Simulating Item Forecasts

This chapter provides an overview of item simulation and discusses how to:

Define forecast simulation options.
View simulation results.

Understanding Item Simulation

Forecast item simulation uses forecasts in a controlled and interactive environment to perform a what-if analysis. Use simulations to quickly and efficiently compare the effects of using different forecast models and parameters for an item. You typically use simulation on an exception basis when trying to resolve problems with an item's forecast.

You can try several different scenarios to determine the best settings for a particular forecast item at any level in the view structure. However, unless you have a working knowledge of forecasting and statistics, you should not make adjustments to forecast models. The system provides most of the statistical input during forecast processing and determines the best-fit model.

Using simulation, you compare forecasts that the system develops for a selected forecast item. You can use one or a combination of different models, forecast effective periods, control groups, and item or group seasonality as part of the simulation. You can also calculate or recalculate the forecast using any combination of model or model components with a future effectivity date.

Note. You cannot simulate life profile items using this feature; however, you can change the life profile and rebuild it using the Life Profile Weight Calculations Process page.

This section discusses:

Forecast models.
Forecast model statistical parameters.
Forecast model accuracy measurements.

Forecast Models

These elements make up forecast simulation:

Forecast Models.

Predefined sets of calculations the system uses to make predictions about the future demand for an item. Along with their calculations, models use statistical values that have been defined at the control-group and item level to perform their calculations. The models define how you control forecasts.
Model Parameters.

Most models that you use with PeopleSoft Demand Planning have parameters that provide you with a way to control how the model processes the forecast during reoptimization of forecasts. Model parameters are also predefined statistical calculations.

Note. It is not necessary for you to maintain model parameters unless you are knowledgeable about using forecasting statistics. The system calculates and simulates these forecast values for you and determines the best-fit model. You normally do not adjust the parameters.
Accuracy Measures.

Measure how well a forecast model is predicting the forecast. Each accuracy measurement is based on different statistical criteria. Understanding how and why the system applies an accuracy measurement helps you to determine which forecast model best meets the business needs. Depending on the forecasting requirements, these measurements provide you with a meaningful measure of model accuracy.

See Forecast Model Accuracy Measurements.

The forecast models, accuracy measures, and statistical parameters that you use in forecasting are provided through ForecastX™. This forecasting tool makes it possible to perform complex forecasting and statistical calculations using PeopleSoft Demand Planning. You can perform statistical analysis on any single or multiple sets of data series. You can also perform statistical analysis on a series of observations or process a set of series in a batch-process or batch mode.

More descriptions of the calculations and formulas used by ForecastX™ are available.

See Business Forecasting with Accompanying Excel-Based ForecastX™ Software, Fourth Edition, 2002

When you simulate item forecast models, the system provides simulation results related to the selected model's statistics.

This section discusses:

Adaptive exponential smoothing
Box Jenkins
Census X11 Additive
Croston
Double exponential smoothing-Brown
Double exponential smoothing-Holt
Exponential smoothing
Holt-Winters (Additive/Multiplicative)
Linear regression
Moving average
Triple exponential smoothing-Brown
Weighted moving average

Adaptive Exponential Smoothing

The adaptive exponential smoothing model is a derivative of exponential smoothing. In adaptive exponential smoothing, the alpha value changes systematically from period to period to allow for pattern changes in the historical data. Exponential smoothing uses one level (alpha value) to create the forecast, which adapts to data changes during the life of the forecast.

These model-specific statistical parameters are available with the adaptive exponential smoothing forecast model:

Beta (Smoothing)
Tracking Signals 1 - 6

Note. You can change the Beta (Smoothing) statistical parameter before you calculate the simulations.

Box Jenkins

Box Jenkins is a multistep model that builds strategy for analyzing and forecasting time series data by looking for an adequate model in the autoregression moving average and autoregression integrated moving average model group. PeopleSoft Demand Planning optimization automatically chooses the best method within this model that fits the data.

These model-specific statistical parameters are available with the Box Jenkins forecast model:

ACF Boundary (autocorrelation function boundary)
AR Parameter (autoregression parameter)
AR Seasonal Parameter (autoregression seasonal parameter)
Constant Parameter
MA Parameter (moving average parameter)
MA Seasonal Parameter (moving average seasonal parameter)
Non Seasonal D
Non Seasonal P
Non Seasonal P Decay
Non Seasonal P Initial
Non Seasonal Q
Non Seasonal Q Decay
Non Seasonal Q Initial
Number of Lags
Partial Coefficient 1 - x (the number of lags)
Auto Coefficient 1 - x (the number of lags)
Seasonal D
Seasonal Length
Seasonal P
Seasonal P Decay
Seasonal P Initial
Seasonal Q
Seasonal Q Decay
Seasonal Q Initial
Tracking Signal 1 - 6

You can change the statistical parameter for these parameters before you calculate the simulation:

d value - Differencing
p value - Moving Average
q value - Auto Regression
Seasonal d value - Differencing
Seasonal p value - Moving Average
Seasonal q value - Auto Regression

Census X11 (Additive/Multiplicative)

The Census X11 model refines the seasonal-decomposition method. It seasonally adjusts and decomposes forecast data through a series of predefined steps: seasonal, trend, cycle, and random or irregular. A seasonal component of a time series occurs regularly, such as an annual holiday, while cycle duration varies from cycle to cycle. Economic variables are cyclical and a depression occurs at irregular intervals.

