Aller au contenu

ÉconométrieFinanceTime seriesExplicationIntermédiaire

Par · Publiée le · 9 min de lecture

Sommaire
  1. Three parameters, one recursion
  2. Persistence, long-run variance and mean reversion
  3. Positivity and stationarity
  4. Quasi-maximum likelihood
  5. Check your understanding

Après cette fiche, vous saurez

  • Write the conditional variance recursion of a GARCH(1,1) and interpret ω\omega, α\alpha and β\beta.
  • Relate the persistence α+β\alpha+\beta to the unconditional variance, the half-life of a shock and the variance forecast.
  • State the positivity and stationarity constraints, and tell covariance stationarity from strict stationarity.
  • Describe quasi-maximum likelihood estimation and the role of robust standard errors.

Formulaire

Les équations de la fiche, d’un coup d’œil

  1. 01Return and innovationrt=μ+εt,εt=σt zt,zt∼i.i.d.(0, 1)r_t = \mu + \varepsilon_t, \qquad \varepsilon_t = \sigma_t\, z_t, \qquad z_t \overset{\text{i.i.d.}}{\sim} (0,\,1)
  2. 02GARCH(1,1) conditional varianceσt2=ω+α εt−12+β σt−12\sigma_t^2 = \omega + \alpha\,\varepsilon_{t-1}^2 + \beta\,\sigma_{t-1}^2
  3. 03Unconditional varianceσˉ2  =  E ⁣[εt2]  =  ω1−α−β,α+β<1\bar\sigma^2 \;=\; \mathbb{E}\!\left[\varepsilon_t^2\right] \;=\; \frac{\omega}{1-\alpha-\beta}, \qquad \alpha+\beta<1
  4. 04Forecast at horizon hEt ⁣[σt+h2]=σˉ2+(α+β) h−1(σt+12−σˉ2),h1/2=ln⁡(1/2)ln⁡(α+β)\mathbb{E}_t\!\left[\sigma_{t+h}^2\right] = \bar\sigma^2 + (\alpha+\beta)^{\,h-1}\left(\sigma_{t+1}^2-\bar\sigma^2\right), \qquad h_{1/2} = \frac{\ln(1/2)}{\ln(\alpha+\beta)}
  5. 05Gaussian quasi-log-likelihoodθ^=arg⁡max⁡θ  ∑t=1Tℓt(θ),ℓt(θ)=−12[ln⁡(2π)+ln⁡σt2(θ)+εt2(θ)σt2(θ)]\hat\theta = \arg\max_{\theta}\; \sum_{t=1}^{T} \ell_t(\theta), \qquad \ell_t(\theta) = -\frac{1}{2}\left[\ln(2\pi) + \ln \sigma_t^2(\theta) + \frac{\varepsilon_t^2(\theta)}{\sigma_t^2(\theta)}\right]

Demonstration example — not reviewed by the author

Financial econometrics

The GARCH(1,1) model and volatility dynamics

Financial returns are close to unpredictable in level, but their magnitude is not: turbulent periods follow turbulent periods. Engle's ARCH model1, generalised to GARCH by Bollerslev2, captures this by making today's variance a function of yesterday's shock and yesterday's variance.

The model

Three parameters, one recursion

Return and innovation
rt=μ+εt,εt=σt zt,zt∼i.i.d.(0, 1)r_t = \mu + \varepsilon_t, \qquad \varepsilon_t = \sigma_t\, z_t, \qquad z_t \overset{\text{i.i.d.}}{\sim} (0,\,1)

The shock εt\varepsilon_t is a standardised noise ztz_t scaled by the conditional standard deviation σt\sigma_t, which is known at date t−1t-1.

