MD 021 - Management and Operations Forecasting Outline Components of demand Judgment methods Linear regression Time series methods Forecast errors 1 Judgment Methods Sales force estimates Executive opinion Market research Delphi method 2 Linear Regression Yi a bXi where: Y = dependent variable X = independent variable a = Y-intercept of the line b = slope of the line 3 Measures of Forecast Accuracy in Linear Regression Coefficient of correlation Coefficient of determination Standard error of the estimate 4 Regression Analysis Example The manager of Al’s Diner is interested in forecasting the number of potato skin appetizers sold each week. He believes that the number sold has a linear relationship to the price and uses linear regression to determine if this is the case. X Y Week (Price) (Appetizers) 1. $2.70 760 2. 3.50 510 3. 2.00 980 4. 4.20 250 5. 3.10 320 6. 4.05 480 The Excel output is below: Regression Statistics Multiple R 0.843 R Square 0.711 Adjusted R Square 0.639 Standard Error 165.257 Observations 6 ANOVA Regression Residual Total Intercept Price ($) df SS 1 4 5 269160 109239 378400 Coefficients Standard Error 1454.604 295.939 -277.628 88.434 MS F Significance F 269160 9.856 0.035 27309 t Stat Pvalue 4.915 0.008 -3.139 0.035 5 Linear Regression Example A professor is interested in determining whether average study hours per week is a good predictor of test scores. The results of her study are: Hours Score 3.0 90 2.1 95 5.8 65 3.8 80 4.2 95 3.2 60 5.3 85 4.6 70 A student says: "Professor, what can I do to get a B on the next test. The professor asks, "On average, how many hours do you spend studying for this course per week?" The student responds, "About 2 hours." Use linear regression to forecast the student's test score. Regression Statistics Multiple R 0.391 R Square 0.153 Adjusted R Square 0.0121 Standard Error 13.544 Observations 8 ANOVA Regression Residual Total Intercept Study hours df SS 1 6 7 199.246 1100.753 1300 Coefficients Standard Error 97.325 17.301 -4.331 4.156 MS F Significance F 199.246 1.0861 0.3375 183.458 t Stat 5.625 -1.042 P-value 0.0013 0.3375 6 Time Series Methods Naive forecasts Moving averages Weighted moving averages Exponential smoothing Trend-adjusted exponential smoothing Regression Method Multiplicative seasonal method 7 Moving Average Method n Ft MAn At i i 1 n Month 1 2 3 4 Customer arrivals 800 740 810 790 Use a 3-month moving average to forecast customer arrivals for month 5. F5 If the actual demand for month 5 is 805 customers, what is the forecast for month 6? F6 8 Weighted Moving Average Method Ft wn At n wn 1 At ( n 1) ... w1 At 1 Month 1 2 3 4 Customer arrivals 800 740 810 790 Let W1 0.50, W2 0.30, and W3 0.20. Calculate the forecast for month 5. F5 If the actual demand for month 5 is 805 customers, what is the forecast for month 6? F6 9 Exponential Smoothing Ft (1 ) Ft 1 At 1 Month 1 2 3 4 Customer arrivals 800 740 810 790 Suppose F3 783 customers and 0.20. What is the forecast for month 5? F4 F5 If D5 805, what is the forecast for month 6? F6 10 Trend-Adjusted Exponential Smoothing St TAFt ( At TAFt ) Tt Tt 1 (TAFt TAFt 1 Tt 1 ) TAFt 1 St Tt Month 1 2 3 4 5 Patients 48 52 50 54 55 Using months 1-4, an initial estimate of the trend is 2 [(4-2+4)/3 = 2]. The starting forecast for month 5 is 54+2 = 56. Using 0.3 and 0.4 , forecast the number of patients in month 6. S5 T5 TAF6 If the actual number of patients in month 6 is 58, what is the forecast for month 7? S6 T6 TAF7 11 Regression Method Example: Garcia Garage Month (t) Number of Number of Oil time periods Changes (Y) from t = 0 Jan. 1 41 Feb. 2 46 Mar. 3 57 Apr. 4 52 May 5 59 Jun. 6 51 Jul. 7 60 Aug. 8 62 1. Forecast the numbers of oil changes in September, October, and November. 