Time Series Analysis and

Forecasting
 Introduction to Time Series Analysis
• A time-series is a set of observations on a quantitative variable
collected over time.
• Examples
– Dow Jones Industrial Averages
– Historical data on sales, inventory, customer counts, interest
rates, costs, etc
• Businesses are often very interested in forecasting time series
variables.
• Often, independent variables are not available to build a
regression model of a time series variable.
• In time series analysis, we analyze the past behavior of a
variable in order to predict its future behavior.
 Methods used in Forecasting
• Regression Analysis
• Time Series Analysis (TSA)
– A statistical technique that uses time-
series data for explaining the past or
forecasting future events.
– The prediction is a function of time
(days, months, years, etc.)
– No causal variable; examine past behavior
of a variable and and attempt to predict
future behavior
 Components of TSA
• Time Frame (How far can we predict?)
– short-term (1 - 2 periods)
– medium-term (5 - 10 periods)
– long-term (12+ periods)
– No line of demarcation
• Trend
– Gradual, long-term movement (up or down) of
demand.
– Easiest to detect
 Components of TSA (Cont.)
• Cycle
– An up-and-down repetitive movement in demand.
– repeats itself over a long period of time
• Seasonal Variation
– An up-and-down repetitive movement within a trend
occurring periodically.
– Often weather related but could be daily or weekly
occurrence
• Random Variations
– Erratic movements that are not predictable because they
do not follow a pattern
 Time Series Plot Actual Sales

$3,000

$2,500

$2,000
Sales (in $1,000s)

$1,500

$1,000

$500

$0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

Time Period
 Components of TSA (Cont.)
• Difficult to forecast demand because...
– There are no causal variables
– The components (trend, seasonality,
cycles, and random variation) cannot
always be easily or accurately
identified
 Some Time Series Terms
• Stationary Data - a time series variable exhibiting
no significant upward or downward trend over
time.
• Nonstationary Data - a time series variable
exhibiting a significant upward or downward
trend over time.
• Seasonal Data - a time series variable exhibiting
a repeating patterns at regular intervals over
time.
Approaching Time Series Analysis
• There are many, many different time series
techniques.
• It is usually impossible to know which technique
will be best for a particular data set.
• It is customary to try out several different
techniques and select the one that seems to
work best.
• To be an effective time series modeler, you need
to keep several time series techniques in your
“tool box.”
 Measuring Accuracy
• We need a way to compare different time series techniques for a given data set.
• Four common techniques are the:

– mean absolute deviation,

– mean absolute percent error,

– the mean square error, n 

Yi  Y
– root mean square error.
MAD = 
i 1 n
i

100 n Yi  Ŷi
MAPE = 
n i 1 Yi

MSE = 
n 
Yi  Y i  
2

i 1 n
RMSE  MSE
• We will focus on MSE.
 Extrapolation Models
• Extrapolation models try to account for the past behavior
of a time series variable in an effort to predict the future
behavior of the variable.

  f Y , Y , Y ,
Yt 1 t t 1 t 2

 We’ll first talk about several extrapolation techniques

that are appropriate for stationary data.
 An Example
• Electra-City is a retail store that sells audio and video
equipment for the home and car.
• Each month the manager of the store must order
merchandise from a distant warehouse.
• Currently, the manager is trying to estimate how many
VCRs the store is likely to sell in the next month.
• He has collected 24 months of data.
•
 Moving Averages

 Yt  Yt-1  Yt- k +1
Yt 1 
k

 No general method exists for determining k.

 We must try out several k values to see what works best.
Implementing the Model
 A Comment on Comparing MSE
Values
• Care should be taken when comparing MSE
values of two different forecasting techniques.
• The lowest MSE may result from a technique that
fits older values very well but fits recent values
poorly.
• It is sometimes wise to compute the MSE using
only the most recent values.
 Forecasting With The Moving Average
Model
Forecasts for time periods 25 and 26 at time period 24:

 Y24  Y23 36 + 35
Y25    355
.
2 2
 Y
Y 35.5 + 36

Y26  25 24
  35.75
2 2
 Weighted Moving Average
• The moving average technique assigns equal weight
to all previous observations
 1 1 1
Yt 1  Yt  Yt-1  Yt- k -1
k k k
 The weighted moving average technique allows for
different weights to be assigned to previous
observations.
  w Y  w Y  w Y
Yt 1 1 t 2 t-1 k t- k -1

where 0  wi  1 and w
i 1
 We must determine values for k and the wi
Implementing the Model
 Using optimal values for the
weights that minimizes the MSE
 Forecasting With The Weighted
Moving Average Model
Forecasts for time periods 25 and 26 at time period 24:
  w Y  w Y  0.291  36  0.709  35  35.3
Y25 1 24 2 23

Ŷ26  w1Ŷ25  w2 Y24  0.291 35.3  0.709  36  35.80

 Exponential Smoothing

 Y
Y    (Y  Y
 )
t 1 t t t
where 0    1

 It can be shown that the above equation is equivalent to:

