Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data

Analysis

APS 425 Fall 2015

Time Series Analysis:

Trend & Seasonality

Instructor: G. William Schwert

585-275-2470
[email protected]

Topics

Deterministic trends & seasonality

Stochastic trends & seasonality

Differencing & seasonal differencing

(c) Prof. G. William Schwert, 2002-2015 1

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Time Trends

Linear
Zt = a + b t + et
do you really expect such a pattern to persist forever?
e.g., accumulated horse manure in the streets of
Philadelphia in the late 18th century

exponential (linear in logs)

log(Zt) = a + b t + et
Zt = exp{a + b t + et}

In Eviews, time = @trend(K) creates a time trend equal

to 0 in observation K

Linear Trend Example

25
Actual Fitted

20

15

10

0
0 10 20 30 40 50 60 70 80 90 100

(c) Prof. G. William Schwert, 2002-2015 2

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Exponential Trend Example

180
Actual Fitted
160

140

120

100

80

60

40

20

0
0 10 20 30 40 50 60 70 80 90 100

Polynomial Trends

Zt = a + b t + b t2 + e
1 2 t

parabola
if b2 < 0, then it eventually turns negative
if b2 > 0, then it eventually sky-rockets up

you can approximate most samples with polynomials,

but you wouldn't expect to do well forecasting out-of-
sample

(c) Prof. G. William Schwert, 2002-2015 3

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Quadratic Trend Example

18

16

14

12

10

6 Actual Fitted

0
0 10 20 30 40 50 60 70 80 90 100

Other Functions of Time

Cosine or Sine represent fixed cycles

S-curves (logistic) -- new product sales

log((L - Zt)/Zt) = a + b t + et

Zt = L / {1 + exp[ a + b t + et]}
L is the upper limit (asymptote) for sales

can iterate on this by trying different values to see which

fits best

(c) Prof. G. William Schwert, 2002-2015 4

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

S-Curve Trend Example

12000
Actual Fitted

10000

8000

6000

4000

2000

0
0 10 20 30 40 50 60 70 80 90 100

Other Types of
Stationarity-inducing Transformations

per capita
divide by population (if it is growing)

real
divide by price index (e.g., CPI)

market share
divide by industry sales

(c) Prof. G. William Schwert, 2002-2015 5

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Relation to Differencing

Linear Trend:
Zt = a + b t + et
(Zt - Zt-1) = [a + b t + et] - [a + b (t-1) + et-1]
= b + et - et-1

Random Walk with Drift:

Zt = b + et + Zt-1
(Zt - Zt-1) = b + et

The Level of a
Random Walk with Drift

Z =b+e +Z
t t t-1

=b+e + b+e +Z
t t-1 t-2

=b+e + b+e + b+e +Z

t t-1 t-2 t-3

...
=bt+e +e +e +... +e +Z
t t-1 t-2 1 0

which is the initial value of Z0,

plus b times t,
plus the sum of all the shocks from time 1 through t

so Z0 plays the role of a in the trend model, but the

errors follow a random walk

(c) Prof. G. William Schwert, 2002-2015 6

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Seasonality

Any type of predictable behavior that is a function of the

season
originally used for agricultural products
growing/harvests happen at the same time every
year

Also caused by conventions of holidays, etc.

e.g., more sales at Christmas

Can also reflect days of the week

e.g., Fed Funds rate

Seasonality

1) Fixed effects -- dummy variables

2) ARMA model at seasonal frequencies

(c) Prof. G. William Schwert, 2002-2015 7

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Seasonals (Fixed)

Monthly dummy variables

e.g., if Jan = 1; otherwise = 0
allows for different intercepts by month

Day-of-the-week dummy variables for daily data

also, for holidays

In Eviews
Jan = @seas(1) if the workfile is defined as being monthly

Seasonals (Fixed)
Monthly Example

define Jant = 1 if the observation occurs in January, and = 0, otherwise; and so

forth

Zt = 1 Jant + 2 Febt + . . . + 12 Dect + et

so 1 is the average value of Z in Januarys, etc.

This is equivalent to the model where any one of the monthly dummies is left
out (e.g., Jan)

Zt = + 2 Febt + . . . + 12 Dect + et

so 2 is the average difference between the mean in February and the mean in
January (which is )

(c) Prof. G. William Schwert, 2002-2015 8

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Seasonal Differencing

Similar to ordinary differences, except s periods apart (Zt - Zt-s):

note that seasonal differencing removes a linear trend

it also removes different fixed means

i.e., dummy variables

Relations Among Seasonal Variables

Suppose you have a regression model that relates sales to past
advertising for a product that has a seasonal sales cycle -- e.g.,
automobiles
it is easy to imagine that some types of advertising that are
specific to a particular model or new feature would have no
effect on sales beyond this model year
other types of advertising that highlighted general quality
characteristics of the brand might have effects that last beyond
the model year
when building a model you need to think about what effects
seasonality might have on the model

(c) Prof. G. William Schwert, 2002-2015 9

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Seasonal ARIMA(P,D,Q)s Models

Just like ARIMA(p,d,q) models, except the gap between

observations that affect one another is s periods, not 1 period

Example: ARIMA(1,0,0)12 model [monthly AR(1)]

Zt = + 1 Zt-12 + et

so the autocorrelations decay exponentially at lags 12, 24,

36, etc.
partial autocorrelation at lag 12 = 1
after lag 12, they equal 0

Seasonal ARIMA(0,1,1)s Models

This model occurs a lot in real data

note that seasonal differencing removes a linear trend

it also removes different fixed means

i.e., dummy variables

if the MA parameter is close to 1, this "exponential smoothing"

model implies a slowly wandering level for the series that is
different for each period of the year

(c) Prof. G. William Schwert, 2002-2015 10

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Seasonal Exponential
Smoothing Forecasts

ARIMA(0,1,1)12 model:

[Zt - Zt-12] = et - 1 et-12

^ ^
Zt-12(12) = Zt-12 - 1 et-12 = Zt-12 - 1 [Zt-12 - Zt-24(12)]
^
= (1 - 1) Zt-12 + 1 Zt-24(12)

so the forecast for next January is a weighted average of the most recent
observation and the forecast for that observation (for Januarys)

Seasonal Adjustment

The most common method is the X-11 technique developed at

the U.S. Bureau of the Census
a complex (two-sided) filter that involves ratios and moving
averages

I have seen cases (e.g., CPI) where it does not seem to work well

for forecasting purposes, it is difficult, since the final seasonally

adjusted figure will keep getting revised as new data comes in
it is a two-sided filter

(c) Prof. G. William Schwert, 2002-2015 11

Time Series - Trends & Seasonality APS 425 - Advanced Managerial Data
Analysis

Seasonality: Summary

1) If the most interesting variation in the data is at seasonal

lags, then you probably need more data to get a reliable
model
e.g., with monthly dummies, you are trying to estimate the
monthly averages with only T/12 observations per parameter

2) It is probably better to directly model seasonality, rather

than use seasonally adjusted data
mechanical adjustment procedures often don't work well

Links

Return to APS 425 Home Page:

http://schwert.ssb.rochester.edu/a425/a425main.htm

(c) Prof. G. William Schwert, 2002-2015 12