Applied Macro and

Financial Econometrics
Module 1: Introduction

Course Instructor:

Dr. Devasmita Jena

What is Time Series Data?
A series of data points indexed in time order
Data collected over a period of time – either at regular or irregular intervals of
time
For TS modelling, we need data collected at regular intervals of time, i.e.,
observed in same frequency
TS data can occur in different frequencies:
• Annual (yearly)
• Quarterly
• Monthly
• Weekly
• Daily
• Hourly
Why study Time Series Analysis?
An appropriate theory to explain the relationship between variables may not be
available
• consider only statistical relationship of the given series with its past values
Sometimes even when the set of explanatory variables are known, it may not
necessarily be the entire set (or even a subset!) of variables required to estimate
a regression model
• Resort to using single series of dependent variables to forecast the future values
Application of UTSM are mainly to forecast the single series; that too in the
short-run. Examples:
• Forecasting macroeconomic variables (inflation, unemployment rates, FIIs, etc.) in the
near future for policy purposes
• Firms – forecasting demand/ market share of their products, costs etc.
• Forecasting financial variables (gold/ silver prices, stocks etc.)
• Spread of disease (covid)forecast
Time Series Analysis
Univariate TS Analysis:
• Univariate TS models are a class of specifications where one attempts to
model and predict TS variables using their own past values, current values
and/or past values of error terms.
Multivariate TS Analysis
• Multivariate TS models (also called structural models) are used to explain
changes in a variable by reference to the movements in the current or past
values of other variables (explanatory variables)
Why study Time Series Analysis?
If we are using two or more time series variables, it may be possible to use the
variation in one series to explain the variation in another series
• helps us understand the mechanism that generated the given time series
You have Xt Yt and Zt (say): doing a simply multiple regression model won’t
work
• Why?
Multivariate Forecasting/Prediction
• Take control of economic situation!
• Policy
Components of Time Series Data
1. Trend: A long term relatively smooth pattern that usually persists more than a
year
• Example: Population of India (Linear); Fixed Broad band subscriptions (Non Linear)
• Can you give more examples?
Components of Time Series Data
2. Seasonal: A pattern that usually appears in a regular interval wherein the
frequency of occurrence is within a year or even shorter
• Example: Quarterly GDP of India
• Can you give more examples?
Components of Time Series Data
What happens to quarterly GDP data if you adjust for seasonality?
Components of Time Series Data
3. Cyclical: The repeated pattern that usually appears in a TS, but beyond a
frequency of one year
• Example: Real Residential prices
• Can you give more examples?
Components of Time Series Data
3. Cyclical: wave-like (sinusoidal) pattern about a long term trend that is apparent
over a number of year.
Cycles are rarely regular and appears in combination with other components.
Components of Time Series Data
4. Random: The component of TS that is obtained after all the three components
have been “extracted” out of the series
How do you think a plot of random component of TS over time look like?
• Devoid of any pattern
• Indicating only a random pattern around mean
Question?
You want to get an idea of GDP(say) over the next quarter, can you do it by
simply extending the curve?
You can, provided:
• You know the statistical or stochastic mechanism (which is called Data Generating
Process) that generated these curves
But what’s that mechanism?
• Well, that’s the very first question of TS analysis!
Random Stochastic Process
Random/Stochastic Process: collection of random variables over time
• Example: GDP
• How is GDP a stochastic process?
• What’s the GDP growth for 4th quarter, FY23?
• https://www.fortuneindia.com/enterprise/indias-gdp-to-grow-at-71-in-fy23-sbi-ecowrap/
112819
• https://www.thehindu.com/business/Economy/gdp-grows-61-in-march-quarter-72-in-fy2
3/article66915856.ece
• https://www.livemint.com/news/india/india-q4-gdp-live-updates-nso-to-release-gdp-data-
for-january-march-2023-quarter-11685504900602.html
