Time series analysis is a statistical technique that studies time series data. A time series is a sequence of data points indexed in discrete-time order.

A time series is a series of observations xt , observed over a period of time. Typically the observations can be over an entire interval, randomly sampled on an interval or at fixed time points. Different types of time sampling require different approaches to the data analysis.

**The Southern Oscillation Index from 1876-present**

The Southern Oscillation Index (SOI) is an indicator of intensity of the El Nino effect (see

wiki). The SOI measures the fluctuations in air surface pressures between Tahiti and Darwin.

In Figure 1.2 we give a plot of monthly SOI from January 1876 – July 2014 (note that there is some doubt on the reliability of the data before 1930).

A univariate time series is a stochastic (a fancy word for random) process that consists of a sequence of time indexed random variables 𝑋𝑡 = {𝑋1, 𝑋2, … }, where 𝑋1 denotes the random variable at time 𝑡 = 1, 𝑋2 the random variable at time 𝑡 = 2 and so on. The observed values of such a stochastic process, 𝑋1 = 𝑥1, 𝑋2 = 𝑥2 and so on is called a realization of the stochastic process. This sequence {𝑋𝑡} = {𝑥1, 𝑥2, … } is also called a time series. It will be clear from the context whether we are referring to the stochastic process or a particular realization of the process.

We have established that one of the main features that distinguish time series analysis from classical

methods is that observations taken over time (a time series) can be dependent and this dependency

tends to decline the further apart in time these two observations. However, to do any sort of analysis

of this time series we have to assume some sort of invariance in the time series, for example the mean

or variance of the time series does not change over time. If the marginal distributions of the time

series were totally different no sort of inference would be possible (suppose in classical statistics you

were given independent random variables all with different distributions, what parameter would you be estimating, it is not possible to estimate anything!).

There are two definitions of stationarity, weak stationarity which only concerns the covariance of a

process and strict stationarity which is a much stronger condition and supposes the distributions

are invariant over time.

Definition 3.3.1 (Strict stationarity)

The time series {Xt} is said to be strictly stationary if for any finite sequence of integers t1, . . . , tk and shift h the distribution of (Xt1, . . . , Xtk) and (Xt1+h, . . . , Xtk+h) are the same.

The above assumption is often considered to be rather strong (and given a data it is very hard to check). Often it is possible to work under a weaker assumption called weak/second order stationarity.

**Definition 3.3.2 (Second order stationarity/weak stationarity)**

The time series {Xt} is said to be second order stationary if the mean is constant for all t and if for any t and k the covariance between Xt and Xt+k only depends on the lag difference k. In other words there exists a function

c : Z → R such that for all t and k we have

** c(k) = cov(Xt, Xt+k). **

# __Covariance and Correlation:__

## • The covariance is defined as

## **cov(X, Y ) = E ((X − E(X))(Y − E(Y ))) = E(XY ) − E(X)E(Y ).**

## • The variance is **var(X) = E(X − E(X))^2= E(X^2) = E(X)^2.**

## • Observe **var(X) = cov(X, X).**

• Rules of covariances. If a, b, c are finite constants and X, Y, Z are random variables with E(X2) < ∞, E(Y2) < ∞ and E(Z2) < ∞ (which immediately implies their means are finite). Then the covariance satisfies the linearity property

** cov (aX + bY + c, Z) = acov(X, Z) + bcov(Y, Z)**.

Observe the shift c plays no role in the covariance (since it simply shifts the data).

• The variance of vectors. Suppose that A is a matrix and X a random vector with variance/- covariance matrix Σ. Then

** var(AX) = Avar(X)A’ = AΣA**

which can be proved using the linearity property of covariances.

• The correlation between X and Y is

** cor(X, Y ) = cov(X, Y )/((var(X)var(Y ))^0.5) **

and lies between [−1, 1]. If var(X) = var(Y ) then cor(X, Y ) is the coefficient of the best linear predictor of X given Y and visa versa.

There is a special time series called white noise. Let {𝜀𝑡} be a white noise process, then

𝜀𝑡

satisfies

1. E(𝜀𝑡) = 𝜇 < ∞, ∀𝑡

2. var(𝜀𝑡) = 𝜎2 < ∞, ∀𝑡

3. cov(𝜀𝑡, 𝜀𝑠) = 0, ∀𝑡 ≠ s

A random walk is a stochastic process that is a sum of random steps. A summation of 𝜀𝑡 , 𝑊𝑡 = ∑ 𝜀𝑖 𝑡 𝑖=1 , is a random walk. Each next value 𝑊𝑡+1 is a random step away from the previous value. When E(𝜀𝑡) = 0, 𝑊𝑡 is called a random-walk-without-drift model. A Random Walk Down Wall Street suggests that the market is a random walk. This is consistent with the efficient-market hypothesis that past information does not help predicting the future.