Average Irregular Statistic
Trend Cycle 1- x (listed number of periods)
Irregular Statistic - x (listed number of periods)
Tracking Signal 1- 6

Note. You cannot change statistical parameters for the Census X11 model.

Croston

The Croston model handles sporadic demand data. This is demand where a manufacturer does not have precursors when it places an order.

Another example of sporadic data is when demand often goes to zero, even though the average demand can be for several units. For example, in a three-period moving average, February, March, and April actual values evenly determine May's forecast. Depending on the weight given to the data series, May's forecast relies heavily on April's actual values and fitted values. Exponential smoothing is most effective as a forecasting method when cyclical and irregular influences comprise the main effects on the time series values.

These model-specific statistical parameters are available with the Croston forecast model:

Alpha
Smoothing Weight for Non Zero Demand Values
Tracking Signal 1 - 6

Note. You can change the smoothing weight for non zero demand values statistical parameter before you calculate the simulations.

Double Exponential Smoothing-Brown

The double exponential smoothing-Brown model attempts to create a linear equation. The model performs two simple exponential smoothing forecasts and then adjusts for the linear trend in the data. Double exponential smoothing-Brown is similar to double exponential smoothing in the fact that the goal is to create a linear trend, but it does so without adding additional parameters to the equation.

Because forecasts can be expressed as a function of the single and double-smoothed constants, the procedure is known as double exponential smoothing. This model is most appropriate for data that shows a linear trend over time.

The model-specific statistical parameters available with the double exponential smoothing-Brown forecast model are the same as those for the Croston model.

Note. You can change the smoothing weight statistical parameter before you calculate the simulations.

Double Exponential Smoothing-Holt

The double exponential smoothing-Holt model is similar in principle to simple exponential smoothing. It calculates alpha (the level component) to measure the level in the forecast. Double exponential smoothing-Holt also adds the parameter gamma (the trend component) to create a linear trend in the forecast. This equation is similar to a linear regression line and is useful when a product is experiencing an exponential growth or decline, while not exhibiting recognizable seasonality. If the data is dynamic and does not change due to seasonal factors, then this model is beneficial.

These model-specific statistical parameters are available with the double exponential smoothing-Holt forecast model:

Final Level
Alpha
Final Trend
Gamma
Tracking 1 - 6
Level Smoothing
Trend Smoothing

Note. You can change the statistical parameters for level and trend smoothing before you calculate the simulations.

Exponential Smoothing

The exponential smoothing model uses historical and fitted forecast data to generate the next forecast. The system creates the fitted line by the forecasting technique and is used as a base for the forecast. In exponential smoothing, recent values are given more weight in forecasting than older observations.

These model-specific statistical parameters are available with the exponential smoothing forecast model:

Final Level
Alpha (Smoothing)
Tracking 1 - 6

Note. You can change the Alpha (Smoothing) statistical parameter before you calculate the simulations.

Holt-Winters (Additive/Multiplicative)

The Holt-Winters model uses an exponential smoothing technique that incorporates growth and seasonality into the forecast by producing seasonal lift factors for each seasonal period. The model is similar in principle to simple exponential smoothing, where it calculates alpha to measure the level of trend in the forecast. However, the Holt-Winters model also adds the parameter, gamma, to create a linear trend in the forecast as well as the parameter, beta, for seasonality.

These model-specific statistical parameters are available with the Holt-Winters forecast model:

Final Level
Alpha (Smoothing)
Final Trend
Gamma (Trend)
Beta (Seasonal)
Tracking 1 - 6

Note. You can change the statistical parameters for alpha smoothing, beta seasonal, and gamma trend before you calculate the simulation.

Linear Regression

The linear regression model uses an equation to analyze the relationship between two or more quantitative variables in order to predict one from the others. The model measures the relationship between two variables: X and Y — where X is the independent variable and Y is the dependent variable. A particular observation of Y depends on X and an additional random error.

These model-specific statistical parameters are available with the linear regression forecast model:

Intercept
Slope
Elasticity
Overall F-Test
T-Test for Intercept
T-Test for Slope
Tracking 1 - 6

Moving Average

The moving average model averages a time series over a specific number of preceding periods. When you add a new value, the system drops the last value from the calculation, so that the specific number of preceding periods remains a constant. During simulation, the system uses the average to determine the minimum error moving average model.

This model does not effectively handle significant trends in data, and all historical data must be stored to create the moving average. In the case of time series data, use this model to eliminate unwanted fluctuations, thereby smoothing the time series. The system determines the appropriate number of preceding periods by selecting the amount of periods that yields the least amount of error.

These model-specific statistical parameters are available with the moving average forecast model:

Final Level
Alpha
Moving Average Periods

Note. You can change the statistical parameter for moving average periods before you calculate the simulation.

