Title: Random Series
1Time Series and Forecasting
2STEREO.XLS
- Monthly sales for a chain of stereo retailers are
listed in this file. - They cover the period form the beginning of 1995
to the end of 1998, during which there was no
upward or downward trend in sales and no clear
seasonal peaks or valleys. - This behavior is apparent in the time series
chart of sales shown on the next slide. It is
possible that this series is random. - Does a runs test support this conjecture?
3Time Series Plot of Stereo Sales
4Random Model
- The simplest time series is the random model.
- In a random model the observations vary around a
constant mean, have a common variance, and are
probabilistically independent of one another. - How can we tell whether a time series is random?
- There are several checks that can be done
individually or in tandem. - The first of these is to plot the series on a
control chart. If the series is random it should
be in control.
5Runs Test
- The runs test is the second check for a random
series. - A run is a consecutive sequence of 0s and 1s.
- The runs test checks whether this is about the
right number of runs for a random series.
6Calculations
- To do a runs test in Excel we use StatPros Runs
Test procedure. - We must specify the time series variable (Sales)
and the cutoff value for the test, which can be
the mean, median or a user specified value. In
this case we select the mean to obtain this
sample of output.
7Output
- Note that StatPro adds two new variables,
Sales_High and Sales_NewRun, as well as the
elements for the test. - The values in the Sales_High are 1 or 0 depending
on whether the corresponding sales value are
above or below the mean. - The values in the Sales_NewRun column are also 1
or 0, depending on whether a new run starts in
that month.
8Output -- continued
- The rest of the output is fairly straightforward.
- We find the number of observations above the
mean, number of runs, mean for the observed
number of runs, the standard deviation for the
observed number of runs and the Z-value. We then
can find the two-sided p-value. - The output shows that there is some evidence of
not enough runs. - The expected number of runs under randomness is
24.8333 and there are only 20 runs for this
series.
9Conclusion
- The conclusion is that sales do not tend to
zigzag as much as a random series - highs tend
to follow highs and lows tend to follow lows -
but the evidence in favor of nonrandomness is not
overwhelming.
10 11The Problem
- The runs test on the stereo sales data suggests
that the pattern of sales is not completely
random. - Large values tend to follow large values, and
small values tend to follow small values. - Do autocorrelations support this conclusion?
12Autocorrelations
- Recall that successive observations in a random
series are probabilistically independent of one
another. - Many time series violate this property and are
instead autocorrelated. - The auto means that successive observations are
correlated with one other. - To understand autocorrelations it is first
necessary to understand what it means to lag a
time series.
13Autocorrelations
- This concept is easy to understand in
spreadsheets. - To lag by 1 month, we simply push down the
series by one row. - Lags are simply previous observations, removed by
a certain number of periods from the present time.
14Solution
- We use StatPros Autocorrelation procedure.
- This procedure requires us to specify a time
series variable (Sales), the number of lags we
want (we chose 6), and whether we want a chart of
the autocorrelations. This chart is called a
correlogram. - How large is a large autocorrelation?
- If the series is truly random, then only an
occasional autocorrelation should be larger than
two standard errors in magnitude.
15Solution -- continued
- Therefore, any autocorrelation that is larger
than two standard errors in magnitude is worth
our attention. - The only large autocorrelation for the sales
data is the first, or lag 1, the autocorrelation
is 0.3492. - The fact that it is positive indicates once again
that there is some tendency for large sales
values to follow large sales values and for small
sales values to follow small sales values. - The autocorrelations are less than two standard
errors in magnitude and can be considered noise.
16Lags and Autocorrelations for Stereo Sales
17Correlogram for Stereo Sales
18 19DEMAND.XLS
- The dollar demand for a certain class of parts at
a local retail store has been recorded for 82
consecutive days. - This file contains the recorded data.
- The store manager wants to forecast future
demands. - In particular, he wants to know whether there is
any significant time pattern to the historical
demands or whether the series is essentially
random.
