Time Series Cheat Sheet Assumptions and properties Trends and seasonality

By Marcelo Moreno - Universidad Rey Juan Carlos

OLS model assumptions under time series Spurious regression - is when the relation between y and
The Econometrics Cheat Sheet Project
Under this assumptions, the OLS estimator will present x is due to factors that affect y and have correlation with
good properties. Gauss-Markov assumptions extended x, Corr(xj , u) ̸= 0. Is the non-fulfillment of t3.
Basic concepts applied to time series: Trends
Definitions t1. Parameters linearity and weak dependence. Two time series can have the same (or contrary) trend, that
Time series - is a succession of quantitative observations a. yt must be a linear function of the β’s. should lend to a high level of correlation. This, can provoke
of a phenomena ordered in time. b. The stochastic {(xt , yt ) : t = 1, 2, . . . , T } is station- a false appearance of causality, the problem is spurious
There are some variations of time series: ary and weakly dependent. regression. Given the model:
ˆ Panel data - consist of a time series for each observation t2. No perfect collinearity. y t = β0 + β1 x t + u t
of a cross section. ˆ There are no independent variables that are constant: where:
ˆ Pooled cross sections - combines cross sections from Var(xj ) ̸= 0, ∀j = 1, . . . , k yt = α0 + α1 Trend + vt
different time periods. ˆ There is not an exact linear relation between inde- xt = γ0 + γ1 Trend + vt
Stochastic process - sequence of random variables that pendent variables. Adding a trend to the model can solve the problem:
are indexed in time. t3. Conditional mean zero and correlation zero. yt = β0 + β1 xt + β2 Trend + ut
a. There are no systematic errors: E(u | x1 , . . . , xk ) = The trend can be linear or non-linear (quadratic, cubic,
Components of a time series E(u) = 0 → strong exogeneity (a implies b).
ˆ Trend - is the long-term general movement of a series. exponential, etc.)
b. There are no relevant variables left out of the model: Another way is make use of the Hodrick-Prescott filter
ˆ Seasonal variations - are periodic oscillations that are Cov(xj , u) = 0, ∀j = 1, . . . , k → weak exogene- to extract the trend and the cyclical component.
produced in a period equal or inferior to a year, and can ity.
be easily identified on different years (usually are the re- t4. Homoscedasticity. The variability of the residuals is Seasonality
sult of climatology reasons). the same for any x: Var(u | x1 , . . . , xk ) = σu2 A time series with can man- y
ˆ Cyclical variations - are periodic oscillations that are t5. No autocorrelation. Residuals do not contain infor- ifest seasonality. That is,
produced in a period greater than a year (are the result mation about any other residuals: the series is subject to a sea-
of the economic cycle). Corr(ut , us | x1 , . . . , xk ) = 0, ∀t ̸= s sonal variations or pattern,
ˆ Residual variations - are movements that do not fol- t6. Normality. Residuals are independent and identically usually related to climatol-
low a recognizable periodic oscillation (are the result of distributed (i.i.d. so on): u ∼ N (0, σu2 ) ogy conditions.
eventual phenomena). t7. Data size. The number of observations available must For example, GDP (black)
Type of time series models be greater than (k + 1) parameters to estimate. (It is is usually higher in summer
ˆ Static models - the relation between y and x’s is con- already satisfied under asymptotic situations) and lower in winter. Season-
t
temporary. Conceptually: ally adjusted series (red) for
Asymptotic properties of OLS comparison.
yt = β0 + β1 xt + ut Under the econometric model assumptions and the Central ˆ This problem is spurious regression. A seasonal ad-
ˆ Distributed-lag models - the relation between y and Limit Theorem:
x’s is not contemporary. Conceptually: justment can solve it.
ˆ Hold t1 to t3a: OLS is unbiased. E(β̂j ) = βj A simple seasonal adjustment could be creating station-
yt = β0 + β1 xt + β2 xt−1 + · · · + βs xt−(s−1) + ut
ˆ Hold t1 to t3: OLS is consistent. plim(β̂j ) = βj (to t3b ary binary variables and adding them to the model. For
The long term cumulative effect in y when ∆x is:
left out t3a, weak exogeneity, biased but consistent) example, for quarterly series (Qqt are binary variables):
β1 + β2 + · · · + β s
ˆ Hold t1 to t5: asymptotic normality of OLS (then, t6 yt = β0 + β1 Q2t + β2 Q3t + β3 Q4t + β4 x1t + · · · + βk xkt + ut
ˆ Dynamic models - lags of the dependent variable (en-
is necessarily satisfied): u ∼ N (0, σu2 ) Another way is to seasonally adjust (sa) the variables, and
dogeneity). Conceptually: a
yt = β0 + β1 yt−1 + · · · + βs yt−s + ut ˆ Hold t1 to t5: unbiased estimate of σu2 . E(σ̂u2 ) = σu2 then, do the regression with the adjusted variables:
sa
ˆ Combinations of the above, like the rational distributed- ˆ Hold t1 to t5: OLS is BLUE (Best Linear Unbiased Es- zt = β0 + β1 Q2t + β2 Q3t + β3 Q4t + vt → v̂t + E(zt ) = ẑt
sa sa sa
lag models (distributed-lag + dynamic). timator) or efficient. ŷt = β0 + β1 x̂1t + · · · + βk x̂kt + ut
ˆ Hold t1 to t6: hypothesis testing and confidence intervals There are much better and complex methods to seasonally
can be done reliably. adjust a time series, like the X-13ARIMA-SEATS.
TS-25.01-EN - github.com/marcelomijas/econometrics-cheatsheet - CC-BY-4.0 license
Autocorrelation – MA(q) process. ACF: only the first q coefficients Under H0 : No autocorrelation:
are significant, the remaining are abruptly canceled. T · Rû2 t ∼ χ2q or T · Rû2 t ∼ χ2p
The residual of any observation, ut , is correlated with the a a
PACF: attenuated exponential fast decay or sine * H1 : Autocorrelation of order q (or p).
residual of any other observation. The observations are not waves.
independent. Is the non-fulfillment of t5. – Ljung-Box Q test:
– AR(p) process. ACF: attenuated exponential fast * H1 : There is autocorrelation.
Corr(ut , us | x1 , . . . , xk ) = Corr(ut , us ) ̸= 0, ∀t ̸= s decay or sine waves. PACF: only the first p coeffi-
Consequences cients are significant, the remaining are abruptly can- Correction
ˆ OLS estimators still are unbiased. celed. ˆ Use OLS with a variance-covariance matrix estimator
ˆ OLS estimators still are consistent. – ARMA(p, q) process. ACF and PACF: the coef- that is robust to heterocedasticity and autocorre-
ˆ OLS is not efficient anymore, but still a LUE (Linear ficients are not abruptly canceled and presents a fast lation (HAC), for example, the one proposed by Newey-
Unbiased Estimator). decay. West.
ˆ Variance estimations of the estimators are biased: If the ACF coefficients do not decay rapidly, there is ˆ Use Generalized Least Squares (GLS). Supposing
the construction of confidence intervals and the hypoth- a clear indicator of lack of stationarity in mean, which yt = β0 + β1 xt + ut , with ut = ρut−1 + εt , where |ρ| < 1
esis testing is not reliable. would lead to take first differences in the original series. and εt is white noise.
ˆ Formal tests - Generally, H0 : No autocorrelation. – If ρ is known, use a quasi-differentiated model:
Detection Supposing that ut follows an AR(1) process: yt − ρyt−1 = β0 (1 − ρ) + β1 (xt − ρxt−1 ) + ut − ρut−1
ˆ Scatter plots - look for scatter patterns on ut−1 vs. ut . ut = ρ1 ut−1 + εt yt∗ = β0∗ + β1′ x∗t + εt
Ac. Ac. + Ac. − ′
where εt is white noise. where β1 = β1 ; and estimate it by OLS.
ut ut ut
– AR(1) t test (exogenous regressors): – If ρ is not known, estimate it by -for example-
ρ̂1 the Cochrane-Orcutt iterative method (Prais-
ut−1 ut−1 t = se(ρ̂1 ) ∼ tT −k−1,α/2
* H1 : Autocorrelation of order one, AR(1). Winsten method is also good):
ut−1 1. Obtain ût from the original model.
– Durbin-Watson statistic (exogenous regressors and
residual normality): 2. Estimate ût = ρût−1 + εt and obtain ρ̂.
Pn
(û −û )2 3. Create a quasi-differentiated model:
ˆ Correlogram - composed – Y axis: correlation [-1, 1]. d = t=2Pnt û2t−1 ≈ 2 · (1 − ρ̂1 ), 0 ≤ d ≤ 4
t=1 t y t − ρ̂yt−1 = β0 (1 − ρ̂) + β1 (xt − ρ̂xt−1 ) + ut − ρ̂ut−1
of the autocorrelation func- – X axis: lag number. * H1 : Autocorrelation of order one, AR(1). yt∗ = β0∗ + β1′ x∗t + εt
0.5
tion (ACF) and the partial – Blue lines: ±1.96/T d = 0 2 4 ′
where β1 = β1 ; and estimate it by OLS.
ACF (PACF). ρ ≈ 1 0 -1
ACF f (d)
4. Obtain û∗t = yt − (β̂0∗ + β̂1′ xt ) ̸= yt − (β̂0∗ + β̂1′ x∗t ).