The Ljung-Box test is a hypothesis test for zero autocorrelations for a time series or a series of residuals. It tests whether any in a group of autocorrelations of a time series is different from zero. The sample ACF 𝜌̂ are calculated for a number of lags that we want to test for. The null hypothesis is that the first 𝑚 autocorrelations are jointly zero.

White noise process is very important in time series analysis. It is a fundamental building block of any time series. The Wold decomposition theorem says that for any covariance stationary stochastic process {𝑋𝑡}, it can be written as the sum of a deterministic time series and a stochastic time series. Mathematically, if {𝑋𝑡} is a

covariance stationary process and {𝜀𝑡} a white noise zero-mean process, then there exists a unique linear representation

**𝑋𝑡 = 𝜇 + 𝑉𝑡 + 𝜀𝑡 + 𝜓1𝜀𝑡−1 + 𝜓2𝜀𝑡−2 + ⋯ = 𝜇 + 𝑉𝑡 + ∑𝜓𝑗𝜀𝑡−𝑗 (summation for j=0 to infi.)**

The determinist component can be a trend or a seasonal component or both. In general,a time series can be decomposed into three components:

1. trend component

2. seasonal component

3. irregular random component

Trend component exists when there is a long-term increase or decrease in the data. For instance, S&P 500 increases with time. Seasonal component exists when the data exhibit patterns with seasons like months or quarters. For instance, the number of visitors to Disneyland changes with seasons (more people in summer and fewer in winter). Irregular random component, a.k.a. stochastic component or white noise, is the

innovations with time. That is the source of uncertainty of a time series. For instance, the return of the S&P 500 is an irregular process.

This decomposition is called the classical decomposition model. There are two forms of decomposition: additive and multiplicative. They combine trend component 𝑇𝑡, seasonal component 𝑆𝑡 and random component 𝜀𝑡 to build a time series model.

The additive model adds up these components:

** 𝑋𝑡 = 𝑇𝑡 + 𝑆𝑡 + 𝜀t**

The multiplicative model multiplies these components:

** 𝑋𝑡 = 𝑇𝑡 × 𝑆𝑡 × 𝜀t**

In general, we can use a function of time or a smoothing method (like moving average) to model the trend component. We can use a function of dummy variable(s) to model the seasonal component. After removing the trend and seasonal components, we are left with the random residuals. Below, we summarize the general steps of time series modeling.

1. Plot the time series to check for non-stationary behavior, such as trend and seasonal components.

2. Remove the non-stationary components by such as subtracting trend and seasonality, taking first difference and etc., to obtain a stationary time series of random residuals.

3. Choose an appropriate time series model to fit the residuals.

4. Check the goodness-of-fit of the model.

5. Invert the operations in step 2, such as adding back the trend and seasonal components, integrating the residuals and etc., to build a model for the original time series.

There are a number of important time series models to model the random residuals. They are the subjects of this section.

AR, MA, ARMA, and ARIMA models are used to forecast the observation at (t+1) based on the historical data of previous time spots recorded for the same observation. However, it is necessary to make sure that the time series is stationary over the historical data of observation overtime period. If the time series is not stationary then we could apply the differencing factor on the records and see if the graph of the time series is a stationary overtime period.

**ACF (Auto Correlation Function)**

Auto Correlation function takes into consideration of all the past observations irrespective of its effect on the future or present time period. It calculates the correlation between the t and (t-k) time period. It includes all the lags or intervals between t and (t-k) time periods. Correlation is always calculated using the Pearson Correlation formula.

**PACF(Partial Correlation Function)**

The PACF determines the partial correlation between time period t and t-k. It doesn’t take into consideration all the time lags between t and t-k. For e.g. let’s assume that today’s stock price may be dependent on 3 days prior stock price but it might not take into consideration yesterday’s stock price closure. Hence we consider only the time lags having a direct impact on future time period by neglecting the insignificant time lags in between the two-time slots t and t-k.

**How to differentiate when to use ACF and PACF?**

Let’s take an example of sweets sale and income generated in a village over a year. Under the assumption that every 2 months there is a festival in the village, we take out the historical data of sweets sale and income generated for 12 months. If we plot the time as month then we can observe that when it comes to calculating the sweets sale we are interested in only alternate months as the sale of sweets increases every two months. But if we are to consider the income generated next month then we have to take into consideration all the 12 months of last year.