Triple Exponential Smoothing-Brown

The triple exponential smoothing-Brown model attempts to create a linear equation. It performs two simple exponential smoothing forecasts and then adjusts for the linear trend in the data. Triple exponential smoothing-Brown is similar to double exponential smoothing, because the goal is to create a linear trend, but it does so without adding additional parameters to the equation. Forecasts can be expressed as a function of the single and double smoothed constants; therefore, the procedure is known as double exponential smoothing. This model is most appropriate for data that shows a linear trend over time.

The Alpha model-specific statistical parameters is available with the triple exponential smoothing-Brown forecast model:

Note. You can change the smoothing weight statistical parameter before you calculate the simulation.

Weighted Moving Average

The weighted moving average model moves averages of moving averages. Rather than replace the oldest observations within the data, the model replaces the oldest moving average with the most recent moving average. The forecast model helps to overcome the strong affect of extreme values within a time series by assigning current data more weight than older data. The start and history parameters are the same as those in moving averages.

These model-specific statistical parameters are available with the weighted moving average forecast model:

Final Level
Alpha
Moving Average Periods

Note. You can change the statistical parameter for moving average periods before you calculate the simulation.

Forecast Model Statistical Parameters

Model statistical parameters are values that the system can either determine during forecast processing, or in some cases that you can use to tune the forecast manually. Each forecast model has a set of statistical parameters associated with it. When you enter or override these parameters, you limit the forecast model to results that you predefine instead of using the system-defined values.

Note. Unless you are skilled with statistical forecasting, let the system set these values and determine the best-fit model.

Use the Model Parameters page to view the statistical values. In some cases, you can change these values before you calculate the simulation. Click the Statistics link, after you calculate the simulation to view the system-calculated values. You cannot changes these values.

Note. To see model statistical parameters that you can update manually, review the Forecast Models section.

See Forecast Models.

ACF Boundary (autocorrelation function boundary)	Represents the upper and lower bounds on the autocorrelation function chart that the system uses to estimate parameters. Autocorrelation is the correlation of errors in a data series.
Alpha	Represents the level smoothing constant for the exponential smoothing method family. Alpha is a hypothesis that is held to be true until sufficient evidence proves it to be false.
Alpha (Smoothing)	Represents the level smoothing constant for the exponential smoothing method family. The system applies the alpha smoothing constant to the most recent forecast errors.
AR Parameter (autoregression parameter)	Represents the coefficient corresponding to autoregression. Autoregression relates data in one data series to values in previous time periods.
AR Seasonal Parameter (autoregression seasonal parameter)	Represents the seasonal coefficient corresponding to autoregression. The parameter relates data in one data series to values in previous time periods. Seasonality sensitivity helps prevent instability of seasonal forecasts.
Auto Coefficient 1 to Auto Coefficient 4	Represents a standardized measure of correlation between two variables. The value is the dependence between two variables of the same time series at different time periods. Autocorrelation helps to determine whether there is a causal connection between two variables even though there is a time lag between their occurrences.
Beta	Represents the distribution for continuous random variables that are constrained to lie between 0 and 1. Beta is characterized by two parameters: shape and scale.
Beta (Seasonal)	Represents the seasonal smoothing constant for the exponential smoothing method family.
Beta (Smoothing)	Represents the smoothing constant for the exponential smoothing method family.
Constant Parameter	Represents the constant in the Box Jenkins formula.
D Value - Differencing	Represents the degree of nonseasonal differencing. Differencing is when you define a constant mean by using the differences of two data series when a data series is not stationary and does not have a constant mean.
Elasticity	Represents the percentage of change that you want in a dependent variable with a one percent change in an independent variable.
Final Level	Represents the exponential smoothing value.
Final Trend	Represents the final trend value in the double-Holt and Holt-Winters exponential smoothing models.
Gamma	Represents the distribution for continuous random variables that are constrained to be greater or equal to zero. The distribution is characterized by two parameters: shape and scale. The gamma distribution is often used to model data which is positively skewed.
Gamma (Trend)	Represents the trend smoothing constant for the exponential smoothing method family.
Intercept	Represents the constant in the regression equation. This is the point where a regression line intercepts the vertical axis, if the horizontal axis has a true zero origin.
Irregular Statistic	Represents the randomness of historical data in a forecast that uses the Census X 11 Additive model.
Level Smoothing	Represents the level smoothing constant for the exponential smoothing method family.
MA Parameter (moving average parameter)	Represents the coefficient corresponding to the moving average. This is the average of the most recent span of data.
MA Seasonal Parameter (moving average seasonal parameter)	Represents the seasonal coefficient corresponding to the moving average. This is the average of the most recent span of seasonal data.
Moving Average Periods	Represents the number of observations to use for the moving average method.
Number of Lags	Represents the number of periods of lag time to return as the number lag values property for the Box Jenkins model.
Non Seasonal P	Represents the degree in the final selected model. This is the degree of freedom for the nonseasonal autoregression data. Degree of freedom is the number of effective sample points in the data.
Non Seasonal P Decay	Represents the difference between the degrees in the identification stage and in the final selected model. This is the degree of freedom change for the nonseasonal autoregression data.
Non Seasonal P Initial	Represents the degree in the identification stage. This is the initial degree of freedom for the nonseasonal autoregression data.
Non Seasonal Q	Represents the degree of the nonseasonal moving average. The value is the degree of freedom in the final selected model.
Non Seasonal Q Decay	Represents the difference of the degrees in the identification stage and in the final selected model. This is the degree of change for the nonseasonal moving average.
Non Seasonal Q Initial	Represents the degree in the identification stage. This is the initial degree of the nonseasonal moving average.
Non Seasonal D	Represents the degree of nonseasonal differencing.
Overall F-Test	Represents a test value used in analysis of variance or regression. The value is the ratio of the variance between groups to the variance within groups. If there is not a difference between the groups, the statistic follows an F distribution. The test value is expected value is one.
P Value - Moving Average	Represents the degree of freedom for nonseasonal autoregression data in the final selected model.
Partial Coefficient 1 to Partial Coefficient 4	Represents an estimate of the additional correlation between the data value at time 't' and the data value at time (t-k). This is after you adjust for the correlation of the values between time 't' and the data value at time (t-k).
Q Value - Auto Regression	Represents the degree of the nonseasonal moving average in the final selected model.
Seasonal D	Represents the degree of seasonal differencing.
Seasonal d Value - Differencing	Represents the degree of seasonal differencing.
Seasonal Length	Represents the length of the seasonal pattern. For example, seasonal quarterly data might use 4 as the value for seasonal length. For nonseasonal data, values are 0 or 1.
Seasonal P	Represents the degree of seasonal autoregression in the final selected model.
Seasonal P Decay	Represents the difference between the degrees in the identification stage and in the final selected model. This is the degree change of the seasonal autoregression.
Seasonal P Initial	Represents the value for the initial seasonal autoregression parameter the system uses during the identification stage.
Seasonal p Value - Moving Average	Represents the degree of the seasonal autoregression in the final selected model.
Seasonal Q	Represents the degree of the nonseasonal moving average in the final selected model.
Seasonal Q Decay	Represents the difference between the degrees in the identification stage and in the final selected model. This is the degree of change for the seasonal moving average.
Seasonal Q Initial	Represents the initial seasonal moving average parameter the system uses during the identification stage.
Seasonal q Value - Auto Regression	Represents the degree of the seasonal moving average in the final selected model.
Smoothing Weight	Represents the weight that is applied to smoothing alpha, gamma, or beta data.
Smoothing Weight for Non Zero Demand Values	Represents the weight that is applied to smoothing alpha, gamma, or beta data with nonzero demand. Note. Values of the smoothing weight do not affect the forecasted values for nonzero demand exponential smoothing model.
Slope	Represents the change in the dependent variable Y per unit change in the independent variable X.
T-Test for Intercept	Represents a test value for comparing the means of one or more normal distributions in which the variance is not known but must be estimated from the data.
T-Test for Slope	Represents a test value for comparing the means of one or more normal distributions in which the variance is not known but must be estimated from the data.
Tracking 1 - 6	Represents how close a forecast is tracking to what was forecasted and helps you to evaluate structural errors that exist in a forecast. Using tracking signals, you can identify situations where a forecasting model is over or under forecasting. When these situations occur, it is then possible to adjust the forecast to obtain additional accuracy. As a general rule the only time that a forecast should be adjusted is when the tracking signal is at + or - .5.
Trend Smoothing	Represents the trend smoothing constant for the exponential smoothing method family.

