Part VIII · Chapter 25

ARCH Models

ARCH models conditional variance using past squared shocks: today's risk depends on yesterday's surprise. Volatility clusters in calm and turbulent regimes — exactly what ARCH captures.

\( h_t = \alpha_0 + \alpha_1 \varepsilon_{t-1}^2,\;\; \varepsilon_t = \sqrt{h_t} z_t \)

Learning objectives

Simulate an ARCH(1) and observe volatility clustering.
Interpret α₀, α₁; stability requires α₁ < 1.
Identify ARCH effects by squared-return autocorrelation.
Understand that ARCH forecasts variance, not direction.

ARCH Clustering Simulator

Seed:

Status

Persistence α₁:
Long-run var: (if α₁ < 1)

Squared-return ACF should be POSITIVE — that's the ARCH signature.

📐 ARCH(1)

\[ h_t = \alpha_0 + \alpha_1 \varepsilon_{t-1}^2,\quad \alpha_0 > 0,\;\;\alpha_1 \geq 0 \]
\[ \text{Long-run variance: }\frac{\alpha_0}{1-\alpha_1}\quad(\alpha_1<1) \]
α₀ floor variance
α₁ shock sensitivity
Stability α₁ < 1
ARCH LM test T·R² ~ χ²(q) on squared residuals

🔍 What to look for

⚠️ Pro Tip: What to Avoid

Student says

"GARCH/ARCH predicts whether the market will go up tomorrow."

Why this is wrong

ARCH/GARCH forecast VARIANCE (risk magnitude), not the sign of returns. Returns themselves may be ≈ white noise in the conditional mean.

Correct interpretation

"Conditional variance ĥ_{t+1} = α₀ + α₁ ε̂_t². Use for VaR, option pricing, position sizing — not direction."

📝 Mini-quiz

📋 Key Takeaways

QuantityNote
ARCH(1) varianceα₀ + α₁ ε²_{t-1}
Stabilityα₁ < 1
Long-run varianceα₀ / (1 − α₁)
ARCH LM testT·R² on squared residuals ~ χ²(q)
ForecastVariance, NOT direction