Professional Documents
Culture Documents
Time-Series Methods
Causal Methods
Regression Analysis
Multiple Regression
Time-series
models attempt to predict the future based on the past Common time-series models are
Regression
Causal
models use variables or factors that might influence the quantity being forecasted The objective is to build a model with the best statistical relationship between the variable being forecast and the independent variables Regression analysis is the most common technique used in causal modeling
YEAR
1 2 3 4 5 6 7 8 9 10
RADIOS
300 310 320 330 340 350 360 370 380 390
(a)
Annual Sales of Televisions 330
Sales appear to be
250
200 150 100 50
| | | | | | | | | |
constant over time Sales = 250 A good estimate of sales in year 11 is 250 televisions
0 1 2 3 4 5 6 7 8 9 10
Time (Years)
(b)
Sales appear to be
380
360 340 320
300 280
|
0 1 2 3 4 5 6 7 8 9 10
increasing at a constant rate of 10 radios per year Sales = 290 + 10(Year) A reasonable estimate of sales in year 11 is 400 televisions
Time (Years)
| | | | | | |
not be perfectly accurate because of variation from year to year Sales appear to be increasing A forecast would probably be a larger figure each year
0 1 2 3 4 5 6 7 8 9 10
Time (Years)
We compare forecasted values with actual values to see how well one model works or to compare models
YEAR 1 2 3 4 5 6 7 8 9
10
11
190
200
190
|190 200| = 10
Sum of |errors| = 160 MAD = 160/9 = 17.8
10
11
190
200
190
|190 200| = 10
Sum of |errors| = 160 MAD = 160/9 = 17.8
MAD
YEAR
1` 2 3 4 5 6 7 8 9 10 11
SALES
1400 1600 1800 2000 1800 6000 5000 6000 6000 7000
FORECAST
1800 2000 3000 3000 8000 6000 7000 7000 8000 9000
DEVIATION
MAD
There are other popular measures of forecast accuracy The mean squared error
(error)2 MSE n
The mean absolute percent error
time series is a sequence of evenly spaced events Time-series forecasts predict the future based solely of the past values of the variable Other variables are ignored
A time series typically has four components 1. Trend (T) is the gradual upward or downward movement of the data over time 2. Seasonality (S) is a pattern of demand fluctuations above or below trend line that repeats at regular intervals 3. Cycles (C) are patterns in annual data that occur every several years 4. Random variations (R) are blips in the data caused by chance and unusual situations
Trend Component Seasonal Peaks Actual Demand Line Average Demand over 4 Years
Year 1
Year 2
Time
Year 3
Year 4
There are two general forms of time-series models The multiplicative model
Demand = T x S x C x R
The additive model
Demand = T + S + C + R
Models may be combinations of these two
forms Forecasters often assume errors are normally distributed with a mean of zero
relatively steady over time The next forecast is the average of the most recent n data values from the time series This methods tends to smooth out short-term irregularities in the data series
Mathematically
Ft 1
Yt Yt 1 ... Yt n1 n
where
Ft 1 = forecast for time period t + 1 Yt
IIPM wants to forecast demand They have collected data for the past year
ACTUAL ASMISSIONS 10 12 13
April
May June July August September October November December January
16
19 23 26 30 28 18 16 14
July
August September October November December January
260
300 280 180 160 140
emphasis on recent periods Often used when a trend or other pattern is emerging
Ft 1
Mathematically
where
average model to forecast demand They decide on the following weighting scheme
WEIGHTS APPLIED 3 2 1 6 Sum of the weights PERIOD Last month Two months ago Three months ago
3 x Sales last month + 2 x Sales two months ago + 1 X Sales three months ago
[(3 X 13) + (2 X 12) + (10)]/6 = 12.17 [(3 X 16) + (2 X 13) + (12)]/6 = 14.33
May
June July August September
19
23 26 30 28
October
November December January
18
16 14
May
June July August September October November December January
190
230 260 300 280 180 160 140
Exponential smoothing is easy to use and requires little record keeping of data It is a type of moving average
New forecast = Last periods forecast + (Last periods actual demand - Last periods forecast)
Where is a weight (or smoothing constant) with a value between 0 and 1 inclusive
Mathematically
Ft 1 Ft (Yt Ft )
where
Ft+1 = new forecast (for time period t + 1) Ft = previous forecast (for time period t) = smoothing constant (0 1) Yt = pervious periods actual demand
the old estimate plus some fraction of the error in the last period
In January, Februarys demand for a certain car model was predicted to be 142 Actual February demand was 153 autos Using a smoothing constant of = 0.20, what is the forecast for March?