See Also

Business Forecasting with Accompanying Excel-Based ForecastX™ Software,Fourth Edition, 2002.

Forecast Model Accuracy Measurements

Accuracy measurements are forecasting assessments used by ForecastX™ that help you to determine how well the forecast is predicting demand. This section discusses:

Akaike Information Criterion
Bayesian Information Criterion
Durbin-Watson
LJung Box
Mean error
Mean absolute error
Mean absolute percentage error
Mean square error
Normality error
R-Square
Adjusted R-Square
Root mean square error
Standard deviation error
Sum squared error
Sum squared regression
Standard deviation
Theil's statistic
Chi square
T test
Cochrane-Orcutt

Akaike Information Criterion

Akaike Information Criterion (AIC) displays results that are constructed; therefore, as the number of independent variables increases, the simulation measure of accuracy results increase. The model uses a measure of the accuracy of the estimate and a measure of the principle of parsimony.

Bayesian Information Criterion

Bayesian Information Criterion (BIC) displays results based on order-estimation criteria used in the Box-Jenkins specification stage of model building. Models that minimize the BIC value are considered the most appropriate models. The AIC and BIC differ in their second terms. These are penalty functions for extra parameters. Often, AIC and BIC lead to the same model choice.

Durbin-Watson

Durbin-Watson displays results that test whether errors from a forecasting model are correlated. Autocorrelation describes the dependence between two variables of the same time series at different time periods. Autocorrelation occurs when there is dependence between successive error values. This dependence is also called a serial correlation.

Durbin-Watson results are between 0 and 4. Results closer to 0 indicate a positive autocorrelation, while results closer to 4 indicate a negative autocorrelation. Results close to 2 tend to reinforce that a correlation does not exist among the errors.

