1

INTRODUCTION TO
FORECASTING & TIME
SERIES STRUCTURE
Dr. Susan Simmons
Institute for Advanced Analytics
 2

Source: xkcd.com/2620
 3

TIME SERIES DATA

 4

Time Series Data

• A time series is an ordered sequence of observations.
• Ordering is typically through equally spaced time
intervals.
• Possibly through space as well.
• Used in a variety of fields:
• Agriculture: Crop Production
• Economics: Stock Prices
• Engineering: Electric Signals
• Meteorology: Wind Speeds
• Social Sciences: Crime Rates
 5

Time Series Data

• We will begin our time

series discussions with
univariate time series
(only one time
series…one variable,
we will call it Y).
• Multivariate time series
will be in Fall 2.
 6

Time Series Data

• We will begin our time

Date Y
series discussions with January 2000 23
univariate time series February 2000 18
(only one time March 2000 20
series…one variable, April 2000 25

we will call it Y). May 2000 21

• Multivariate time series

will be in Fall 2.
 7

Time Series Data

• We will begin our time

Date Y
series discussions with January 2000 23 Y1
univariate time series February 2000 18
(only one time March 2000 20
series…one variable, April 2000 25

we will call it Y). May 2000 21

• Multivariate time series

will be in Fall 2.
 8

Time Series Data

• We will begin our time

Date Y
series discussions with January 2000 23
univariate time series February 2000 18 Y2
(only one time March 2000 20
series…one variable, April 2000 25

we will call it Y). May 2000 21

• Multivariate time series

will be in Fall 2.
 9

Time Series Data

• We will begin our time

Date Y
series discussions with January 2000 23
univariate time series February 2000 18
(only one time March 2000 20 Y3
series…one variable, April 2000 25

we will call it Y). May 2000 21

• Multivariate time series

will be in Fall 2.
 10

Time Series Data

• We will begin our time

Date Y
series discussions with January 2000 23
univariate time series February 2000 18
(only one time March 2000 20 Y3
series…one variable, April 2000 25

we will call it Y). May 2000 21

• Multivariate time series Yt

will be in Fall 2.
CAREFUL: Since we are assuming equally
spaced, you will need to take care of missing
values !!
 11

Example 1: Iron and Steel Exports

 12

Example 2: Amazon.com Stock

 13

Example 2: Amazon.com Stock

Time series can have a

trend – an overall
pattern to the data
(linear, quadratic;
positive, negative)
 14

Example 3: Airlines Passengers

 15

Example 3: Airlines Passengers

Time series can

have a seasonal
pattern – a
systematic up and
down pattern
 16

Example 5: Airline Passengers Again

 17

Temperature over the past century for

Tuscaloosa, Alabama

Source: Dr. Robert Lund

 18

Temperature over the past century for

Tuscaloosa, Alabama

Station relocated and

instrumentation was
changed

Thermometer was
Station was changed
relocated

Source: Dr. Robert Lund

 19

Time Series to Forecast

 20

Forecasting Process

Propose
Fit Model
Model

Data Cleaning Forecasts

Diagnose
Repeat
Model
 21

SIGNAL AND NOISE

 22

Statistical Forecasting

Time
Signal Noise
Series
 23

Statistical Forecasting

Time
Signal Noise
Series

Explained Variation Unexplained Variation

 24

Statistical Forecasting

Time
Signal Noise
Series

Explained variation: Error

Trend/Cycle
Seasonality
 25

Statistical Forecasting

Time
Signal Noise
Series

Explained variation: Error

Trend/Cycle
Seasonality This will be
Dependence Structure discussed in
ARIMA
 26

Statistical Forecasting

Time
Signal Noise
Series

Forecasts extrapolate Confidence intervals

signal portion of model. account for uncertainty.
 27

Time Series Decomposition

• If a time series only has trend/cycle patterns, there is no
need to decompose
• If a time series has both trend/cycle patterns AND
seasonal variation, we can decompose series into these
individual parts:
• Trend/Cycle patterns
• Seasonal variation
• Error
 28

Time Series Decomposition

• The signal part of the time series can typically be broken
down into two components:

Time
Signal Noise
Series

Trend / Cycle and Seasonal Error / Remainder / Irregular

 29

Time Series Decomposition

• The whole time series can now be thought of like the
equations below.
• Additive:

𝑌𝑡 = 𝑇𝑡 + 𝑆𝑡 + 𝐸𝑡

• Multiplicative:

𝑌𝑡 = 𝑇𝑡 × 𝑆𝑡 × 𝐸𝑡
 30

Time Series Decomposition

• The whole time series can now be thought of like the
equations below.
• Additive:

𝑌𝑡 = 𝑇𝑡 + 𝑆𝑡 + 𝐸𝑡

Trend / Cycle
• Multiplicative:

𝑌𝑡 = 𝑇𝑡 × 𝑆𝑡 × 𝐸𝑡
 31

Time Series Decomposition

• The whole time series can now be thought of like the
equations below.
• Additive:

𝑌𝑡 = 𝑇𝑡 + 𝑆𝑡 + 𝐸𝑡

Seasonal
• Multiplicative:

