Professional Documents
Culture Documents
Causal Analysis
Chris Caplice
ESD.260/15.770/1.260 Logistics Systems
Sept 2006
Agenda
Forecasting Evaluation
Use of Causal Models in Forecasting
Approach and Methods
Forecast Evaluation
How do we determine what is a good forecast?
MD =
2. Mean Deviation:
t =1
MAD =
t =1
MSE =
n
t =1
2
t
2
t
t =1
RMSE =
et
D
MPE = t =1 t
n
et
D
t =1
n
4
MD
MAD
MSE
RMSE
MAPE
112.00
111.00
110.00
MA3
0.05
0.56
0.47
0.68
0.50%
MA10
0.21
1.07
1.67
1.29
0.96%
MA20
0.35
1.41
2.71
1.65
1.27%
109.00
108.00
-
20.00
MA3
40.00
MA10
60.00
MA20
80.00
100.00
120.00
ActDemand
Frequency
10
errors?
What is the variance of the errors?
.1
9)
0.
12
0.
42
0.
72
1.
03
1.
33
M
or
e
(0
.4
9)
.1
0)
.7
9)
(0
(1
.4
0)
0
(0
15
(1
20
error
Errors
2.00
e = 0.05
e= 0.69
D= 1.478
1.50
1.00
0.50
(0.50)
20
40
60
80
100
120
(1.00)
(1.50)
(2.00)
Tt=zt/MADt
Where zt= et + (1-)zt-1 (smoothing constant)
Corrective Actions
Adaptive Forecasting
Human Intervention
E (Y | x) = 0 + 1 x
yi = 0 + 1 xi
Yi = 0 + 1 xi + i
Observed
for i = 1, 2,...n
StdDev(Y | x) =
Unknown
yi = b0 + b1 xi for i = 1, 2,...n
ei = yi yi = yi b0 + b1 xi for i = 1, 2,...n
10
( e ) = ( y y ) = ( y b
n
i =1
i =1
i =1
b1 xi )
b0 = y b1 x
b1
n
i =1
( xi x )( yi y )
2
(
x
x
)
i =1 i
11
for i = 1, 2,...n
E (Y | x1 , x2 ,..., xk ) = 0 + 1 x1 + 2 x2 + ... + k xk
StdDev(Y | x1 , x2 ,..., xk ) =
( e ) = ( y y ) = ( y b
n
i =1
i =1
i =1
12
OLS Example
4,500
4,000
3,500
3,000
2,500
Ju
l
ay
M
ar
M
Ja
n
No
v
Se
p
Ju
l
ay
M
ar
2,000
M
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Demand
3,025
3,047
3,079
3,136
3,454
3,661
3,554
3,692
3,407
3,410
3,499
3,598
3,596
3,721
3,745
3,650
4,157
4,221
4,238
4,008
Ja
n
Month
13
OLS Example
Month
Establish relationship
Points to consider:
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Demand Period
3,025
1
3,047
2
3,079
3
3,136
4
3,454
5
3,661
6
3,554
7
3,692
8
3,407
9
3,410
10
3,499
11
3,598
12
3,596
13
3,721
14
3,745
15
3,650
16
4,157
17
4,221
18
4,238
19
4,008
20
Summer
0
0
0
0
1
1
1
1
0
0
0
0
0
0
0
0
1
1
1
1
14
4,100
3,900
3,700
Actual
Predicted
3,500
SUMMARY OUTPUT
3,300
0.979
0.958
0.953
79.21
20
3,100
Aug
Jul
Jun
May
Apr
Mar
Feb
Jan
Dec
Nov
Oct
Sep
Jul
Aug
Jun
May
Apr
Mar
Feb
2,900
Jan
Regression Statistics
Multiple R
R Square
Adjusted R Square
Standard Error
Observations
ANOVA
df
Regression
Residual
Total
Intercept
Period
Summer
2
17
19
Coefficients
2,969.14
48.03
303.51
SS
MS
F
Significance F
2442766.966 1221383.483 194.6730408
1.91955E-12
106658.4214 6274.024786
2549425.387
Standard Error
37.21
3.20
37.70
t Stat
79.79
15.00
8.05
P-value
0.0000
0.0000
0.0000
Lower 95%
Upper 95% Lower 95.0% Upper 95.0%
2,890.62
3,047.65
2,890.62
3,047.65
41.27
54.79
41.27
54.79
223.97
383.04
223.97
383.04
15
ANOVA
df
Regression
Residual
Total
Intercept
Period
Summer
Regression
Coefficients
2
17
19
SS
MS
F
Significance F
2442766.966 1221383.483 194.6730408
1.91955E-12
106658.4214 6274.024786
2549425.387
Degrees of
Freedom = n-k-1
t Stat
79.79
15.00
8.05
P-value
0.0000
0.0000
0.0000
95%
Confidence
Intervals
Lower 95.0%
Upper 95.0%
2,890.62
41.27
223.97
3,047.65
54.79
383.04
t-Statistic (bm/sbm)
Is bm different from 0?
P-value tells you % conf.
Chris Caplice, MIT
R2=1-ESS/TSS = RSS/TSS
TSS = ESS + RSS
Variation of observed around mean = Variation of observed
17
1
Yi =
1 + e X i
X
MIT Center for Transportation & Logistics ESD.260
18
X
Chris Caplice, MIT
Expert Opinions
44.8%
37.3%
14.9%
Sales Force
Executives
Industry Surveys
Level
Statistical Models
30.6%
20.9%
11.2%
6.0%
3.7%
Nave Model
Moving Average
Exp. Smoothing
Regression
Box-Jenkins
Industry
11
15
Corporate
11
18
Product Group
10
15
20
Product Line
11
16
20
Product
16
21
26
19
Bullwhip effect
Collaborative Planning, Forecasting, and Replenishment
(CPFR)
20
Questions, Comments,
Suggestions?