to obtaining a good forecast The objective is always to generate an accurate forecast The general approach is to develop trial forecasts with different values of and select the that results in the lowest MAD
is key
QUARTER
1 2 3 4 5 6 7
FORECAST USING =0.50 175 177.5 172.75 165.88 170.44 180.22 192.61
8
9
182
?
186.30
184.15
QUARTER
1 2 3 4 5 6 7 8
ACTUAL TONNAGE UNLOADED 180 168 159 175 190 205 180 182
FORECAST WITH = 0.10 175 175.5 174.75 173.18 173.36 175.02 178.02 178.22 |deviations| n
ABSOLUTE DEVIATIONS FOR = 0.10 5.. 7.5.. 15.75 1.82 16.64 29.98 1.98 3.78 82.45 = 10.31
FORECAST WITH = 0.50 175 177.5 172.75 165.88 170.44 180.22 192.61 186.30
ABSOLUTE DEVIATIONS FOR = 0.50 5. 9.5.. 13.75 9.12 19.56 24.78 12.61 4.3.. 98.63
MAD =
12.33
Best choice
QUARTER 1 2 3 4 5
FORECAST USING =
FORECAST USING =
6 7
8 9
600 800
700 ?
Like all averaging techniques, exponential smoothing does not respond to trends A more complex model can be used that adjusts for trends The basic approach is to develop an exponential smoothing forecast then adjust it for the trend
The equation for the trend correction uses a new smoothing constant Tt is computed by
Tt 1 (1 )T1 ( Ft 1 Ft )
where Tt+1 = smoothed trend for period t + 1 Tt = smoothed trend for preceding period = trend smooth constant that we select Ft+1 = simple exponential smoothed forecast for period t + 1 Ft = forecast for pervious period
makes the forecast more responsive to changes in trend A low value of gives less weight to the recent trend and tends to smooth out the trend Values are generally selected using a trial-and-error approach based on the value of the MAD for different values of Simple exponential smoothing is often referred to as first-order smoothing Trend-adjusted smoothing is called second-order, double smoothing, or Holts method
Trend Projection
Trend projection fits a trend line to a series
of historical data points The line is projected into the future for medium- to long-range forecasts Several trend equations can be developed based on exponential or quadratic models The simplest is a linear model developed using regression analysis
Trend Projection
The mathematical form is
Y b0 b1 X
where
Y = predicted value
Trend Projection
Dist7
*
Dist1
Dist5 Dist3
Dist6
Dist2
Dist4
Time
the following demand for its electrical generators over the period of 2001 2007
YEAR ELECTRICAL GENERATORS SOLD
r2 says model predicts about 80% of the variability in demand Significance level for F-test indicates a definite relationship
Y 56.71 10.54 X
To project demand for 2008, we use the coding
system to define X = 8
Actual Demand Line Trend Line Y 56.71 10.54 X
100
90 80 70 60
50
| | | | | | | | |
2001 2002 2003 2004 2005 2006 2007 2008 2009 Year
Recurring
variations over time may indicate the need for seasonal adjustments in the trend line A seasonal index indicates how a particular season compares with an average season When no trend is present, the seasonal index can be found by dividing the average value for a particular season by the average of all the data
Eichler
Supplies sells telephone answering machines Data has been collected for the past two years sales of one particular model They want to create a forecast this includes seasonality
MONTHLY DEMAND 94 94 94
April
May June July August
110
115 120 100 110
90
131 110 110 90
100
123 115 105 100
94
94 94 94 94
1.064
1.309 1.223 1.117 1.064
September
October November December
85
75 85 80
95
85 75 80 1,128 = 94 12 months
90
80 80 80
94
94 94 94
0.957
0.851 0.851 0.851
July
Aug. Sept. Oct. Nov.
