ARCH Models
Volatility
clustering hands us a demand and a clue. The demand: build a model whose variance
moves over time. The clue: the squared returns are autocorrelated, so a big shock yesterday should
predict a big variance today. In 1982 Robert Engle
wrote that clue down as an equation — and it was such a good idea that it won him the 2003 Nobel Prize
in Economics. He called it ARCH: Autoregressive Conditional
Heteroskedasticity. Unpack the name and you have the whole model:
heteroskedasticity = non-constant variance; conditional = the variance for today
depends on what we already know; autoregressive = it depends on the recent past of the series
itself.
The model
Write each return as a time-varying standard deviation times a standardized shock:
r_t = \sigma_t\, z_t, \qquad z_t \overset{\text{iid}}{\sim} (0, 1),
where the z_t are independent draws with mean 0
and \operatorname{Var}(z_t)=1 (often standard normal). All the action is in
\sigma_t, the conditional standard deviation. The
ARCH(q) model makes the conditional variance
\sigma_t^2 a linear function of the last q squared
returns:
\sigma_t^2 = \alpha_0 + \alpha_1 r_{t-1}^2 + \alpha_2 r_{t-2}^2 + \dots + \alpha_q r_{t-q}^2,
with \alpha_0 > 0 and every \alpha_i \ge 0 (so
the conditional variance can never go negative — a
variance must be
non-negative). Read the ARCH(1) special case aloud and the mechanism is obvious:
\sigma_t^2 = \alpha_0 + \alpha_1 r_{t-1}^2.
A big shock r_{t-1} (of either sign — it is squared) pushes up
today's variance; a big variance makes a big r_t likely; that big
r_t feeds the next day's variance. The feedback loop is volatility
clustering, written in one line.
The news-impact curve, live
Fix \alpha_0 = 0.2 and watch how today's variance
\sigma_t^2 = \alpha_0 + \alpha_1 r_{t-1}^2 responds to yesterday's shock
r_{t-1}. It is a parabola — the news-impact curve — symmetric in the
sign of the news: a -3\% day and a +3\% day raise
tomorrow's variance by exactly the same amount. Drag \alpha_1 and the bowl
steepens: larger \alpha_1 means the market overreacts more violently to
recent shocks.
The flat bottom at r_{t-1}=0 is the floor variance
\alpha_0: even after a perfectly quiet day the model keeps a baseline of
risk. That floor is what stops volatility from ever collapsing to zero.
Watching the band breathe
Here is an ARCH(1) series (\alpha_0=0.3,\ \alpha_1=0.6) with its
\pm 2\sigma_t band drawn around it. The band is not fixed: the instant a
large return lands, the very next day's band flares outward, and then — because each quiet day feeds a
smaller square into the recursion — it narrows back toward the baseline. That breathing band is the
conditional variance doing its job.
Notice the band widens one step after the shock, never before: ARCH is strictly backward
looking, reacting to yesterday's news, never anticipating tomorrow's.
Worked example: big shock versus small shock
Take an ARCH(1) with \alpha_0 = 0.2 and
\alpha_1 = 0.5 (returns in %).
- After a turbulent day with r_{t-1} = 4:
\sigma_t^2 = 0.2 + 0.5\,(4)^2 = 0.2 + 8.0 = 8.2, so
\sigma_t = \sqrt{8.2} \approx 2.86\%. Tomorrow is expected to be wild.
- After a calm day with r_{t-1} = 0.5:
\sigma_t^2 = 0.2 + 0.5\,(0.5)^2 = 0.2 + 0.125 = 0.325, so
\sigma_t = \sqrt{0.325} \approx 0.57\%. Tomorrow is expected to be quiet.
Same model, same parameters — a five-fold difference in tomorrow's expected volatility, driven entirely
by the size of yesterday's move. That is exactly the clustering we set out to capture.
Conditionally wild, unconditionally tame
A beautiful tension sits inside ARCH. Conditionally — given the recent past — the variance
jumps around; the process is heteroskedastic by construction. Yet unconditionally — averaged
over all history — the process is a well-behaved, stationary series with a single
finite variance. Taking the long-run expectation of both sides of the ARCH(q) recursion (and using that
\mathbb{E}[r_{t-i}^2] equals the constant unconditional variance in the
stationary state) gives:
- The long-run (unconditional) variance is
\operatorname{Var}(r_t) = \frac{\alpha_0}{1 - (\alpha_1 + \alpha_2 + \dots + \alpha_q)}.
- This is positive and finite — the process is
stationary —
only if \alpha_1 + \dots + \alpha_q < 1.
- If the coefficients sum to 1 or more, the variance is infinite and the
model is not covariance-stationary.
For the worked ARCH(1) above (\alpha_1 = 0.5) the long-run variance is
0.2/(1-0.5) = 0.4, i.e. an unconditional volatility of
\sqrt{0.4}\approx 0.63\%. Every wild and quiet day averages out to that.
ARCH captures clustering, but in practice it is clumsy: real volatility fades slowly, so a
single lagged square (ARCH(1)) forgets far too fast, and matching the long, gentle decay of the squared
returns' ACF needs a long recipe — ARCH(5), ARCH(8), sometimes more — each with its own
\alpha_i to estimate. That is a lot of parameters, they must all stay
non-negative, and the fits get unstable. The classic mistake is to keep bolting on more lags. The
fix is not more ARCH terms but a smarter recursion that lets past variance carry the memory —
the GARCH
model, which reproduces a long ARCH with just a handful of parameters. If you find yourself fitting
ARCH(8), you almost certainly want GARCH(1,1).
Before Engle, econometricians treated variance as a nuisance to be assumed constant and swept into a
standard error. Engle's leap, working on UK inflation data, was to make the variance a modelled
quantity in its own right — to say that uncertainty itself has dynamics you can estimate and forecast.
Suddenly "how risky is tomorrow?" became a question with a data-driven numerical answer, which is the
beating heart of option pricing, Value-at-Risk, and modern risk management. The 2003 Nobel citation
paired him with Clive Granger; the committee singled out ARCH as the tool that let economists
measure and predict volatility. Not bad for the idea of paying attention to the squares.