So in the above situation, we will use ACF to find out the income generated in the future but we will be using PACF to find out the sweets sold in the next month.

## AR** (Auto-Regressive) Model**

The time period at t is impacted by the observation at various slots t-1, t-2, t-3, ….., t-k. The impact of previous time spots is decided by the coefficient factor at that particular period of time. The price of a share of any particular company X may depend on all the previous share prices in the time series. This kind of model calculates the regression of past time series and calculates the present or future values in the series in know as Auto Regression (AR) model.

**Yt = β₁* y-₁ + β₂* yₜ-₂ + β₃ * yₜ-₃ + ………… + βₖ * yₜ-ₖ**

Consider an example of a milk distribution company that produces milk every month in the country. We want to calculate the amount of milk to be produced current month considering the milk generated in the last year. We begin by calculating the PACF values of all the 12 lags with respect to the current month. If the value of the PACF of any particular month is more than a significant value only those values will be considered for the model analysis.

For e.g in the above figure the values 1,2, 3 up to 12 displays the direct effect(PACF) of the milk production in the current month w.r.t the given the lag t. If we consider two significant values above the threshold then the model will be termed as AR(2).

**MA (Moving Average) Model**

The time period at t is impacted by the unexpected external factors at various slots t-1, t-2, t-3, ….., t-k. These unexpected impacts are known as Errors or Residuals. The impact of previous time spots is decided by the coefficient factor α at that particular period of time. The price of a share of any particular company X may depend on some company merger that happened overnight or maybe the company resulted in shutdown due to bankruptcy. This kind of model calculates the residuals or errors of past time series and calculates the present or future values in the series in know as Moving Average (MA) model.

Yt = α₁* Ɛₜ-₁ + α₂ * Ɛₜ-₂ + α₃ * Ɛₜ-₃ + ………… + αₖ * Ɛₜ-ₖ

Consider an example of Cake distribution during my birthday. Let’s assume that your mom asks you to bring pastries to the party. Every year you miss judging the no of invites to the party and end upbringing more or less no of cakes as per requirement. The difference in the actual and expected results in the error. So you want to avoid the error for this year hence we apply the moving average model on the time series and calculate the no of pastries needed this year based on past collective errors. Next, calculate the ACF values of all the lags in the time series. If the value of the ACF of any particular month is more than a significant value only those values will be considered for the model analysis.

For e.g in the above figure the values 1,2, 3 up to 12 displays the total error(ACF) of count in pastries current month w.r.t the given the lag t by considering all the in-between lags between time t and current month. If we consider two significant values above the threshold then the model will be termed as MA(2).

**ARMA (Auto Regressive Moving Average) Model**

This is a model that is combined from the AR and MA models. In this model, the impact of previous lags along with the residuals is considered for forecasting the future values of the time series. Here β represents the coefficients of the AR model and α represents the coefficients of the MA model.

Yt = β₁* yₜ-₁ + α₁* Ɛₜ-₁ + β₂* yₜ-₂ + α₂ * Ɛₜ-₂ + β₃ * yₜ-₃ + α₃ * Ɛₜ-₃ +………… + βₖ * yₜ-ₖ + αₖ * Ɛₜ-ₖ

Consider the above graphs where the MA and AR values are plotted with their respective significant values. Let’s assume that we consider only 1 significant value from the AR model and likewise 1 significant value from the MA model. So the ARMA model will be obtained from the combined values of the other two models will be of the order of ARMA(1,1).

**ARIMA (Auto-Regressive Integrated Moving Average) Model**

We know that in order to apply the various models we must in the beginning convert the series into Stationary Time Series. In order to achieve the same, we apply the differencing or Integrated method where we subtract the t-1 value from t values of time series. After applying the first differencing if we are still unable to get the Stationary time series then we again apply the second-order differencing.

The ARIMA model is quite similar to the ARMA model other than the fact that it includes one more factor known as Integrated( I ) i.e. differencing which stands for I in the ARIMA model. So in short ARIMA model is a combination of a number of differences already applied on the model in order to make it stationary, the number of previous lags along with residuals errors in order to forecast future values.

Consider the above graphs where the MA and AR values are plotted with their respective significant values. Let’s assume that we consider only 1 significant value from the AR model and likewise 1 significant value from the MA model. Also, the graph was initially non-stationary and we had to perform differencing operation once in order to convert into a stationary set. Hence the ARIMA model which will be obtained from the combined values of the other two models along with the Integral operator can be displayed as ARIMA(1,1,1).