20Time Series Plot of Demand for Parts
21Solution
- A visual inspection of the time series graph
shows that demands vary randomly around the
sample mean of 247.54 (shown as the horizontal
centerline). - The variance appears to be constant through time,
and there are no obvious time series patterns. - To check formally whether this apparent
randomness holds, we perform the runs test and
calculate the first 10 autocorrelations. The
numerical output and associated correlogram are
shown on the next slides.
22Autocorrelations and Runs Test for Demand Data
23Correlogram for Demand Data
24Solution -- continued
- The p-value for the run test is relatively large,
0.118 - although these are somewhat more runs
than expected - and none of the autocorrelations
is significantly large. - These findings are consistent with randomness.
For all practical purposes there is no time
series pattern to these demand data. - The mean is 247.54 and the standard deviation is
47.78.
25Solution -- continued
- The manager might as well forecast that demand
for any day in the future will be 247.54. If he
does so about 95 of his forecast should be
within two standard deviations (about 95) of the
actual demands.
26 27DOW.XLS
- Given the monthly Dow Jones data in this file,
check that it satisfies the assumptions of a
random walk, and use the random walk model to
forecast the value for April 1992.
28Random Walk Model
- Random series are sometimes building blocks for
other time series models. - The random walk model is an example of this.
- In the random walk model the series itself is not
random. However, its differences - that is the
changes from one period to the next - are random. - This type of behavior is typical of stock price
data.
29Solution
- The Dow Jones series itself is not random, due to
upward trend, so we form the differences in
Column C with the formula B7-B6 which is copied
down column C. The difference can be seen on the
next slide. - A graph of the differences (see graph following
data) show the series to be a much more random
series, varying around the mean difference 26.00.
- The runs test appears in column H and shows that
there is absolutely no evidence of nonrandom
differences the observed number of runs is
almost identical to the expected number.
30Differences for Dow Jones Data
31Time Series Plot of Dow Differences
32Solution -- continued
- Similarly, the autocorrelations are all small
except for a random blip at lag 11. - Because the values are 11 months apart we would
tend to ignore this autocorrelation. - Assuming the random walk model is adequate, the
forecast of April 1992 made in March 1992 is the
observed March value, 3247.42, plus the mean
difference, 26.00 or 3273.42. - A measure of the forecast accuracy is the
standard deviation of 84.65. We can be 95
certain that our forecast will be within the
standard deviations.
33Additional Forecasting
- If we wanted to forecast further into the future,
say 3 months, based on the data through March
1992, we would add the most recent value,
3247.42, to three times the mean difference,
26.00. - That is, we just project the trend that far into
the future. - We caution about forecasting too far into the
future for such a volatile series as the Dow.
34 35HAMMERS.XLS
- A retailer has recorded its weekly sales of
hammers (units purchased) for the past 42 weeks. - The data are found in the file.
- The graph of this time series appears below and
reveals a meandering behavior.
36The Plot and Data
- The values begin high and stay high awhile, then
get lower and stay lower awhile, then get higher
again. - This behavior could be caused by any number of
things. - How useful is autoregression for modeling these
data and how would it be used for forecasting?
37Autocorrelations
- A good place to start is with the
autocorrelations of the series. - These indicate whether the Sales variable is
linearly related to any of its lags. - The first six autocorrelations are shown below.
38Autocorrelations -- continued
- The first three of them are significantly
positive, and then they decrease. - Based on this information, we create three lags
of Sales and run a regression of Sales versus
these three lags. - Here is the output from this regression
39Autoregression Output with Three Lagged Variables
40Autocorrelations -- continued
- We see that R2 is fairly high, about 57, and
that se is about 15.7. - However, the p-values for lags 2 and 3 are both
quite large. - It appears that once the first lag is included in
the regression equation, the other two are not
really needed. - Therefore we reran the regression with only the
first lag include.
41Autoregression Output with a Single Lagged
Variable
42Forecasts from Aggression
- This graph shows the original Sales variable and
its forecasts
43Regression Equation
- The estimated regression equation is
Forecasted Salest 13.763 0.793Salest-1 - The associated R2 and se values are approximately
65 and 155.4. The R2 is a measure of the
reasonably good fit we see in the previous graph,
whereas the se is a measure of the likely
forecast error for short-term forecasts. - It implies that a short-term forecast could
easily be off by as much as two standard errors,
or about 31 hammers.
