Time series plot
Shows the underlying inflation data.
Series with 2 Standard Deviations
Overlays conditional volatility, clearly revealing clusters of high and low variance.
These plots help verify if the model captures volatility patterns effectively.

Conclusion: A Powerful Tool for Real-World Volatility
ARCH and GARCH models fill a critical gap in time-series modeling by allowing variance itself to evolve with time. Whether you are analyzing currencies, stocks, commodities, or economic indicators, understanding volatility is essential for forecasting, risk management, and scenario analysis.
This tutorial walked through:
Fundamental concepts of ARCH/GARCH
Volatility clustering & persistence
Parameter interpretation
Half-life of shocks
Practical steps to build a GARCH model in R
With these foundations and tools like fGarch, you can start modeling real-world financial volatility and gain insights into complex time-dependent behaviors.

Perceptive Analytics provides end-to-end BI solutions through its team of expert Microsoft Power BI consultants. As a trusted Power BI consulting company, we help organizations modernize reporting, automate workflows, and build scalable analytics systems tailored to their business goals.