IN

IN
1 5. Repeat from step 2. The algorithm ends when the
CO

CO
Rej. H0 Accept H0 Rej. H0
estimated parameters vary very little between iter-
NC

NC
AR + LU No AR AR − ations.

LU
0
ˆ If not solved, look for high dependence in the series.
S

S
IV

IV
E

E
-1
PACF 0 dL dU 2 4−
d
4−
d
4
Exponential smoothing
1 U L
ft = αyt + (1 − α)ft−1
– Durbin’s h (endogenous regressors):
q where 0 < α < 1 is the smoothing parameter.
T
0 h = ρ̂ · 1−T ·υ
where υ is the estimated variance of the coefficient as- Predictions
sociated to the endogenous variable.
-1 Two types of predictions:
* H1 : Autocorrelation of order one, AR(1).
Conclusions differ between autocorrelation processes. ˆ Of the mean value of y for a specific value of x.
– Breusch-Godfrey test (endogenous regressors): it
ˆ Of an individual value of y for a specific value of x.
can detect MA(q) and AR(p) processes (εt is w. noise):
* MA(q): ut = εt − m1 ut−1 − · · · − mq ut−q
* AR(p): ut = ρ1 ut−1 + · · · + ρp ut−p + εt
TS-25.01-EN - github.com/marcelomijas/econometrics-cheatsheet - CC-BY-4.0 license
Stationarity Unit roots From unit root to percentage change
When an I(1) series is strictly positive, it is usually con-
Stationarity allows to correctly identify relations –that stay A process is I(d), that is, integrated of order d, if applying verted to logarithms before taking the first difference to
unchanged with time– between variables. differences d times makes the process stationary. obtain the (approx.) percentage change of the series:
ˆ Stationary process (strict stationarity) - if any collec- When d ≥ 1, the process is called a unit root process or yt − yt−1
tion of random variables is taken and shifted h periods it is said to have an unit root. ∆ log(yt ) = log(yt ) − log(yt−1 ) ≈
yt−1
(time changes), the joint probability distribution should A process have an unit root when the stability condition is
stay unchanged. not met (there are roots on the unit circle). Cointegration
ˆ Non-stationary process - for example, a series with
trend, where at least the mean changes with time.
Strong dependence When two series are I(1), but a linear combination of
Most of the time, economics series are strongly dependent
ˆ Covariance stationary process - its a weaker form of them is I(0). If the case, the regression of one series over
(or high persistent). Some examples of unit root I(1): the other is not spurious, but expresses something about
stationarity:
ˆ Random walk - an AR(1) process with ρ1 = 1. the long term relation. Variables are called cointegrated if
– E(xt ) is constant. – Var(xt ) is constant.
yt = yt−1 + et they have a common stochastic trend.
– For any t, h ≥ 1, the Cov(xt , xt+h ) depends only of h,
where {et : t = 1, 2, . . . , T } is an i.i.d. sequence with zero For example, {xt } and {yt } are I(1), but yt − βxt = ut
not of t.
mean and σe2 variance. where {ut } is I(0). (β is the cointegrating parameter).
ˆ Random walk with a drift - an AR(1) process with