44Regression Equation -- continued
- To use the regression equation for forecasting
future sales values, we substitute known or
forecasted sales values in the right hand side of
the equation. - Specifically, the forecast for week 43, the first
week after the data period, is approximately 98.6
using the equation ForecastedSales43
13.763 0.793Sales42 - The forecast for week 44 is approximately 92.0
and requires the forecasted value of sales in
week 43 in the equation ForecastedSales44
13.763 0.793ForecastedSales43
45Forecasts
- Perhaps these two forecasts of future sales are
on the mark and perhaps they are not. - The only way to know for certain is to observe
future sales values. - However, it is interesting that in spite of the
upward movement in the series, the forecasts for
weeks 43 and 44 are downward movements.
46Regression Equation Properties
- The downward trend is caused by a combination of
the two properties of the regression equation. - First, the coefficient of Salest-1, 0.793, is
positive. Therefore the equation forecasts that
large sales will be followed by large sales (that
is, positive autocorrelation). - Second, however, this coefficient is less than 1,
and this provides a dampening effect. - The equation forecasts that a large will follow a
large, but not that large.
47- Regression-Based Trend Models
48REEBOK.XLS
- This file includes quarterly sales data for
Reebok from first quarter 1986 through second
quarter 1996. - The following screen shows the time series plot
of these data. - Sales increase from 174.52 million in the first
quarter to 817.57 million in the final quarter. - How well does a linear trend fit these data?
- Are the residuals from this fit random?
49Time Series Plot of Reebok Sales
50Linear Trend
- A linear trend means that the time series
variable changes by a constant amount each time
period. - The relevant equation is Yt a bt Et where a
is the intercept, b is the slope and Et is an
error term. - If b is positive the trend is upward, if b is
negative then the trend is downward. - The graph of the time series is a good place to
start. It indicates whether a linear trend model
is likely to provide a good fit.
51Solution
- The plot indicates an obvious upward trend with
little or no curvature. - Therefore, a linear trend is certainly plausible.
- We use regression to estimate the linear fit,
where Sales is the response variable and Time is
the single explanatory variable. - The Time variable is coded 1-42 and is used as
the explanatory variable in the regression.
52Solution -- continued
- The Quarter variable simply labels the quarters
(Q1-86 to Q2-96) and is used only to label the
horizontal axis. - The following regression output shows that the
estimated equation is Forecasted Sales 244.82
16.53Time with R2 and se values of 83.8 and
90.38 million.
53Regression Output for Linear Trend
54Time Series Plot with Linear Trend Superimposed
- The linear trendline, superimposed on the sales
data, appears to be a decent fit.
55Solution -- continued
- The trendline implies that sales are increasing
by about 16.53 million per quarter during this
period. - The fit is far from perfect, however.
- First, the se value 90.38 million is an
indication of the typical forecast error. This is
substantial, approximately equal to 11 of the
final quarters sales - Furthermore, there is some regularity to the
forecast errors shown in the following plot.
56Time Series Plot of Forecasted Errors
57Plot Interpretation
- They zigzag more than a random series.
- There is probably some seasonal pattern in the
sales data, which we might be able to pick up
with a more sophisticated forecasting method. - However, the basic linear trend is sufficient as
a first approximation to the behavior of sales.
58- Regression-Based Trend Models
59INTEL.XLS
- This file contains quarterly sales data for the
chip manufacturing firm Intel from the beginning
of 1986 through the second quarter of 1996. - Each sales value is expressed in millions of
dollars. - Check that an exponential trend fits these sales
data fairly well. - Then estimate the relationship and interpret it.
60Time Series Plot of Sales with Exponential Trend
Superimposed
61The Time Series Plot of Sales
- The time series plot shows that sales are clearly
increasing at an increasing rate, which a linear
trend would not capture. - The smooth curve of the plot is an exponential
trendline, which appears to be an adequate fit. - Alternatively, we can try to straighten out the
data by taking the log of sales with Excels LN
function. - The following is a plot of the log data.
