Upload
ata-turan
View
227
Download
0
Embed Size (px)
Citation preview
8/3/2019 Chapter Regression
1/19
01.07.2011
1
CHAPTER 4
CAUSAL FORECASTING WITH REGRESSION
TIME SERIES METHODS
SIMPLE LINEAR REGRESSION
We wish to forecast a dependent variable. The value
of dependent variable is related to an observable
value of one or more independent variables.
We call this process causal forecasting, because the
value of the dependent variable is often caused by,
or at least highly correlated with, the value of the
independent variable.
8/3/2019 Chapter Regression
2/19
01.07.2011
2
SIMPLE LINEAR REGRESSION
we minimize the sum of the squared differencesbetween the actual sales and the sales indicated by
the model. The difference is the error of the
forecast.
SIMPLE LINEAR REGRESSION
8/3/2019 Chapter Regression
3/19
01.07.2011
3
Example
Example
8/3/2019 Chapter Regression
4/19
01.07.2011
4
Example
Example
8/3/2019 Chapter Regression
5/19
01.07.2011
5
Example
If there are 23 housing starts in January of 1996,
we would expect to sell about
24.17 +1.83 23 66 fixtures in February.
Coefficient of determination
8/3/2019 Chapter Regression
6/19
01.07.2011
6
Comments on Regression Regression models are very useful for forecasting when
there is a strong relationship and a time lag between the
dependent variable and the iindependent variable.
If there is no time lag between dependent and
independent variables, i.e., they occur in the same time
period, we cannot forecast future values of the dependent
value unless we use a forecast of the independent
variable, which may introduce additional error in theforecast of the dependent variable.
If causal relationships do not exist, regression is not the
best forecasting method.
Time Series Methods
For short-term forecasting, time series methods are favored.
A time series is simply a time-ordered list of historical data, theunderlying assumption which is that history is a reasonable predictor ofthe future.
There are several time series models and methods to choose from,including a constant, trend, or seasonal model, depending on thehistorical data and our understanding of the underlying process. Constant process
Moving Average
Simple Exponential Smoothing
Trend process
Double Exponential Smoothing
Double Moving Average (Regression)
Seasonal process
Winters Method
8/3/2019 Chapter Regression
7/19
01.07.2011
7
Constant Process-Last Data Point (LPD)
Constant Process- Average all past data
Given T periods ofdata, the average at time T is
8/3/2019 Chapter Regression
8/19
01.07.2011
8
Example
Example
8/3/2019 Chapter Regression
9/19
01.07.2011
9
Constant Process- Moving Average
Rather than take an average of all data points, we might choose
to average only some of the more recent data. This method,
called a moving average, is a compromise between the last data
point and average methods. It averages recent data to reduce
the effect of random fluctuations.
Constant Process- Moving Average
8/3/2019 Chapter Regression
10/19
01.07.2011
10
Constant Process-Simple Exponential Smoothing
Constant Process-Simple Exponential Smoothing
8/3/2019 Chapter Regression
11/19
01.07.2011
11
Constant Process-Simple Exponential Smoothing
Trend Process-Double Exponential Smoothing
8/3/2019 Chapter Regression
12/19
01.07.2011
12
Trend Process-Double Exponential Smoothing
Trend Process-Double Exponential Smoothing
8/3/2019 Chapter Regression
13/19
01.07.2011
13
Solution:
First, compute the averages of the months 1 to 12, and
13 to 24.
Trend Process-Double Exponential Smoothing
Trend Process-Double Exponential Smoothing
8/3/2019 Chapter Regression
14/19
01.07.2011
14
Trend Process-Double Exponential Smoothing
Trend Process- Other Method
Regression, with time as the independent variable,
can be used. Let dt be the demand in period t.
T=1,2,..T.
222
1 1
)1((4
1)12)(1((
6
1
)1((
2
1
+++
+
=
= =
TTTTT
dTTtdT
b
T
t
T
t
tt
)1(2
1
1
+= =
Tb
dT
a
T
t
t
kbaFkt
+=+
8/3/2019 Chapter Regression
15/19
01.07.2011
15
Seasonal Process-Winters Method
Many processes naturally have some number of seasons in a year.
If the time periods are weeks, the year would have 52 seasons.
Periods of months and quarters have 12 and 4 seasons in a year,
respectively.
A good model must consider the constant portion of demand, the
trend and seasonality.
Seasonal Process-Winters Method
8/3/2019 Chapter Regression
16/19
01.07.2011
16
Seasonal Process-Winters Method
Seasonal Process-Winters Method
8/3/2019 Chapter Regression
17/19
01.07.2011
17
Seasonal Process-Winters Method
Seasonal Process-Winters Method
8/3/2019 Chapter Regression
18/19
01.07.2011
18
Seasonal Process-Winters Method
Seasonal Process-Winters Method
8/3/2019 Chapter Regression
19/19
01.07.2011
Seasonal Process-Winters Method