LJung Box

LJung Box displays results that test the overall autocorrelation of the fitted errors of a model and determine how a variable relates to itself when it is lagged one or more periods. This model evaluates the magnitude of a forecasting model's residual autocorrelations. If the system estimates the correct Auto Regression Integrated Moving Average (ARIMA) model, the LJung Box statistic tends to be smaller. A model with errors tends to increase the LJung Box statistic value.

Mean Error

The mean error displays results that are calculated as the average error value. The mean error is an average of the difference between the actual and forecasted values in time-series data. This model is less reliable in some cases than other accuracy statistic methods because there is the risk that large outliers can cancel each other out, producing a mean error near zero, which normally indicates a perfect fit.

Mean Absolute Error

The mean absolute error (MAE) displays results that consider the absolute values of errors from a forecasting model. Similar to the mean error model, use the mean absolute error model in situations where there is a great variance in the magnitude of errors.

This model takes negative values and replaces them with their absolute values. It de-emphasizes large outliers; therefore, negative and positive results do not cancel each other out. A zero is a perfect mean absolute error fit.

Mean Absolute Percentage Error

The mean absolute percentage error (MAPE) displays simulation results that produce a measure of relative overall fit. The system sums the absolute values of all percentage errors and the average is computed. Compared with the mean error and MAE models, the MAPE is a more meaningful measurement model.

The mean error is determined as the average error value affected by outliers, while the MAE model de-emphasizes outliers by their average. The MAPE also de-emphasizes outliers, but produces results more easily interpreted and calculated as the average absolute error in percentage terms.

Mean Square Error

The mean square error (MSE) is one of the traditional measures of forecast accuracy. Use the MSE when all of the errors are similar in magnitude. If the data contains one or two large errors, calculate the MAE, because using sum squares magnifies these errors. Also use the MAE or MSE to select the appropriate forecasting model by choosing the model that results in the smallest MAE or MSE.

Note. You cannot compare forecast models that use different data transformations, and you cannot compare MSE to MAE.

Normality Error

The Normality error displays simulation results for errors associated with the single time series normal distribution. A standard assumption of forecasting models is that the error terms are distributed normally.

R-Square

R-square displays results that evaluate what percentage of the up and down variation in the dependent variable is explained by variation in the independent variable. The system makes the evaluation by interpreting the R-square value that is reported in regression output.

R-square is the determining coefficient that indicates the fraction of the variation in the dependent variable that is explained by variation in the independent variable. The value can range between zero and one. The greater the value, the more likely that the variation in Y (the dependent variable) is explained by the variation in X (the independent variable).

Adjusted R-Square

Adjusted R-square displays results that indicate when negative effects of variables outweigh the positive effects. To achieve the value, the statistic uses a coefficient of determination adjusted for the number of parameters in the model. The coefficient establishes that when an explanatory variable is added to a model, R-square never decreases; therefore, R-square is the fraction of variance explained by the model.

Ideally, the measure of fit would decrease when useless variables are entered into the model as explanatory variables. If the measure of fit decreases every time a useless variable is entered into the model, then you can measure and determine which variables to keep and which to delete.

Root Mean Square Error

The root mean square error displays results that measure the data dispersion of a data series from its mean. The root mean square error is the square root of the variance. The variance is calculated by finding out how far each data point is from the mean (average) data point. The greater the value, the greater the dispersion of data.

Standard Deviation Error

The standard deviation error displays simulation results that provide the best estimate of the standard error, or deviation, of the residuals about the regression line. It is calculated by taking the square root of the residual mean squared error. Typically, a smaller standard error implies a better fit for the regression line. The standard deviation of error is the square root of the mean square error.

Sum Squared Error

The sum squared error displays results that measure accuracy by squaring and then adding the errors. The model is used to determine the accuracy of the forecasting model when the data points are similar in magnitude. The lower the sum squared error, the more accurate the forecast.

Sum Squared Regression

The sum squared regression (SSR) statistic is a sum of squared deviations. Specifically, it is the deviation between the forecasted value and the mean value, and is used in regression analysis. To calculate the SSR, sum the square of the difference between the forecasted value and mean. If the regression line is horizontal, the SSR is equal to zero. In all other cases, the SSR is positive. As the SSR increases, so does the relation between the observations and the independent variables (the regression line).

Standard Deviation

The standard deviation error is used in regression analysis to measure the average variation of the observed values around the forecast values. It verifies how far the forecasted values are from the actual values in the data set.

Theil's Statistic

Theil's statistic displays results that compare the accuracy of a forecast model to the actual model. The model uses the actual value of the last time period as the forecast. The closer Theil's statistic is to zero the more accurate the forecasting model. For example, suppose that Theil's statistic is equal to 1, the forecast is completely inaccurate. If it is equal to 0, then the forecasting model is a perfect fit.

Chi Square

Chi square is a commonly used test to compare actual data with fitted data according to a specific hypothesis, namely, the null hypothesis. The null hypothesis states that there is no significant difference between the fitted value and actual data. This statistics return the p-value of the test instead of the value of Chi square test. The range of p-value is from 0 to 1. The closer to 1, the better fit.