62Time Series Plot of Log Sales with Linear Trend
Superimposed
63The Time Series Plot of Log Sales
- This plot goes together logically with the time
series plot of Sales in the sense that if an
exponential trendline fits the original data
well, then a linear trendline will fit the
transformed data well, and vice versa. - Either is evidence of an exponential trend in the
sales data.
64Estimating the Exponential Trend
- To estimate the exponential trend, we run a
regression of the log of sales, LnSales, versus
Time. - A portion of the resulting data and output
appears below.
65Data Setup for Regression of Exponential Trend
66Regression Output for Exponential Trend
67Regression Output
- The regression output shows that the estimated
log of sales is given by Forecasted LnSales
5.6883 0.0657Time - Looking at the coefficient of Time, we can say
that Intels sales are increasing by
approximately 6.6 per quarter during this
period. - This translates to an annual percentage increase
of about 29. Perhaps the slight tailing off that
we see at the right indicates that Intel cant
keep up this fantastic rate forever.
68Regression Output -- continued
- It is important to view the R2 and se values with
caution. Each is based in log units not original
units. - To produce similar measures in original units, we
need to forecast sales in Column E. This is a two
step process. - First, we forecast the log sales.
- Then we take the antilog with Excels EXP
function. The specific formula is
EXP(J18J19A4).
69Regression Output -- continued
- As usual, R2 is the square of the correlation
between actual and fitted sales values, so the
formula in cell J22 is CORREL(Sales,FittedSales
)ˆ2. - Then se is the square root of the sum of squared
residuals divided by n-2. We can calculate this
in cell J23 by using Excels SUMSQ(sum of
squares) function SQRT(SUMSQ(ResidSAles)/40). - The R2 value of 0.988 indicates that there is a
very high correlation between the actual and
fitted sales values. In other words, the
exponential fit is a very good one.
70Regression Output -- continued
- However, the se value if 159.698 (in millions of
dollars) indicates the forecasts based on this
exponential fit could still be fairly far off.
71 72DOW.XLS
- We again look at the Dow Jones monthly data from
January 1988 through March 1992 contained in this
file. - How well do moving averages track this series
when the span is 23 months when the span is 12
months? - What about future forecasts, that is, beyond
March 1992?
73Moving Averages
- Perhaps the simplest and one of the most
frequently used extrapolation methods is the
method of moving averages. - To implement the moving averages method, we first
choose a span, the number of terms in each moving
average. - The role of span is very important. If the span
is large - say 12 months - then many observations
go into each average, and extreme values have
relatively little effect on the forecasts.
74Moving Averages -- continued
- The resulting series forecasts will be much
smoother than the original series. - For this reason the moving average method is
called a smoothing method.
75Moving Averages Method in Excel
- Although the moving averages method is quite easy
to implement with Excel, it can be tedious. - Therefore we can use the Forecasting procedure of
StatPro. This procedure lets us forecast with
many methods. - Well go through the entire procedure step by
step.
76Forecasting Procedure
- To use the StatPro Forecasting procedure, the
cursor needs to be in a data set with time series
data. - We use the StatPro/Forecasting menu item and
eventually choose Dow as the variable to analyze.
- We then see several dialog boxes, the first of
which is where we specify the timing.
77Timing Dialog Box
- In the next dialog box, we specify which
forecasting method to use and any parameters of
that method.
78Method Dialog Box
- We next see a dialog box that allows us to
request various time series plots, and finally we
get the usual choice of where to report the
output .
79The Output
- The output consists of several parts.
- First, the forecasts and forecast errors are
shown for the historical period of data. - Actually, with moving averages we lose some
forecasts at the beginning of the period. - If we ask for future forecasts, they are shown in
red at the bottom of the data series. - There are no forecast errors and to the left we
see the summary measures.
80Moving Averages with Output Span 3
81Moving Averages with Output Span 12
82The Output -- continued
- The essence of the forecasting method is very
simple and is captured in column F of the output.
It used the formula AVERAGE(E2E4) in cell F5,
which is then copied down.
83The Plots
- The plots show the behavior of the forecasts.
- The forecasts with span 3 appear to track the
data better, whereas the forecasts with span 12
is considerably smoother - it reacts less to ups
and downs of the series.