  Y   (1   )Y   (1   ) 2 Y   (1   ) n Y 
Yt 1 t t 1 t 2 t n
 Examples of Two
Exponential Smoothing Functions
42

40

38
Units Sold

36

34

32
Number of VCRs Sold
30 Exp. Smoothing alpha=0.1
Exp. Smoothing alpha=0.9

28
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Time Period
Implementing the Model
 Forecasting With The
Exponential Smoothing Model
Forecasts for time periods 25 and 26 at time period 24:

 Y
Y    (Y  Y
 )  35.74  0.268( 36  35.74)  35.81
25 24 24 24

 Y
Y    (Y  Y
 )Y
   (Y
 Y )Y
  3581
.
26 25 25 25 25 25 25 25

Note that,
  35.81, for t = 25, 26, 27, 
Yt
 Seasonality
• Seasonality is a regular, repeating
pattern in time series data.
• May be additive or multiplicative in
nature...
Stationary Seasonal Effects
Additive Seasonal Effects

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Tim e Pe r iod

Multiplicative Seasonal Effects

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Tim e Pe r iod
 Stationary Data With
Additive Seasonal Effects
Ŷt  n  E t  St  n  p
where
E t   (Yt - St-p )  (1-  )Et 1
St   (Yt - E t )  (1-  )St  p
0  1
0   1
p represents the number of seasonal periods

• Et is the expected level at time period t.

• St is the seasonal factor for time period t.
Implementing the Model
 Forecasting With The Additive
Seasonal Effects Model
Forecasts for time periods 25 - 28 at time period 24:

Ŷ24 n  E 24  S24 n  4

Ŷ25  E 24  S21  354.55  8.45  363.00

Ŷ26  E 24  S22  354.55  17.82  336.73
Ŷ27  E 24  S23  354.55  46.58  401.13
Ŷ28  E 24  S24  354.55  31.73  322.82
 Stationary Data With
Multiplicative Seasonal Effects
Ŷt  n  E t  St  n  p
where
E t   (Yt /St-p )  (1-  )Et 1
St   (Yt /E t )  (1-  )St  p
0  1
0   1
p represents the number of seasonal periods

• Et is the expected level at time period t.

• St is the seasonal factor for time period t.
Implementing the Model
 Forecasting With The Multiplicative
Seasonal Effects Model
Forecasts for time periods 25 - 28 at time period 24:

Ŷ24 n  E 24  S24 n  4

Ŷ25  E 24  S21  353.95 1.015  359.26

Ŷ26  E 24  S22  353.95  0.946  334.83
Ŷ27  E 24  S23  353.95 1.133  401.03
Ŷ28  E 24  S24  353.95  0.912  322.80
 Trend Models
• Trend is the long-term sweep or general
direction of movement in a time series.
• We’ll now consider some nonstationary time
series techniques that are appropriate for
data exhibiting upward or downward trends.
 An Example
• WaterCraft Inc. is a manufacturer of personal water crafts (also known as jet
skis).
• The company has enjoyed a fairly steady growth in sales of its products.
• The officers of the company are preparing sales and manufacturing plans
for the coming year.
• Forecasts are needed of the level of sales that the company expects to
achieve each quarter.
 Double Exponential Smoothing
(Holt’s Method)
Ŷt  n  E t  nTt
where
Et = Yt + (1-)(Et-1+ Tt-1)
Tt = (Et Et-1) + (1-) Tt-1
0    1 and 0    1

• Et is the expected base level at time period t.

• Tt is the expected trend at time period t.
• n is the period-ahead forecast made in period t.
Implementing the Model
Using optimal values for α and ß that
minimizes the MSE
 Forecasting With Holt’s Model
Forecasts for time periods 21 through 24 at time period 20:
Ŷ20 n  E 20  nT20
  E  1T  2336.8  1  152.1  2488.9
Y21 20 20
  E  2T  2336.8  2  152.1  2641.0
Y22 20 20
  E  3T  2336.8  3  152.1  27931
Y .
23 20 20
  E  4T  2336.8  4  152.1  2945.2
Y24 20 20
Holt-Winter’s Method For Additive
Seasonal Effects
Ŷt  n  E t  nTt  St  n  p
where