T Test

The T test indicates whether the parameter or data series contributes to the accuracy of the model. If the T test statistic is greater than 1.96, the parameter or series does contribute to the accuracy of the model. The T test is used to make a judgment about the variable m, or the slope coefficient of the linear regression model. This test indicates whether m is significantly different than zero. Again, if m is significantly different than zero, there is a relationship between the dependent and independent variable.

Cochrane-Orcutt

When there is a strong indication of autocorrelation errors in a forecasted model (especially, a regression model), it is not suitable to proceed with the ordinary least squares model. However, the alternative estimation to the new data set Yt - r × Y(t-1) will have better results for appropriated r. Cochrane-Orcutt is the model to estimate r.

Defining Forecast Simulation Options

Before you run a simulation, define the parameters that you want to use for the simulation.

This section discusses how to:

Establish simulation options.
Define simulation model parameters.

Pages Used to Define Forecast Simulation Options

Page Name	Object Name	Navigation	Usage
Simulation Options	DP_FORECASTSIM	Demand Planning, Process Forecast, Forecast Items, Simulation, Simulation Options	Establish simulation options.
Model Parameters	DP_FORECASTSIM2	Demand Planning, Process Forecast, Forecast Items, Simulation, Model Parameters	Define simulation model parameters that the system uses when it is not reoptimizing a forecast.

Establishing Simulation Options

Access the Simulation Options page.

Perform forecast simulations for one item at a time.

To define and apply simulation options for an item:

Make any changes to the basic information.

This might include selecting another control group or limiting the number of demand periods to the more recent periods.
If you are working with a seasonal item, select the Uses check box in the Seasonal group box.

This activates the Seasonal Group field, where you can select the group that you want the system to use for the forecast simulation. This can be useful when the item that you were simulating does not have sufficient demand to determine a good seasonality pattern. Selecting a seasonal group that has similar seasonal characteristics helps to stabilize the forecast and produce better results.
Select an option in the Forecast Options group box to indicate the type of forecast simulation that you want to use in the simulation.

Depending on the selection you make, the system creates models or groups of models available for selection in the Models group box. The default is Reoptimize Current. You do not have to make any additional selections when you select this option. However, if you select to reoptimize the current model, you can define applicable model parameters by accessing the Model Parameters page.
Select the models or groups of models in the Models group box that you want the system to use for simulations, when selecting the Best-Fit or Select Models options.
Click the Calculate button.

The system does the calculations and displays the simulated results on the Accuracy page. The results include accuracy measures for each model included in the simulation.

The system populates the Model To Use field with the best model. This selection is based on the accuracy measure that you select to yield the least amount of error. Define the accuracy measure in the Best-Fit Accuracy Statistic field for control groups in the forecast view. To compare a different model to the current model, select the model from the Model To Use field and click the Refresh button. If you ran the simulation using the Reoptimize Current option, the system displays the current model only.
Click the additional links on the Accuracy page to chart, review, and compare simulated data for the forecast item.

Depending on the number of models that the system uses in the simulation, you can switch from model to model to compare different forecasting techniques and make decisions about the item's forecast model.
Click the Apply button when you determine the forecast model that you want to use.

The system applies the simulated values to the item record. The model and its parameters also become the current model for the item. Click the Return link to go back to the selection page to select a different model or group of models, forecast options, or change the number of demand periods to use in the simulation.

Suspended Item	Displays whether the item has had zero demand for the most recent consecutive number of periods defined on the control group. If the item has been suspended, the box is selected.
Control Group	Select the control group from which you want to use parameters for this item. The forecast model uses these parameters when you perform the simulation. Control groups determine the major forecast process options and are required for all of the forecast items. The system associates the group with the forecast item using the Control Group field. To access the field, select Define Forecast Elements, Forecast Items, Maintain Forecast Items, Control. See Maintaining Forecast Items.
Demand Use	Defines the number of effective demand periods to use for this item. The demand use value cannot exceed the value in the Demand Total field for this item. These periods represent the span of time from the item's earliest recognized demand to the current period that you decide to use in simulating an item's forecast. You can change the value. Note. If the demand use value is 0, you must use the Select Models option in the Forecast Options group box to select a forecast model for simulation. You must also select Linear Regression, Exponential Smoothing, or Holt-Winters Additive/Multiplicative models as model types for the simulation.
Demand Total	Displays the total number of historical periods of demand for this item. The value represents the number of periods from the earliest demand period to the current period.
Reason Code	Enter a reason for the forecast adjustment if you use the simulation to adjust the forecast.
Models to Display	Enter the number of models that you want to display in the simulation results. For example, if you enter three, only three models appear on the results page. The system displays the models in ascending order according to the value in the Best Fit Accuracy Statistic field on the control group. That value is highlighted by an asterisk on the results page.
Reason Comment	Enter a comment associated with the reason code.
Calculate	Click to initialize the simulation. After processing, the Accuracy page appears. Use this page to access and review different types of simulation data.

Seasonality

Seasonality is a repetitive pattern of demand for an item during certain periods of a forecast. A seasonality group is a flexible, user-defined grouping of forecast items that helps you to analyze and understand patterns of demand over time and is made up of a set of forecast items that all have similar seasonality. The system aggregates the demand for the contributors to the group to come up with a seasonality profile.

