Chapter 11-Forecasting Models.ppt

《Chapter 11-Forecasting Models.ppt》由会员分享，可在线阅读，更多相关《Chapter 11-Forecasting Models.ppt（53页珍藏版）》请在麦多课文档分享上搜索。

1、Chapter 11: Forecasting Models, 2007 Pearson Education,Forecasting,Forecasting is attempting to predict the futureDecision makers want to reduce uncertainty by predicting future values such as sales or investment return,Steps in Forecasting,Determine the objective of the forecast Identify items to b

2、e forecast Determine time horizon Select the forecasting model(s) Gather data Validate model Make forecast and implement results,Types of Forecasts,Qualitative - subjective methods based on intuition and experience Time Series based on historical data and assume the past indicates the future Causal

3、Models data based where there may be a cause and effect relation between variables,Qualitative Forecasting Models,Delphi Method an iterative group process where a group of experts attempt to reach consensusJury of Executive Opinion uses opinions of high level managers often combined with statistical

4、 models,Qualitative Forecasting Models,Sales Force Composite each salesperson estimates sale in his/her own region and forecast are combined for an overall forecastConsumer Market Survey future purchase plans are solicited from customers,Measuring Forecast Error,Measures how accurate the forecast wa

5、sFor time period t: Forecast error = Actual value Forecast value= At - Ft,Methods of Measuring Overall Forecast Error,Mean Absolute Deviation (MAD) MAD = |At Ft| / Twhere T = the number of time periodsMean Squared Error (MSE) MSE = (At Ft)2 / T,Methods of Measuring Overall Forecast Error,Mean Absolu

6、te Percent Error (MAPE) Measure error as a percent of actual values MAPE = 100 |At Ft| / At / T,Time Series,A time series is where the same value is recorded at regular time intervals Examples: daily stock price, monthly sales, annual revenue, etc.,Components of a Time Series,Trend long term upward

7、or downward movement Seasonality the pattern that occurs every year Cycles the pattern that occurs over a period of years Random variations caused by chance and unusual events,Time Series Components,Time Series Decomposition,A time series can be broken down into its individual components Two approac

8、hes: Multiplicative decomposition Forecast = Trend x Seasonality x Cycles x RandomAdditive decomposition Forecast = Trend + Seasonality + Cycles + Random,Stationary and Nonstationary Time Series Data,If a time series has an upward or downward trend, it is nonstationaryIf it has no trend, it is stati

9、onary,Moving Averages,Smooth out variations in a time series when values are fairly steady Some number (k) of consecutive periods are averagedk-period moving average = (actual values in previous k periods) k,Wallace Garden Supply Example,Weighted Moving Averages,A moving average where some periods a

10、re weighted more heavily than othersK-period weighted moving average = (wi Ai) / (wi)where, wi = weight for period iAi = actual value for period i,Wallace Garden Supply With Weighted Moving Averages,Period Weightslast month 32 month ago 23 months ago 1,3-Month Weighted Moving Average,Using Solver to

11、 Find the Optimal Weights,The weights are the decision variables (changing cells) Minimize some measure of forecast error (MAD, MSE, or MAPE) as the Target cell Note this is a nonlinear objective Weights must be nonnegative Go to file 11-3.xls,Exponential Smoothing,Another smoothing method Does not

12、require extensive past data Ft+1 = Ft + x (At Ft)Ft+1 = forecast for period (t+1) Ft = forecast for period t = a weight (smoothing constant) At = actual value for period t,Wallace Garden Supply With Exponential Smoothing,Assume the smoothed value for the first month is the actual value Use = 0.1 and

13、 also = 0.9,Trend Analysis,Fits a straight or curved line through a time series We will cover only linear trends A scatter diagram shows the trend Excel can both create the scatter diagram and fit the linear trend line,Midwestern Electric Co. Example,Go to file 11-5.xls,The Trend Equation, = b0 + b1

14、X where, = forecast average dependent valueX = independent value (time)b0 = Y-interceptb1 = slope of the line,Least Squares Method,The b0 and b1 values are found using the least squares method, which seeks to minimize the sum of squared errorsSSE = (Y )2 Where,Error = Y - ,Least Squares Method for B

15、est-Fitting Line,Least Squares Line With Excel,Can use regression in the Analysis ToolPak add-in The time (X) values are transformed to 1, 2, 3, etc.Go to file 11-6.xls,Seasonality Analysis,When a seasonal pattern repeats yearly, this can be used for future forecasts Need monthly or quarterly data A

16、 seasonal index is the ratio of the average value in that season, over the annual average,Eichler Supplies Seasonality Example,Have monthly demand data for 24 months Calculate overall average monthly demand Calculate ratio for each monthGo to file 11-7.xls,Decomposition of a Time Series,Decompositio