𝑌𝑡 = 𝑇𝑡 × 𝑆𝑡 × 𝐸𝑡
 32

Time Series Decomposition

• The whole time series can now be thought of like the
equations below.
• Additive:

𝑌𝑡 = 𝑇𝑡 + 𝑆𝑡 + 𝐸𝑡

Error
• Multiplicative:

𝑌𝑡 = 𝑇𝑡 × 𝑆𝑡 × 𝐸𝑡
 33

Time Series Decomposition

• The whole time series can now be thought of like the
equations below.
• Additive:

𝑌𝑡 = 𝑇𝑡 + 𝑆𝑡 + 𝐸𝑡

• Multiplicative:

𝑌𝑡 = 𝑇𝑡 × 𝑆𝑡 × 𝐸𝑡
OR
log(𝑌𝑡 ) = log(𝑇𝑡 ) + log(𝑆𝑡 ) + log(𝐸𝑡 )
 34

Additive vs. Multiplicative

• Additive – magnitude of • Multiplicative – magnitude
variation around trend / of the variation around
cycle remains constant. trend / cycle proportionally
changes.
 35

Seasonally Adjusted Data

One advantage of time series decomposition is that we are
able to create seasonally adjusted data (i.e. remove the “effect
of Seasonality”)

This allows analysts to understand the trend of the series

𝑌𝑡 = 𝑇𝑡 + 𝑆𝑡 + 𝐸𝑡
𝑌𝑡 − 𝑆𝑡 (𝑇𝑡 + 𝐸𝑡 )

The seasonal length of the time series is the length of one

season (how long til the series repeats the “pattern”)
36
 37

Airline data set

• Data contains number of US airline passengers from
January 1990 – March 2008
• Data is monthly (length of season is 12…repeats pattern
every 12 observations)
 38

Time Series Decomposition

𝑌𝑡
=

𝑆𝑡
+
𝑇𝑡
+
𝐸𝑡
 39

Time Series Decomposition

𝑌𝑡
=

𝑆𝑡 Original
+ data

𝑇𝑡
+
𝐸𝑡
 40

Time Series Decomposition

𝑌𝑡
=

𝑆𝑡
+
Seasonal
𝑇𝑡 Component
+
𝐸𝑡
 41

Time Series Decomposition

𝑌𝑡
=

𝑆𝑡
+
𝑇𝑡
+
𝐸𝑡 Trend
Component
 42

Time Series Decomposition

𝑌𝑡
=

𝑆𝑡
+
𝑇𝑡
Remainder
+ (What’s left
𝐸𝑡 over)
 43

Time Series Decomposition

# Time Series Decomposition ...STL#

Passenger <- ts(USAirlines$Passengers, start = 1990, frequency =12)
decomp_stl <- stl(Passenger, s.window = 7)

# Plot the individual components of the time series

plot(decomp_stl)

autoplot(decomp_stl)
44
 45

Time Series Decomposition

autoplot(Passenger)+
geom_line(aes(y=decomp_stl$time.series[,2]),
color="blue")

Overlay the trend component

46
 47

Time Series Decomposition

seas_adj=Passenger-decomp_stl$time.series[,1]

autoplot(Passenger) +
geom_line(aes(y=decomp_stl$time.series[,2]),color="blue")
+ geom_line(aes(y=seas_adj),color="orange")

Overlay the trend component

Overlay seasonally adjusted
 48

Time Series Decomposition

𝑇𝑡

𝑌𝑡 − 𝑆𝑡 = 𝑇𝑡 + 𝐸𝑡
 49

Time Series Decomposition – R

ggsubseriesplot(Passenger)
 50

Time Series Decomposition

 51

Decomposition Techniques
• There are many different ways to calculate the trend/cycle
and seasonal effects inside time series data.
• Here are 3 common techniques:
1. Classical Decomposition
 52

Decomposition Techniques
• There are many different ways to calculate the trend/cycle
and seasonal effects inside time series data.
• Here are 3 common techniques:
1. Classical Decomposition
a. Default in SAS (Can be done in R)
b. Trend – Uses Moving / Rolling Average Smoothing
c. Seasonal – Average De-trended Values Across Seasons
 53

Decomposition Techniques
• There are many different ways to calculate the trend/cycle
and seasonal effects inside time series data.
• Here are 3 common techniques:
1. Classical Decomposition
2. X-13 ARIMA Decomposition (self study)
a. Trend – Uses Moving / Rolling Average Smoothing
b. Seasonal – Uses Moving / Rolling Average Smoothing
c. Iteratively Repeats Above Methods and ARIMA Modeling
d. Can handle outliers
 54

Decomposition Techniques
• There are many different ways to calculate the
trend/cycle, and seasonal effects inside time series data.
• Here are 3 common techniques:
1. Classical Decomposition
2. X-12 ARIMA Decomposition
3. STL (Seasonal and Trend using LOESS estimation)
Decomposition
a. Default of stl Function in R (Not available in SAS)
b. Uses LOcal regrESSion Techniques to Estimate Trend
and Seasonality
c. Allows Changing Effects for Trend and Season
d. Adapted to Handle Outliers
 55

Comparison of Classical versus STL

seasonal decomposition