84Moving Averages Forecasts with Span 3
85Moving Averages with Forecasts Span 12
86In Summary
- The summary measures MAE, RMSE, and MAPE confirm
that moving averages with span 3 forecast the
known observations better. - For example, the forecasts are off by about 3.6
with span 3, versus 7.7 with span 12. - Nevertheless, there is no guarantee that a span
of 3 is better for forecasting future
observations.
87 88EXXON.XLS
- This file contains data on quarterly sales (in
millions of dollars) for the period from 1986
through the second quarter of 1996. - The following chart is the time series chart of
these sales and shows that there is some evidence
of an upward trend in the early years, but that
there is no obvious trend during the 1990s. - Does a simple exponential smoothing model track
these data well? How do the forecasts depend on
the smoothing constant, alpha?
89Time Series Plot of Exxon Sales
90StatPros Exponential Smoothing Model
- We start by selecting the StatPro/Forecasting
menu item. - We first specify that the data are quarterly,
beginning in quarter 1 of 1986, we do not hold
out any of the data for validation, and we ask
for 8 quarters of future forecasts. - We then fill out the next dialog box like this
91Method Dialog Box
- That is, we select the exponential smoothing
option, elect the Simple option choose smoothing
constant (0.2 was chosen here) and elect not to
optimize, and specify that the data are not
seasonal.
92StatPros Exponential Smoothing Model -- continued
- On the next dialog sheet we ask for time series
charts of the series with the forecasts
superimposed and the series of forecast errors. - The results appear in the following three
figures. - The heart of the method takes place in the
columns F, G, and H of the first figure. The
following formulas are used in row 6 of these
columns. AlphaE6(1-Alpha)F5
F5 E6-G6
93StatPros Exponential Smoothing Model -- continued
- The one exception to this scheme is in row 2.
- Every exponential smoothing method requires
initial values, in this case the initial smoothed
level in cell F2. - There is no way to calculate this value because
the previous value is unknown. - Note that 8 future forecasts are all equal to the
last calculated smoothed level in cell F43. - The fact that these remain constant is a
consequence of the assumption behind simple
exponential smoothing, namely, that the series is
not really going anywhere. Therefore, the last
smoothed level is the best indication of future
values of the series we have.
94Simple Exponential Smoothing Output
95Forecast Series Error Charts
- The next figure shows the forecast series
superimposed on the original series. - We see the obvious smoothing effect of a
relatively small alpha level. - The forecasts dont track the series well but if
the zig zags are just random noise, then we dont
want the forecasts to track these random ups and
downs too closely. - A plot of the forecast errors shows some quite
large errors, yet the errors do appear to be
fairly random.
96Plot of Forecasts from Simple Exponential
Smoothing
97Plot of Forecast Errors from Simple Exponential
Smoothing
98Summary Measures
- We see several summary measures of the forecast
errors. - The RMSE and MAE indicate that the forecasts from
this model are typically off by a magnitude of
about 2300, and the MAPE indicates that this
magnitude is about 7.4 of sales. - This is a fairly sizable error. One way to try to
reduce it is to use a different smoothing
constant.
99Summary Measures -- continued
- The optimal alpha level for this example is
somewhere between 0.8 and 0.9. This figure shows
the forecast series with alpha 0.85.
100Summary Measures -- continued
- The forecast series now appears to tack the
original series very well - or does it? - A closer look shows that we are essentially
forecasting each quarters sales value by the
previous sales value. - There is not doubt that this gives lower summary
measures for the forecast errors, but it is
possibly reacting too quickly to random noise and
might not really be showing us the basic
underlying patter of sales that we see with alpha
0.2.
101 102DOW.XLS
- We return to the Dow Jones data found in this
file. - Again, these are average monthly closing prices
from January 1988 through March 1992. - Recall that there is a definite upward trend in
this series. - In this example, we investigate whether simple
exponential smoothing can capture the upward
trend. - The we see whether Holts exponential smoothing
method can make an improvement.