Assigning a seasonality group for simulation is not required. You associate the seasonality group with the item on the Seasonality page when maintaining forecast items. You can also change groups.

To add a seasonality group for the simulation:

Select Uses.

This activates the Seasonality Group field. If the forecast item already uses a seasonality group, the check box is selected.
Select a seasonality group.
Select the Contributes check box to indicate that the item's demand should contribute to the aggregation of demand for the group.

Forecast Options

Defines how you want the system to process an item's forecast. The system processes the simulation based on the selections you specify.

Reoptimize Current	Optimizes the forecast using the model that is currently defined as the forecast model for the forecast item. When the system reoptimizes a forecast, it updates the item's forecast with any changes that you have made to the current model's parameters or to the item's data. During optimization, the system analyzes the data series and model components and produces simulation results for each of the model parameters. You can set the current model using the Forecast Model field on the Model Control page when maintaining forecast items. When you click the Calculate button, the system displays the simulation results for each accuracy statistic associated with the forecast model. Note. If you change the forecast model for the item, you might have to update values for the item using the Model Parameters page before the system performs the simulation.
Best Fit	Best-fit analysis is an automated system function that searches through various forecasting models to find a model that best fits a particular data set. When you select Best Fit, values include: Seasonal Models Non Seasonal Models All Models except Box Jenkins All Models Note. The default value for best-fit analysis is determined by the value of the Best Fit Seasonal Model field defined for the control group.
Select Models	Displays a list of models from which you can select one model or a combination of models to use in the item simulation. The models initially appear with those models that have been defined for the control group. Select or clear the check boxes to change the list of models that you want to use with the simulation. When you click the Calculate button simulation results appear on the Accuracy page for the models that you select. When you select Census X11 Additive, you cannot select another model.
Inhibited	Select to stop calculating a forecast for an item and set the forecast to zero. Select this option, for example, when you have a slow-moving item that you want to forecast manually. You can set up Work Queue messages to produce alerts when inhibited items have demand, so you do not have to include the item in forecasting. You might also select Inhibited when you have a high-frequency item that might be phased out in three months. By inhibiting the item, you can simulate a zero forecast that does not automatically replenish the item. Note. If you selected Inhibited for an item on the Model Control page, the option is not available for selection on this page; however, you can set the forecast to zero by simulating and applying the results.

Models

You can select which models you want to include in item simulation using this grid.

See Forecast Model Statistical Parameters.

Defining Simulation Model Parameters

Access the Model Parameters page.

Model statistical parameters are values that you enter or that the system creates during simulation. Define the parameters when you select Reoptimize Current to simulate a forecast. Make sure that you have processed the forecast, such as midperiod or period-end processing, so the system creates initial values for the page.

If you are simulating an item's forecast for the first time, and have not processed the forecast, the parameter fields are empty. You can enter values for simulation or select to let the system generate the values when it calculates values for the simulation. Different parameters exist for different models.

The system uses the values that you enter as input for processing models that use, for example, the gamma trend in producing simulation results. The value that you enter overrides existing values. When the system generates the values for you, it computes the value based on forecast settings for the control group and item.

Note. Although you can tune the item's forecast by entering different values, you should have a good understanding of forecasting statistics before making model parameter changes.

Select the Reoptimize All Parameters check box to reset the current parameters for the model. For example, suppose that the current value for Alpha (Smoothing) is 0.650, the system removes the value and recalculates the value when you simulate the forecast.

Note. If you click the Calculate button without completing required parameters for models, the system displays a warning message. When you click OK, the system displays the page with the missing parameter. Complete that field, select the Simulations Option tab, and click the Calculate button again.

See Also

Forecast Model Statistical Parameters

Forecast Model Accuracy Measurements

Viewing Simulation Results

After you calculate the simulation, you can review the changes and then apply them to the forecast item. Use simulation results to review item information period-by-period and to compare the statistical forecast, adjusted demand, and seasonality profile. You can also make decisions about which accuracy measurements are better for use with the forecast.

This section discusses how to:

Review accuracy measurements for forecast models.
Use charts to simulate item forecasts.
Compare simulated and current item forecasts.
Review simulation statistics.
Change historical or evaluated forecasts for simulation.

Common Elements Used in This Section

Accuracy	Select to view simulation results for the forecast models that you selected or were selected by the system. The results include rows of accuracy statistics for each model included in the simulation. See Forecast Model Accuracy Measurements.
Apply	Click to apply the simulation results to the item record. This permanently updates item forecast information.
Chart	Select to view the simulation results in a graphical format.
Model to Use	Initially displays the best-ranked model. Each model that you select for simulation is available for selection.
Options	Select to define the historical and evaluated forecast values that you want to update when you apply the results of the simulation to the item's record.
Refresh	Click to update the statistics for the forecast model in the Model to Use field. This option is available when you selected multiple models on which to simulate the forecast. To refresh the page, select a model and click the Refresh button.
Return	Click to access the Simulation Options page, where you can change options and model parameters.
Statistics	Select to review the simulated values for each accuracy measurement statistic method for the current forecast model.
Tables	Select to compare demand and forecast data for the current forecast model with that of the simulated model.