Weak dependence ρ1 = 1 and a constant. Cointegration test
Weak dependence replaces the random sampling assump- yt = β0 + yt−1 + et Following the example above:
tion for time series. where {et : t = 1, 2, . . . , T } is an i.i.d. sequence with zero 1. Estimate yt = α + βxt + εt and obtain ε̂t .
ˆ An stationary process {xt } is weakly dependent when mean and σe2 variance. 2. Perform an ADF test on ε̂t with a modified distribution.
xt and xt+h are almost independent as h increases with- Unit root tests The result of this test is equivalent to:
out a limit. ˆ H0 : β = 0 (no cointegration)
ˆ A covariance stationary process is weakly dependent
Test H0 Reject H0 ˆ H1 : β ̸= 0 (cointegration)
ADF I(1) tau < Critical value if test statistic > critical value, reject H0 .
if the correlation between xt and xt+h tends to 0 fast
enough when h → ∞ (they are not asymptotically corre- KPSS I(0) level mu > Critical value
lated). I(0) trend tau > Critical value Heterocedasticity on time series
Weakly dependent processes are known as integrated of Phillips-Perron I(1) Z-tau < Critical value
Zivot-Andrews I(1) tau < Critical value The assumption affected is t4, which leads OLS to be
order zero, I(0). Some examples: not efficient.
ˆ Moving average - {xt } is a moving average of order q, From unit root to weak dependence Use tests like Breusch-Pagan or White’s, where H0 : No het-
MA(q): Integrated of order one, I(1), means that the first dif- erocedasticity. It is important for the tests to work that
xt = et + m1 et−1 + · · · + mq et−q ference of the process is weakly dependent or I(0) (and there is no autocorrelation.
where {et : t = 0, 1, . . . , T } is an i.i.d. sequence with zero usually, stationary). For example, let {yt } be a random
mean and σe2 variance. walk: ARCH
ˆ Autoregressive process - {xt } is an autoregressive pro- ∆yt = yt − yt−1 = et y, ∆y An autoregressive conditional heterocedasticity (ARCH),
cess of order p, AR(p): where {et } = {∆yt } is i.i.d. is a model to analyze a form of dynamic heterocedasticity,
xt = ρ1 xt−1 + · · · + ρp xt−p + et where the error variance follows an AR(p) process.
where {et : t = 1, 2, . . . , T } is an i.i.d. sequence with zero Note: Given the model: yt = β0 + β1 zt + ut where, there is AR(1)
mean and σe2 variance. ˆ The first difference of a se- and heterocedasticity:
Stability condition: if 1−ρ1 z−· · ·−ρp z p = 0 for |z| > 1 ries removes its trend. E(u2t | ut−1 ) = α0 + α1 u2t−1
then {xt } is an AR(p) stable process that is weakly de- ˆ Logarithms of a series sta- GARCH
pendent. For AR(1), the condition is: |ρ1 | < 1. bilizes its variance. A general autoregressive conditional heterocedasticity
ˆ ARMA process - is a combination of AR(p) and (GARCH), is a model similar to ARCH, but in this case,
MA(q); {xt } is an ARMA(p, q): t
the error variance follows an ARMA(p, q) process.
xt = et + m1 et−1 + · · · + mq et−q + ρ1 xt−1 + · · · + ρp xt−p

TS-25.01-EN - github.com/marcelomijas/econometrics-cheatsheet - CC-BY-4.0 license