Variables
SymbolMeaningUnit
rtr_t(log) return over period tt% per period
μ\muconditional mean, constant here
εt\varepsilon_tinnovation: unexpected return
σt\sigma_tconditional standard deviation (volatility), a function of information at t−1t-1
ztz_ti.i.d. noise with zero mean and unit variance
GARCH(1,1) conditional variance
σt2=ω+α εt−12+β σt−12\sigma_t^2 = \omega + \alpha\,\varepsilon_{t-1}^2 + \beta\,\sigma_{t-1}^2

Today's variance combines a baseline level, yesterday's news (squared shock) and the memory of yesterday's variance. Setting β=0\beta=0 gives back ARCH(1).

Variables
SymbolMeaningUnit
σt2\sigma_t^2conditional variance of εt\varepsilon_t given the past%² per period
ω\omegaintercept, ω>0\omega>0
α\alphareaction to the previous shock (ARCH effect), α≥0\alpha\ge 0
β\betainertia of the variance (GARCH effect), β≥0\beta\ge 0

Sources:2

The volatility feedback loop

Loop diagram: the constant omega, yesterday's squared shock weighted by alpha and yesterday's variance weighted by beta add up to today's conditional variance; that variance scales a standardised noise to produce the innovation and the return; the squared innovation and the variance flow back to the next date. A second group shows estimation: observed returns, Gaussian likelihood, optimisation, estimated parameters and robust standard errors.INFORMATION AT T−1DATE TESTIMATIONωbaseline levelε²(t−1)yesterday's newsσ²(t−1)yesterday's varianceσ²(t)conditional variancez(t)i.i.d. noise (0, 1)ε(t) = σ(t)·z(t)innovationr(t) = μ + ε(t)observed returnReturns r(1)…r(T)Gaussianquasi-likelihoodNumerical maximisationunder constraintsEstimated ω, α, βrobust standard errors+ ω× α× βscalessquared → t+1→ t+1observed
  • Solid arrow: computation at date t
  • Dashed arrow: carried over to the next date
  • Amber: shock channel (α)
  • Blue: memory channel (β)
  • Violet: estimation from the data
At each date the variance is computed from the past, then a shock is drawn; that shock and that variance feed the next date.
  • level

    ω

    Sets, together with α+β\alpha+\beta, the average level of the variance. Must be strictly positive.

  • reaction

    α

    How strongly a recent shock lifts the variance. A high α\alpha gives a jumpy volatility.

  • memory

    β

    Share of yesterday's variance carried into today. A high β\beta gives a smooth volatility.

  • persistence

    α + β

    How fast a variance shock dies out. Close to 1: turbulent episodes last.

Dynamics

Persistence, long-run variance and mean reversion

Unconditional variance
σˉ2  =  E ⁣[εt2]  =  ω1−α−β,α+β<1\bar\sigma^2 \;=\; \mathbb{E}\!\left[\varepsilon_t^2\right] \;=\; \frac{\omega}{1-\alpha-\beta}, \qquad \alpha+\beta<1

Take expectations of the recursion using E[εt−12]=E[σt−12]=σˉ2\mathbb{E}[\varepsilon_{t-1}^2]=\mathbb{E}[\sigma_{t-1}^2]=\bar\sigma^2. The formula only makes sense when α+β<1\alpha+\beta<1.

Variables
SymbolMeaningUnit
σˉ2\bar\sigma^2long-run variance towards which σt2\sigma_t^2 reverts%² per period
α+β\alpha+\betapersistence of the variance

Sources:2

Forecast at horizon h
Et ⁣[σt+h2]=σˉ2+(α+β) h−1(σt+12−σˉ2),h1/2=ln⁡(1/2)ln⁡(α+β)\mathbb{E}_t\!\left[\sigma_{t+h}^2\right] = \bar\sigma^2 + (\alpha+\beta)^{\,h-1}\left(\sigma_{t+1}^2-\bar\sigma^2\right), \qquad h_{1/2} = \frac{\ln(1/2)}{\ln(\alpha+\beta)}

Result derived in this sheet

Obtained by iterating Et[σt+h2]=ω+(α+β) Et[σt+h−12]\mathbb{E}_t[\sigma_{t+h}^2]=\omega+(\alpha+\beta)\,\mathbb{E}_t[\sigma_{t+h-1}^2]. The gap to the long-run variance decays geometrically; h1/2h_{1/2} is the number of periods after which it is halved.