• https://currentaffairs.adda247.com/rbi-projects-robust-economic-growth-in-india-expects
-q1-gdp-growth-at-7-6/
We use realization to draw inference about DGP
• Corollary with Cross Section Data
Stationarity
Stationary Stochastic Process: process is stationary if its mean and variance are
constant over time and the value of covariance between 2 time periods depends
only on the lag between two time periods and not on actual time at which the
covariance is computed, i.e., Cov(Yt, Yt-k)=f(k).
Yt is said to be a stochastic TS with:
• Mean: E(Yt) = µ
• Variance: V(Yt) = E[ Yt - µ]2 = σ2
• Covariance/ Auto covariance at lag k: Ɣk = E[(Yt - µ)(Yt-k - µ)]
• Ɣ0= ? and Ɣ1 is the cov b/w two adjacent values of Y
• Suppose you shift origin of Y from Yt to Yt+m
o For Yt to be stationary, the mean, variance and the autocovariances ( at various lags) remain the same
no matter at what point we measure them
Stationarity
The mean, variance, autocovariances of stationary TS process are time invariant
Mean reverting: TS tend to return to mean and fluctuations around this mean
(variance) will broadly have constant amplitude.
In TS nomenclature, such stochastic process is known as:
• weakly stationary
• covariance stationary
• 2nd order stationary
Strictly stationary series: if all moments of probability distribution and not just the 1st
two (i.e., mean and variance) are time invariant
Weak stationarity suffices for most TS analysis
Example:
• Normal Distribution?
• White noise/ pure random process/Error term?
Does strict stationarity always mean weak stationarity?
Non Stationary Time Series
Non stationary TS => TS that is not stationary!
TS with time-varying mean and/or time-varying variance
Example?
• TS data with trend
• TS data with seasonality
What about TS data with cyclic behavior?
• A TS data with cyclic behavior, but with no trend or seasonality, is stationary.
• Cycles are not of fixed length
• We can’t be sure where the crests and troughs will be before we observe the series
• Randomness
Non Stationary Time Series
•
Why is Stationarity important?
Non stationary TS => behavior of TS can be studied only for the time period
under consideration
Generalization of “distributional trait” not possible for the entire time period :
each set of TS data will be particular episode
Forecasting not possible
Descriptive Techniques in TS: How to check for stationarity?
What descriptive techniques do you apply in cross-section data?
Will this work for TS data?
Time plot
• Shows important features of the series: trend, seasonality, outliers or discontinuities
• Describe data
• Check for stationarity
• Decision on transformation of data
• Formulate better models
What are the other ways of checking for stationarity?
Time plot
ACF/PACF plots
Unit root tests (DF, ADF, PP, KPSS, DF-GLS)
Trend Stationary and Difference Stationary
•
Trend Stationary and Difference Stationary (Contd…) & How to
make a TS stationary?
•
Few points to note
If there is a trend in the series and the size off seasonal effect appears to increase
with mean, i.e. seasonal effect is additive
• Transform the data to make seasonal effects constant y-o-y/m-o-m
If size of seasonal effect is directly proportional to mean, i.e., seasonal effect is
multiplicative
• Log transformation can make it additive
• However only variance gets stabilized, if at all
• May be impossible to achieve all the stationarity conditions
Model constructed with transformed data may be less than helpful: more
difficult to interpret
Forecasts obtained by transformed data may have to be transformed back to be
of use: may add bias
Use your judgement – comes with more data handling and experience!
Other practices to make data stationary?
• Differencing:
o Works well for linear trend
o In case of quadratic trend: take difference of already differenced data
o Used to remove seasonality as well
Taking logarithmic transformation of the data
• Helps when only the variance is changing over time
• One should be careful to use this method: transformed variable should have a direct
physical interpretation
• Example: percentage changes, growth etc
Exponential smoothing
• Method depends on the weighted average of the past observations, with weights decaying
exponentially as observations get older
• Prediction is weighted sum of past observations, but the model explicitly uses
exponentially decaying weights
Integrated Stochastic Process
•