103Solution
- This first graph shows how a simple exponential
smoothing model handles this trend, using alpha
0.2. - The graphs summary error messages are not bad
(MAPE is 5.38), but the forecasted series is
obviously lagging behind the original series. - Also, the forecasts for the next 12 months are
constant, because no trend is built into the
model. - In contrast, the following graph shows forecasts
from Holts model with alpha beta 0.2. The
forecasts are still far from perfect (MAPE is now
4.01), but at least the upward trend has been
captured
104Plot of Forecasts from Simple Exponential
Smoothing
105Plot of Forecasts from Holts Model
106Holts Method
- The exponential smoothing method generally works
well if there is no obvious trend in the series.
But if there is a trend, then this method lags
behind. - Holts model rectifies this by dealing with trend
explicitly. - Holts model includes a trend term and a
corresponding smoothing constant. This new
smoothing constant (beta) controls how quickly
the method reacts to perceived changes in the
trend.
107Using Holts Method
- To produce the output from Holts method with
StatPro we proceed exactly as with the simple
exponential procedure. The only difference is
that we now get to choose two smoothing
parameters. - The output is also very similar to simple
exponential smoothing output, except that there
is now an extra column (column G) for the
estimated trend.
108Portion of Output from Holts Method
109Smoothing Constants
- It was mentioned that the smoothing constants
used above are not optimal. - If we use an StatPros optimize option to find
the best alpha for simple exponential smoothing
or the best alpha and beta for the Holts
method. - In this case we find 1.0 and 0.0 for the
smoothing constants. - Therefore, the best forecast for next months
value is the months value plus a constant trend.
110 111COCACOLA.XLS
- The data in this spreadsheet represents quarterly
sales for Coca Cola from the first quarter of
1986 through the second quarter of 1996. - As we might expect there has been an upward trend
in sales during this period and there is also a
fairly regular seasonal pattern as shown in the
time series plot of sales. - Sales in warmer quarters, 2 and 3, are
consistently higher than in the colder quarters,
1 and 4. - How well can Winters method track this upward
tend and seasonal pattern?
112Time Series Plot of Coca Cola Sales
113Seasonality
- Seasonality if defined as the consistent
month-to-month (or quarter-to-quarter)
differences that occur each year. - The easiest way to check if there is seasonality
in a time series is to look at a plot of the
times series to see if it has a regular pattern
of up and/or downs in particular months or
quarters. - There are basically two extrapolation methods for
dealing with seasonality - We can use a model that takes seasonality into
account or - We can deseasonalize the data, forecast the data,
and then adjust the forecasts for seasonality.
114Seasonality -- continued
- Winters model is of the first type. It attacks
seasonality directly. - Seasonality models are usually classified as
additive or multiplicative. - An additive model finds seasonal indexes, one for
each month, that we add to the monthly average to
get a particular months value. - A multiplicative model also finds seasonal
indexes, but we multiply the monthly average by
these indexes to get a particular months value. - Either model can be used but multiplicative
models are somewhat easier to interpret.
115Winters Model of Seasonality
- Winters model is very similar to Holts model -
it has level and trend terms and corresponding
smoothing constants alpha and beta - but it also
has seasonal indexes and a corresponding
smoothing constant. - The new smoothing constant controls how quickly
the method reacts to perceived changes in the
pattern of seasonality. - If the constant is small, the method reacts
slowly if the constant is large, it reacts more
quickly.
116Using Winters Method
- To produce the output from Winters method with
StatPro we proceed exactly as with the other
exponential methods. - In particular, we fill out the second main dialog
box as shown below.
117Portion of Output from Winters Method
118The Output
- The optimal smoothing constants (those that
minimize RMSE) are 1.0, 0.0 and 0.244.
Intuitively, these mean react right away to
changes in level, never react to changes in
trend, and react fairly slowly to changes in the
seasonal pattern. - If we ignore seasonality, the series is trending
upward at a rate of 67.107 per quarter. - The seasonal pattern stays constant throughout
this 10-year period. - The forecast series tracks the actual series
quite well.
119Plot of the Forecasts from Winters Method
- The plot indicates that Winters method clearly
picks up the seasonal pattern and the upward
trend and projects both of these into the future.