**Conclusion :**

All these models give us an insight or at least close enough prediction about any particular time series. Also, it depends on the users that which model perfectly suffices their needs. If the chances of error rate are less in any one model compared to other models then it’s preferred that we choose the one which gives us the closest estimation.

An ARMAX model simply adds in the covariate on the right hand side:<span id="MathJax-Element-5-Frame" class="mjx-chtml MathJax_CHTML" style="margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;background: 0px 0px;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="{y}_{t}=\beta {x}_{t}+{\varphi}_{1}{y}_{t-1}+\cdots +{\varphi}_{p}{y}_{t-p}-{\theta}_{1}{z}_{t-1}-\cdots -{\theta}_{q}{z}_{t-q}+{z}_{t}“>yt=βxt+ϕ1yt−1+⋯+ϕpyt−p−θ1zt−1−⋯−θqzt−q+zt

$${y}_{t}=\beta {x}_{t}+{\varphi}_{1}{y}_{t-1}+\cdots +{\varphi}_{p}{y}_{t-p}-{\theta}_{1}{z}_{t-1}-\cdots -{\theta}_{q}{z}_{t-q}+{z}_{t}$$

where <span id="MathJax-Element-6-Frame" class="mjx-chtml MathJax_CHTML" style="margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;background: 0px 0px;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="${x}_{t}$“>xt is a covariate at time <span id="MathJax-Element-7-Frame" class="mjx-chtml MathJax_CHTML" style="background: 0px 0px;margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="$t$“>t and <span id="MathJax-Element-8-Frame" class="mjx-chtml MathJax_CHTML" style="background: 0px 0px;margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="$\beta $“>β is its coefficient. While this looks straight-forward, one disadvantage is that the covariate coefficient is hard to interpret. The value of <span id="MathJax-Element-9-Frame" class="mjx-chtml MathJax_CHTML" style="background: 0px 0px;margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="$\beta $“>β is *not* the effect on <span id="MathJax-Element-10-Frame" class="mjx-chtml MathJax_CHTML" style="background: 0px 0px;margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="${y}_{t}$“>yt

when the <span id="MathJax-Element-11-Frame" class="mjx-chtml MathJax_CHTML" style="margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;background: 0px 0px;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="${x}_{t}$“>xt is increased by one (as it is in regression). The presence of lagged values of the response variable on the right hand side of the equation mean that <span id="MathJax-Element-12-Frame" class="mjx-chtml MathJax_CHTML" style="background: 0px 0px;margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="$\beta $“>β can only be interpreted conditional on the value of previous values of the response variable, which is hardly intuitive.

If we write the model using backshift operators, the ARMAX model is given by<span id="MathJax-Element-13-Frame" class="mjx-chtml MathJax_CHTML" style="margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;background: 0px 0px;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="\varphi \left(B\right){y}_{t}=\beta {x}_{t}+\theta \left(B\right){z}_{t}\phantom{\rule{"2em"}{0ex}}\text{or}\phantom{\rule{"2em"}{0ex}}{y}_{t}=\frac{\beta}{\varphi \left(B\right)}{x}_{t}+\frac{\theta \left(B\right)}{\varphi \left(B\right)}{z}_{t},“>ϕ(B)yt=βxt+θ(B)ztoryt=βϕ(B)xt+θ(B)ϕ(B)zt,

$$\varphi (B){y}_{t}=\beta {x}_{t}+\theta (B){z}_{t}\phantom{\rule{2em}{0ex}}\text{or}\phantom{\rule{2em}{0ex}}{y}_{t}=\frac{\beta}{\varphi (B)}{x}_{t}+\frac{\theta (B)}{\varphi (B)}{z}_{t},$$

where <span id="MathJax-Element-14-Frame" class="mjx-chtml MathJax_CHTML" style="margin: 0px;padding: 1px 0px;border: 0px;font-size: 17px;vertical-align: baseline;background: 0px 0px;line-height: 0;float: none;direction: ltr;max-width: none;max-height: none;min-width: 0px;min-height: 0px" role="presentation" data-mathml="$\varphi \left(B\right)=1-{\varphi}_{1}B-\cdots -{\varphi}_{p}{B}^{p}$“>ϕ(B)=1−ϕ1B−⋯−ϕpBp

$\theta (B)=1-{\theta}_{1}B-\cdots -{\theta}_{q}{B}^{\mathrm{q.}}$