1,200 1.117 112 12 1,200 1.064 106 12 1,200 0.957 96 12 1,200 0.851 85 12 1,200 0.851 85 12 1,200 0.851 85 12
June
Dec.
SALES DEMAND MONTH January February March YEAR 1 800 800 800 YEAR 2 700 750 900
MONTHLY DEMAND
April
May June July August
1000
1500 1200 1000 1100
900
1300 1300 1200 800
September
October November December
800
700 950 900
950
800 750 850
When both trend and seasonal components are present, the forecasting task is more complex Seasonal indices should be computed using a centered moving average (CMA) approach There are four steps in computing CMAs 1. Compute the CMA for each observation (where possible) 2. Compute the seasonal ratio = Observation/CMA for that observation 3. Average seasonal ratios to get seasonal indices 4. If seasonal indices do not add to the number of seasons, multiply each index by (Number of seasons)/(Sum of indices)
QUARTER 1 2 3 4 Average
Definite trend
Seasonal pattern
compare the actual sales with an average quarter centered on that time period We will use 1.5 quarters before quarter 3 and 1.5 quarters after quarter 3 that is we take quarters 2, 3, and 4 and one half of quarters 1, year 1 and quarter 1, year 2
0.5(108) + 125 + 150 + 141 + 0.5(116) CMA(q3, y1) = 4
= 132.00
these are averaged to get the seasonal index Index Index Index Index for for for for quarter quarter quarter quarter 1 2 3 4 = = = = I1 I2 I3 I4 = = = = (0.851 (0.965 (1.136 (1.051 + + + + 0.848)/2 0.960)/2 1.127)/2 1.063)/2
= = = =
CMAs
200 150 Sales 100 CMA
50
0
| |
5 6 7 Time Period
10
11
12
Decomposition is the process of isolating linear trend and seasonal factors to develop more accurate forecasts There are five steps to decomposition 1. Compute seasonal indices using CMAs 2. Deseasonalize the data by dividing each number by its seasonal index 3. Find the equation of a trend line using the deseasonalized data 4. Forecast for future periods using the trend line 5. Multiply the trend line forecast by the appropriate seasonal index
Problem
YEAR 1 QUARTER 1 2 3 4 1 2 3 4 1 2 3 4 SALES 100 120 135 140 160 140 150 150 120 140 180 160 CMA SEASONAL RATIO
SALES ($1,000,000s) 108 125 150 141 116 134 159 152 123 142 168 165
SEASONAL INDEX 0.85 0.96 1.13 1.06 0.85 0.96 1.13 1.06 0.85 0.96 1.13 1.06
DESEASONALIZED SALES ($1,000,000s) 127.059 130.208 132.743 133.019 136.471 139.583 140.708 143.396 144.706 147.917 148.673 155.660
b1 = 2.34
b0 = 124.78
Y = 124.78 + 2.34X = 124.78 + 2.34(13) = 155.2 (forecast before adjustment for seasonality) Y x I1 = 155.2 x 0.85 = 131.92
March
April May June
1.0203
1.0087 0.9935 0.9906
September
October November December
0.9653
1.0048 0.9598 0.9805
equation where
Multiple regression can be used to forecast both trend and seasonal components in a time series
One independent variable is time Dummy independent variables are used to represent the seasons
Y 104.1 2.3(13) 15.7(0) 38.7(0) 30.1(0) 134 Y 104.1 2.3(14 ) 15.7(1) 38.7(0) 30.1(0) 152
These are different from the results obtained using
the multiplicative decomposition method Use MAD and MSE to determine the best model
performance of a forecast Tacking signals are computed using the following equation
RSFE Tracking signal MAD
where
Acceptable Range
Time
than forecast Negative tracking signals indicate demand is less than forecast Some variation is expected, but a good forecast will have about as much positive error as negative error Problems are indicated when the signal trips either the upper or lower predetermined limits This indicates there has been an unacceptable amount of variation Limits should be reasonable and may vary from item to item
Adaptive smoothing is the computer monitoring of tracking signals and self-adjustment if a limit is tripped In exponential smoothing, the values of and are adjusted when the computer detects an excessive amount of variation