120In Conclusion
- Some analysts would suggest using more typical
values for the constants such as alphabeta0.2
and 0.5 for the seasonality constant. - To see how these smoothing constants would affect
the results, we can simply substitute their
values into the range B6B8. - The summary measures get worse, yet the plot
still indicates a very good fit.
121- Deseasonalizing The Ratio-to-Moving-Averages
Method
122COCACOLA.XLS
- We return to this data file that contains the
sales history from 1986 to quarter 2 of 1996. - Is it possible to obtain the same forecast
accuracy with the ratio-to-moving-averages method
as we obtained with the Winters method?
123Ratio-to-Moving-Averages Method
- There are many varieties of sophisticated methods
for deseasonalizing time series data but they are
all variations of the ratio-to-moving-averages
method. - This method is applicable when we believe that
seasonality is multiplicative. - The goal is to find the seasonal indexes, which
can then be used to deseasonalize the data. - The method is not meant for hand calculations and
is straightforward to implement with StatPro.
124Solution
- The answer to the question posed earlier depends
on which forecasting method we use to forecast
the deseasonalized data. - The ratio-to-moving-averages method only provides
a means for deseasonalizing the data and
providing seasonal indexes. Beyond this, any
method can be used to forecast the deseasonalized
data, and some methods work better than others. - For this example, we will compare two methods
the moving averages method with a span of 4
quarters, and Holts exponential smoothing method
optimized.
125Solution -- continued
- Because the deseasonalized data still has a a
clear upward trend, we would expect Holts method
to do well and we would expect the moving
averages forecasts to lag behind the trend. - This is exactly what occurred.
- To implement the latter method in StatPro, we
proceed exactly as before, but this time select
Holts method and be sure to check Use this
deseasonalizing method. We get a large selection
of optional charts.
126Ration-to-Moving-Averages Output
- Here are the summary measures for forecast errors.
- This output shows the seasonal indexes from the
ratio-to-moving-averages method. They are
virtually identical to the indexes found using
Winters method.
127Ratio-to-Moving Averages Output
128Forecast Plot of Deseasonalized Series
- Here we see only the smooth upward trend with no
seasonality, which Holts method is able to track
very well.
129The Results of Reseasonalizing
130Summary Measures
- The summary measures of forecast errors below are
quite comparable to those from Winters method. - The reason is that both arrive at virtually the
same seasonal pattern.
131 - Estimating Seasonality with Regression
132COCACOLA.XLS
- We return to this data file which contains the
sales history of Coca Cola from 1986 to quarter 2
of 1996. - Does a regression approach provide forecasts that
are as accurate as those provided by the other
seasonal methods in this chapter?
133Solution
- We illustrate a multiplicative approach, although
an additive approach is also possible. - The data setup is as follows
134Solution
- Besides the Sales and Time variables, we need
dummy variables for three of the four quarters
and a Log_Sales variable. - We then can use multiple regression, with the
Log_sales as the response variable and Time, Q1,
Q2, and Q3 as the explanatory variables. - The regression output appears as follows
135Regression Output
136Interpreting the Output
- Of particular interest are the coefficients of
the explanatory variables. - Recall that for a log response variable, these
coefficients can be interpreted as percent
changes in the original sales variable. - Specifically, the coefficient of Time means that
deseasonalized sales increase by 2.4 per
quarter. - This pattern is quite comparable to the pattern
of seasonal indexes we saw in the last two
examples.
137Forecast Accuracy
- To compare the forecast accuracy of this method
with earlier examples, we must go through several
steps manually. - The multiple regression procedure in StatPRo
provide fitted values and residuals for the log
of sales. - We need to take these antilogs and obtain
forecasts of the original sales data, and
subtract these from the sales data to obtain
forecast errors in Column K. - We can then use the formulas that were used in
StatPros forecasting procedure to obtain the
summary measures MAE, RMSE, and MAPE.
138Forecast Errors and Summary Measures
139Forecast Accuracy -- continued
- From the summary measures it appears that the
forecast are not quite as accurate. - However, looking at the plot below of the
forecasts superimposed on the original data shows
us that the method again tracks the data very
well.