Variables
SymbolMeaningUnit
Et[⋅]\mathbb{E}_t[\cdot]expectation conditional on information available at tt
σt+12\sigma_{t+1}^2next-period variance, known at tt
hhforecast horizonperiods
h1/2h_{1/2}half-life of a variance shockperiods

Forecast variance reverting to its long-run level

Two decreasing curves starting at 3 and converging to 1. With a persistence of 0.90 the curve is almost back to 1 after about thirty periods; with a persistence of 0.98 it is still near 1.6 at horizon 59.11.522.531020304050Forecast horizon h (periods)Forecast variance / long-run variancelong-run varianceα + β = 0.90α + β = 0.98ILLUSTRATIVE DATA
Two decreasing curves starting at 3 and converging to 1. With a persistence of 0.90 the curve is almost back to 1 after about thirty periods; with a persistence of 0.98 it is still near 1.6 at horizon 59.1232040Forecast horizon h (periods)Forecast variance / long-run variancelong-run varianceα + β = 0.90α + β = 0.98ILLUSTRATIVE DATA
Curves computed from the forecast formula, starting from a variance three times the long-run variance (normalised to 1). Values chosen to show the shape, not estimated from data.
View the data
Two decreasing curves starting at 3 and converging to 1. With a persistence of 0.90 the curve is almost back to 1 after about thirty periods; with a persistence of 0.98 it is still near 1.6 at horizon 59. — Illustrative data
SeriesForecast horizon h (periods)Forecast variance / long-run variance
α + β = 0.9013
α + β = 0.9032.62
α + β = 0.9052.3122
α + β = 0.9072.0629
α + β = 0.9091.8609
α + β = 0.90111.6974
α + β = 0.90131.5649
α + β = 0.90151.4575
α + β = 0.90171.3706
α + β = 0.90191.3002
α + β = 0.90211.2432
α + β = 0.90231.197
α + β = 0.90251.1595
α + β = 0.90271.1292
α + β = 0.90291.1047
α + β = 0.90311.0848
α + β = 0.90331.0687
α + β = 0.90351.0556
α + β = 0.90371.0451
α + β = 0.90391.0365
α + β = 0.90411.0296
α + β = 0.90431.0239
α + β = 0.90451.0194
α + β = 0.90471.0157
α + β = 0.90491.0127
α + β = 0.90511.0103
α + β = 0.90531.0083
α + β = 0.90551.0068
α + β = 0.90571.0055
α + β = 0.90591.0044
α + β = 0.9813
α + β = 0.9832.9208
α + β = 0.9852.8447
α + β = 0.9872.7717
α + β = 0.9892.7015
α + β = 0.98112.6341
α + β = 0.98132.5694
α + β = 0.98152.5073
α + β = 0.98172.4476
α + β = 0.98192.3903
α + β = 0.98212.3352
α + β = 0.98232.2823
α + β = 0.98252.2316
α + β = 0.98272.1828
α + β = 0.98292.136
α + β = 0.98312.091
α + β = 0.98332.0478
α + β = 0.98352.0063
α + β = 0.98371.9664
α + β = 0.98391.9282
α + β = 0.98411.8914
α + β = 0.98431.8561
α + β = 0.98451.8222
α + β = 0.98471.7896
α + β = 0.98491.7584
α + β = 0.98511.7283
α + β = 0.98531.6995
α + β = 0.98551.6718
α + β = 0.98571.6452
α + β = 0.98591.6196

Simulate a GARCH(1,1) path

Synthetic data — seed 7

rt=σtztr_t = \sigma_t z_t with zt∼N(0,1)z_t \sim \mathcal{N}(0, 1) and σt2=ω+α rt−12+β σt−12\sigma_t^2 = \omega + \alpha\, r_{t-1}^2 + \beta\, \sigma_{t-1}^2. The long-run variance ω/(1−α−β)\omega/(1 - \alpha - \beta) exists only when α+β<1\alpha + \beta < 1: beyond that, β\beta is reduced automatically. Gaussian innovations, zero mean: real returns have fatter tails.