17、n breaks a time series down into its components (Trend, Seasonal, Cyclical, and Random) Two types of models Multiplicative Additive,Multiplicative Decomposition Sawyer Piano House Example,Want to forecast sales of grand pianos Have quarterly data for the past 5 years Steps: Find the seasonal indices

18、 First smooth data with moving averages Seasonal ratio = actual value / smoothed value Average the seasonal ratios for each quarter Unseasonalized value = actual value / seasonal index,Steps Continued,Find the trend equation using the unseasonalized values Calculate forecasts Use the trend equation

19、to make an unseasonalized forecast Multiply the unseasonalized forecast by the seasonal index Calculate forecast errorGo to file 11-8.xls,Causal Forecasting Models,Forecasting a dependent variable based on other (independent) variables Uses simple or multiple regression Example: Dependent variable:

20、Swimwear sales Independent variables: selling price, competitors prices, temperature, whether schools are in session, advertising,Causal Simple Regression Model,Want to predict selling price of homes (Y) based on the square footage (X) Have data on 12 homes recently sold in a specific neighborhood U

21、se scatter diagram to check for linear relation Find least squares equation = b0 + b1X,Causal Simple Regression With ExcelModules,Given the X and Y data, it will automatically: Calculate the regression equation Calculate forecast error Produce a scatter plot with the regression line Go to file 11-9.

22、xls,The Regression Equation,Forecast average selling price =-8.125 + 97.789(Home size)Slope interpretation: On average the price of a home will increase by $97.789 thousand per additional thousand sq. ft. Intercept interpretation: When X=0 the average selling price is -$8.125 thousand (has no practi

23、cal meaning since no houses have 0 square feet),Standard Error and Correlation,Standard Error (Sy,x) the standard deviation of the regression equation (useful for confidence intervals on forecasts) Correlation Coefficient (r) measures the strength of the linear relation -1 r 1,Correlation Coefficien

24、t Examples,Coefficient of Determination (R2),Measures the proportion of variation in the dependent variable (Y) that can be explained with the independent variable (X) 0 R2 1 It is the correlation squared,Using the Causal Simple Regression Model,To forecast the average selling price of a 3100 sq. fo

25、ot home, use X = 3.10 Can use ExcelModules to produce forecast Forecast = -8.125 + 97.789(3.1) = 295which is $295,000,Potential Weaknesses of Causal Forecasting With Regression,We need to provide the value(s) of the independent variable(s) Individual values of Y may be much higher or lower than the

26、forecast average Model is generally valid only for X values within the range of the data set,Approximate Confidence Interval,Helpful for showing how high or low an individual value might be Approximate confidence interval formula: + Z/2 (Sy,x) Approximate 95% interval example: 295 + 1.96 (42.602) Wh

27、ich is $211,500 to $378,500,Causal Simple Regression Using Excels Analysis ToolPak,An add-in that includes regression Appears as “Data Analysis” at the bottom of the “Tools” menu Go to file 11-9.xls,Statistical Significance Tests,If the true value of the slope (1) does not differ significantly from

28、0, then Y does not change as X changes Hypotheses:H0: 1=0 (X is not significantly related to Y)H1: 10 (X is significantly related to Y) Tested by both F-test and t-test Reject H0 if p-value ,Hypothesis Test Results for Home Selling Prices,From either F-test or t-test (they are equivalent for simple

29、regression):Reject H0 because p=0.011 is 0.05 (alpha)Home size is statistically significant in having ability to predict home selling price,Causal Multiple Regression Model,More than one independent variable = b0 + b1X1 + b2X2 + + bpXpWhere,b0 = Y-axis intercept (all Xs =0)bi = slope for Xip = numbe

30、r of independent variables (Xs),Causal Multiple Regression Using Excels Analysis ToolPak,All columns of independent variables must be adjacent to one another (no gaps) The Analysis ToolPak add-in (Data Analysis) must be “turned on”Go to file 11-10.xls,Statistical Significance Test Of the Overall Mod

31、el (F-test),If all true slope values (i) equal 0, then the model has no ability to predict Y Hypotheses:H0: 1=2=0 (model has no ability to predict Y)H1: at least one i 0 (at least one variable has ability to predict Y) F-test is used,Statistical Significance Test of Individual Variables (t-test),Tes

32、ts whether an individual X is helping to predict Y (in the presence of the other Xs) Hypotheses for each Xi:H0: i=0 (Xi adds no ability to predict Y, given the other Xs in the model)H1: i 0 (Xi adds ability to predict Y, given the other Xs in the model),Hypothesis Test Results for Home Selling Price

33、s,F-test: Overall model has significant ability to predict home prices T-tests:Land area is significant, given the presence of home sizeHome size is not significant, given the presence of land area,Multicollinearity,Why did home size become nonsignificant when land area was added? Multicollinearity exists when 2 or more independent variables are highly correlated Correlations among Xs can be used to detect multicollinearity,