 
E t   Yt  St  p  (1-  )(Et 1  Tt 1 )
Tt   E t  E t 1   (1 -  )Tt 1
St   Yt  E t   (1-  )St  p
0  1
0   1
0   1
Implementing the Model
The Linear Trend Model
  b b X
Yt 0 1 1t

where X1t  t

For example:
X11  1, X12  2, X13  3, 
Implementing the Model
 Forecasting With The Linear Trend Model
Forecasts for time periods 21 through 24 at time period 20:
  b  b X  3751
Y .  92.6255  21  2320.3
21 0 1 121
  b  b X  3751
Y .  92.6255  22  2412.9
22 0 1 122

  b  b X  3751
Y .  92.6255  23  2505.6
23 0 1 123
  b  b X  3751
Y .  92.6255  24  2598.2
24 0 1 124
 The TREND() Function
TREND(Y-range, X-range, X-value for prediction)
where:
Y-range is the spreadsheet range containing the dependent Y
variable,
X-range is the spreadsheet range containing the independent X
variable(s),
X-value for prediction is a cell (or cells) containing the values for
the independent X variable(s) for which we want an estimated value
of Y.
Note: The TREND( ) function is dynamically updated whenever any inputs to
the function change. However, it does not provide the statistical information
provided by the regression tool. It is best two use these two different
approaches to doing regression in conjunction with one another.
The Quadratic Trend Model
  b b X b X
Yt 0 1 1t 2 2t

where X1t  t and X 2t  t 2

Implementing the Model
 Forecasting With The Quadratic Trend
Model
Forecasts for time periods 21 through 24 at time period 20:
Ŷ21  b0  b1X12 1  b2 X 22 1  653.67  16.671 21  3.617  212  2598.8

Ŷ22  b0  b1X12 2  b2 X 2 2 2  653.67  16.671 22  3.617  22 2  2771.0

Ŷ23  b0  b1X12 3  b2 X 2 23  653.67  16.671 23  3.617  232  2950.4

Ŷ24  b0  b1X12 4  b2 X 2 2 4  653.67  16.671 24  3.617  24 2  3137.1
 Computing Multiplicative
Seasonal Indices
• We can compute multiplicative seasonal
adjustment indices for period p as follows:

Yi
i Y
i
Sp  , for all i occuring in season p
np
 The final forecast for period i is then
 adjusted = Y
Y   S , for any i occuring in season p
i i p
Implementing the Model
The Plot
 Forecasting With Seasonal Adjustments
Applied To Our Quadratic Trend Model
Forecasts for time periods 21 through 24 at time period 20:
Ŷ21  (b0  b1X121  b2 X 2 2 1 ) S1  2598.9  105.7%  2747.8
Ŷ22  (b0  b1X12 2  b2 X 2 2 2 ) S 2  2771.1 80.1%  2219.6
Ŷ23  (b0  b1X12 3  b2 X 2 2 3 ) S3  2950.5  103.1%  3041.4

Ŷ24  (b0  b1X12 4  b2 X 22 4 ) S 4  3137.2  111.1%  3486.1

 Summary of the Calculation and Use
of Seasonal Indices
1. Create a trend model and calculate the estimated value for Ŷt
each observation in the sample.
2. For each observation, calculate the ratio of the actual value
to the predicted trend value: Y / Ŷ .
t t
(For additive effects, compute the difference:
Yt  Ŷt ).
3. For each season, compute the average of the ratios calculated
in step 2. These are the seasonal indices.
4. Multiply any forecast produced by the trend model by the
appropriate seasonal index calculated in step 3. (For additive
seasonal effects, add the appropriate factor to the forecast.)
 Refining the Seasonal Indices
• Note that Solver can be used to
simultaneously determine the optimal
values of the seasonal indices and the
parameters of the trend model being
used.
• There is no guarantee that this will
produce a better forecast, but it should
produce a model that fits the data better
in terms of the MSE.
 Seasonal Regression Models
• Indicator variables may also be used in regression models
to represent seasonal effects.
• If there are p seasons, we need p-1 indicator variables.
 Our example problem involves quarterly data, so p=4
and we define the following 3 indicator variables:
1, if Yt is an observation from quarter 1
X 3t  
0, otherwise
1, if Yt is an observation from quarter 2
X4t  
0, otherwise
1, if Yt is an observation from quarter 3
X5t  
0, otherwise
 Implementing the Model
• The regression function is:
Y  b b X b X b X b X b X
t 0 1 1t 2 2t 3 3t 4 4t 5 5t

where X1t  t and X 2t  t 2

The Data
The Regression Results
The Plot
 Forecasting With The
Seasonal Regression Model
Forecasts for time periods 21 through 24 at time period 20:

Ŷ21  824.471  17.319(21)  3.485(21) 2  86.805(1)  424.736(0)  123.453(0)  2638.5

Ŷ22  824.471  17.319(22)  3.485(22) 2  86.805(0)  424.736(1)  123.453(0)  2467.7

Ŷ23  824.471  17.319(23)  3.485(23) 2  86.805(0)  424.736(0)  123.453(1)  2943.2

Ŷ24  824.471  17.319(24)  3.485(24) 2  86.805(0)  424.736(0)  123.453(0)  3247.8

 Combining Forecasts
• It is also possible to combine forecasts to create a composite forecast.
• Suppose we used three different forecasting methods on a given data
set.

 Denote the predicted value of time period t using

each method as follows:
F1t , F2 t , and F3t
 We could create a composite forecast as follows:
  b b F b F b F
Yt 0 1 1t 2 2t 3 3t