Pages Used to View Simulation Results

Page Name	Object Name	Navigation	Usage
Accuracy Measures	DP_ FCASTSIMRES3	Demand Planning, Process Forecast, Forecast Items, Simulation, Simulation Options Click the Calculate button on the Simulation Options page.	Review accuracy measures for all of the simulated forecast models that you selected on the Simulation Options page.
Chart	DP_ FCASTSIMRES2	Click the Chart link.	Use charts to simulate item forecasts and compare the current forecast model with simulated model that you select.
Tables	DP_ FCASTSIMRES1	Click the Tables link.	Compare simulated and current item forecasts. If you selected multiple models on the Simulation Options page, you can change the model to use in the comparison.
Statistics	DP_ FCASTSIMRES5	Click the Statistics link.	Review simulation statistics. Displays statistical measurements for a forecast model, including descriptive, model parameters and accuracy measures.
Options	DP_ FCASTSIMRES4	Click the Options link.	Change historical or evaluated forecast values. The change is temporary. The system does not update the settings when you apply simulation results.

Reviewing Accuracy Measurements for Forecast Models

Access the Accuracy Measures page.

This page displays accuracy statistics for each of the options that you defined for the simulation using the Simulation Options page. The current forecast model appears in the second column, and the system identifies it with two asterisks (**). The statistical accuracy method is defined in the Best Fit Accuracy Statistic field at the control-group level. The system identifies the value with an asterisk (*), and it is always the first row in the grid.

You can:

View simulation results based on the simulation options defined on the Simulation Options page.
Change the model and click the Refresh button to view the affects of using a different model.

You can view the affects of changing models if you click the links.
Update the current model when you click the Apply button.

This updates the model that appears in the Model in Use field and it then becomes the current model for the view.

The statistic accuracy methods appear in the Description column. Values appear for accuracy measurements when data is available for calculations and when the models use a specific accuracy measurement.

See Forecast Model Accuracy Measurements.

Using Charts to Simulate Item Forecasts

Access the Chart page.

This page shows a graphical representation of up to five different period-based fields for an item. You can compare adjusted demand and forecast and simulated and current statistical forecast along with numerous forecast models.

Use the Periods to Display and Periods to Scroll fields to define the amount of data that appears in the chart at one time.

The seasonality graph shows a base line of 100, the current seasonality values, and the simulated seasonality values for the current forecast model.

Note. Seasonality colors use the same color scheme as the main chart. Blue represents the current profile and yellow represents the simulated profile.

Comparing Simulated and Current Item Forecasts

Access the Tables page.

This page displays period-based data for the current values of actual and current adjusted demand, current adjusted forecast, statistical forecast and seasonality, as well as the simulated values for adjusted demand, statistical forecast, and seasonality factors. Twelve periods of detail appear at once.

Reviewing Simulation Statistics

Access the Statistics page.

This page displays statistical results created during the item forecast simulation. Using the results, you can determine which model and model parameters values are best suited for the organization's forecasting.

Use these group boxes to display the current model's statistics:

Descriptive Statistics

A group of measurements made up of tracking signals. Tracking signals is a forecasting tool that detects biases in the forecast and provides an early warning of an unstable forecast. Use the signals to evaluate structural errors that exist within a forecast by identifying situations where a forecasting model is over or under forecasting.

The tracking value is always a number in a range from 0 to 1. The closer it is to 1, the more likely it is that the forecast is biased. When the tracking value exceeds the bias test, which is stored on the control group, the system activates the tracking signal for the current period.

The system uses six tracking signals, starting with the current forecast period and progressing back five additional periods.

Model Specific Parameters

A group of measurements used by specific forecast models. You can normally set these parameters using the Model Parameters page when you set up the simulation.

Accuracy Measures

A group of measurements that are available for all of the models except Census X11 Additive.

See Also

Forecast Model Statistical Parameters

Changing Historical or Evaluated Forecasts for Simulation

Access the Options page.

Use this page to change the options temporarily for resetting history and evaluation modes. The system overwrites all of the historical and evaluated forecast data for the simulation according to the settings you define. Select a field to replace the historical or evaluated forecast with a statistical forecast, which is recalculated based on the new model.

The History RESET Mode and Evaluated RESET Mode group box contain fields maintained at the control-group level. The fields control which historical and evaluated data series the system replaces during the simulation. The history fields control which historical forecast data series is reset or overwritten when you apply the simulation.

Evaluated forecasts are data series that you can consider frozen in time. This provides a technique so that as the you develop and possibly adjust the forecast, its values do not change after they are rolled into the periods defined as evaluation periods on the forecast view. Although you can change the adjusted forecast as the forecast period moves closer to the current period, the evaluated forecast does not change after it is within the evaluation period range.

You can compare how well you are forecasting for periods outside of the evaluation period range to those that are in the adjusted forecast in the nearer term. This is important, because the item replenishment decisions are sometimes based on this evaluated forecast and its accuracy.

Note. Changes that you make to the historical or evaluated statistical, adjusted, or prorated forecasts are in effect during the simulation only.

See Also

Controlling Forecast Item Reset Data