Interactive simulation: a series of simulated returns and, below it, the matching conditional volatility with its long-run level; sliders set omega, alpha, beta and the number of periods, and high-amplitude periods appear grouped together.

Returns r_t−6−4−2024650100150200250300350400Period tReturnSYNTHETIC DATA
Conditional volatility σ_t0.511.522.550100150200250300350400Period tσₜlong-run σSYNTHETIC DATA
Persistence α + β
0.98
Unconditional variance ω/(1−α−β)
2.5
Unconditional standard deviation
1.5811
Sample variance
1.6963
Shock half-life (periods)
34.3
Synthetic data: returns drawn from the model with the chosen parameters and a fixed seed. Move α+β\alpha+\beta towards 1 to see the clusters lengthen.

Conditions

Positivity and stationarity

Parameter constraints of a GARCH(1,1) and what they guarantee.
PropertyConditionConsequence
Positivityω>0, α≥0, β≥0\omega>0,\ \alpha\ge 0,\ \beta\ge 0σt2>0\sigma_t^2>0 at every date, whatever the path of shocks.
Covariance stationarityα+β<1\alpha+\beta<1Finite unconditional variance ω/(1−α−β)\omega/(1-\alpha-\beta); variance forecasts converge2.
Strict stationarityE ⁣[ln⁡(β+αzt2)]<0\mathbb{E}\!\left[\ln(\beta+\alpha z_t^2)\right]<0Weaker than the previous condition: it allows α+β≥1\alpha+\beta\ge 13.
Finite fourth moment (Gaussian ztz_t)3α2+2αβ+β2<13\alpha^2+2\alpha\beta+\beta^2<1Kurtosis 3 1−(α+β)21−(α+β)2−2α2>33\,\dfrac{1-(\alpha+\beta)^2}{1-(\alpha+\beta)^2-2\alpha^2}>3: tails fatter than the normal distribution2.

Estimation

Quasi-maximum likelihood

Gaussian quasi-log-likelihood
θ^=arg⁡max⁡θ  ∑t=1Tℓt(θ),ℓt(θ)=−12[ln⁡(2π)+ln⁡σt2(θ)+εt2(θ)σt2(θ)]\hat\theta = \arg\max_{\theta}\; \sum_{t=1}^{T} \ell_t(\theta), \qquad \ell_t(\theta) = -\frac{1}{2}\left[\ln(2\pi) + \ln \sigma_t^2(\theta) + \frac{\varepsilon_t^2(\theta)}{\sigma_t^2(\theta)}\right]

The likelihood is maximised as if ztz_t were Gaussian. If the conditional mean and variance are correctly specified, the estimator remains consistent even when ztz_t is not; only the standard errors need correcting4.

Variables
SymbolMeaningUnit
θ\thetaparameter vector (μ,ω,α,β)(\mu,\omega,\alpha,\beta)
ℓt(θ)\ell_t(\theta)contribution of observation tt to the log-likelihood
εt(θ)\varepsilon_t(\theta)residual rt−μr_t-\mu
σt2(θ)\sigma_t^2(\theta)variance obtained by running the recursion from a starting value
TTnumber of observations

Sources:4

  1. Filter the mean

    Compute the residuals εt=rt−μ\varepsilon_t=r_t-\mu (or those of a richer mean model).

  2. Initialise and run the recursion

    Set σ12\sigma_1^2, for instance to the sample variance of the residuals, then compute σt2\sigma_t^2 recursively.

  3. Maximise the quasi-likelihood

    Numerical optimisation under the constraints ω>0\omega>0, α,β≥0\alpha,\beta\ge 0.

  4. Robust standard errors

    Use the Bollerslev–Wooldridge "sandwich" covariance matrix, valid without normality4.

  5. Check

    Standardised residuals z^t=ε^t/σ^t\hat z_t=\hat\varepsilon_t/\hat\sigma_t and their squares should no longer be autocorrelated.

Takeaways

Check your understanding

QuestionA GARCH(1,1) has parameters ω=0.02\omega=0.02, α=0.08\alpha=0.08 and β=0.90\beta=0.90. What is its unconditional variance?

  1. A1
  2. B0.02
  3. C0.98
  4. DIt is infinite
Show the answer
  • A — Correct. ω/(1−α−β)=0.02/0.02=1\omega/(1-\alpha-\beta)=0.02/0.02=1.
  • B — Incorrect. That is ω\omega, the intercept of the recursion, not the long-run variance.
  • C — Incorrect. That is the persistence α+β\alpha+\beta.
  • D — Incorrect. It would be infinite only if α+β≥1\alpha+\beta\ge 1.
Confusing conditional and unconditional variance

σt2\sigma_t^2 changes at every date and depends on the past; σˉ2\bar\sigma^2 is a constant. A stationary GARCH is unconditionally homoskedastic and conditionally heteroskedastic.

Believing the model predicts the sign of returns

GARCH describes the expected magnitude, not the direction: Et−1[εt]=0\mathbb{E}_{t-1}[\varepsilon_t]=0 whatever σt\sigma_t.

Reading α + β ≥ 1 as an explosive path

The condition α+β<1\alpha+\beta<1 is about the existence of the variance. Strict stationarity obeys a weaker condition3.

Using Gaussian maximum-likelihood standard errors uncorrected

If the standardised innovations are not Gaussian, the inverse Hessian no longer estimates the variance of the estimator correctly: the robust form is required4.

Forgetting the time unit

Parameters depend on the sampling frequency: a GARCH estimated on daily returns does not apply as is to monthly returns.

References

  1. Robert F. Engle (1982). Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation. Econometrica, 50(4), 987. doi:10.2307/1912773
  2. Tim Bollerslev (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 31(3), 307–327. doi:10.1016/0304-4076(86)90063-1
  3. Daniel B. Nelson (1990). Stationarity and Persistence in the GARCH(1,1) Model. Econometric Theory, 6(3), 318–334. doi:10.1017/S0266466600005296
  4. Tim Bollerslev, Jeffrey M. Wooldridge (1992). Quasi-maximum likelihood estimation and inference in dynamic models with time-varying covariances. Econometric Reviews, 11(2), 143–172. doi:10.1080/07474939208800229
  5. Christian Francq, Jean-Michel Zakoian (2019). GARCH Models. Wiley. doi:10.1002/9781119313472

Aussi disponible en français

Partager cette fiche

Mots-clés

En vedetteIARecherche d’information

Le pipeline RAG, de l’indexation à la réponse

La génération augmentée par récupération donne à un modèle de langage des passages pertinents au moment de répondre. Sa qualité se joue d’abord à la récupération : ce qui n’est pas retrouvé ne peut pas être cité.

Intermédiaire · 11 min de lecture ·

En vedetteIAApprentissage profond

Mélange d’experts et routage top-k

Un mélange d’experts confie chaque entrée à quelques sous-réseaux choisis par un routeur. Le nombre de paramètres augmente avec le nombre d’experts, mais le calcul par jeton ne dépend que de kk.

Avancé · 11 min de lecture ·

LLM agentsContext engineeringEN

Keeping the global agent lean: memory compression and sub-agents

LLM agents re-read an ever-growing context of thoughts, tool calls and tool results. This sheet maps two ways to keep the global agent's window bounded: compressing what is already inside (compaction, tool-result clearing, external memory, prompt compression) and isolating exploration in sub-agents that return condensed reports. It includes a token model and trade-offs.

Avancé · 12 min de lecture ·