2. What is the average value of the trend? Regression Statistics Multiple R 0.817 R Square 0.668 Adjusted R Square 0.613 Standard Error 4.572 Observations 8.000 ANOVA Regression Residual Total Intercept X Variable 1 df 1.000 6.000 7.000 Coefficients 42.464 2.452 SS MS F 252.595 252.595 12.085 125.405 20.901 378.000 Standard Error 3.562 0.705 t Stat 11.921 3.476 Pvalue 0.000 0.013 Significance F 0.013 Lower 95% 33.748 0.726 Upper 95% 51.181 4.179 12 Multiplicative Seasonal Method Step 1: Step 2: Step 3: Step 4: Step 5: Step 4: Calculate the trend line based on the available data using regression. Calculate the centered moving average, with the number of periods equal to the number of seasons. Calculate the seasonal relative for a period by dividing the actual demand for the period by the corresponding centered moving average. Calculate the overall estimated seasonal relative by averaging the seasonal relatives from the same periods over the cycle. Calculate the trend values for each of the periods to be forecast based on the trend line determined in Step 1. To get a forecast for a given period in a future cycle, multiply the seasonal factor by the trend values. 13 Multiplicative Seasonal Method Application Quarter Demand CMA (4 seasons) 100 1 2 MA (2 periods) Seasonal Relatives Normalized S.R. 261.5 1.147227533 1.171002862 274 0.729927007 0.745054133 285.5 0.672504378 0.686441468 298 1.369127517 1.397501537 3.918786436 4 400 250 3 300 273 4 200 275 5 192 296 6 408 300 7 384 8 216 9 10 11 12 331 344 356 369 Total (trend value*) (trend value*) (trend value*) (trend value*) 227 480 417 275 (forecast) (forecast) (forecast) (forecast) * Using regression, the trend line is 218.86 + 12.48t. 14 Forecast Errors Bias--- systematic errors Random Errors --- variability Example: Day 1 100 Actual Demand Forecast 105 1 Forecast 50 2 Day 2 100 Day 3 100 Day 4 100 105 105 105 150 50 150 15 Forecast Error Measures Bias: n et t 1 n Average error Variability: n et 2 Mean squared error MSE t 1 n 1 Standard deviation s MSE n et Mean absolute error t 1 n MAD n [ Mean percent absolute error MAPE t 1 et (100)] At n 16 Control Chart for Forecast Errors Upper Control Limit: UCL 0 z MSE Lower Control Limit: LCL 0 z MSE Z = the number of standard deviations from the mean *Where to find “z” given the percentage of the control chart, P0 ? Where to find “z” given the probability for type I error, ? Normal Distribution Table (page 850, Table B.2) Look for “z” corresponds to the probability: p P{Z<= z} = 1 2 = 0.5+ 2 , 0 P0 =1- e.g. A 95% control chart has = 1-95% = 5%, which means its probability for type I error is 5%. Thus probability in the table should be 0.975 (P = 1-0.025 or P = 0.5+ 0.475), which corresponds to z = 1.96. 17 Summarizing Forecast Accuracy Period 1 2 3 4 5 6 7 8 9 Actual (A) 113 85 96 86 121 100 142 92 72 Forecast (F) Error (E=A-F) Abs Error 95 18 18 80 5 5 103 -7 7 119 -33 33 117 4 4 125 -25 25 67 75 75 96 -4 4 116 -44 44 Total MAD = MSE = s= MAPE = -11 215 Error Sq 324 25 49 1089 16 625 5625 16 1936 [(Abs E)/A] x 100 15.93 5.88 7.29 38.37 3.31 25.00 52.82 4.35 61.11 9705 214.06 23.9 1213.1 34.8 23.8% 18 Tracking and Analyzing Forecast Errors Period 10 11 Actual (A) 102 107 Forecast (F) 130 102 Error (E=A-F) -28 5 12 13 14 15 112 118 89 142 89 97 115 82 23 21 -26 60 16 17 18 100 94 111 130 137 89 -30 -43 22 Average error (periods 1-18)= Standard deviation (periods 1-9) = -0.39 34.8 2s control limits: 0 +/- 2(34.8) = 0 +/- 69.6 Total 4 80 UCL = 69.6 60 40 20 0 -20 10 11 12 13 14 15 16 17 18 -40 -60 LCL = -69.6 -80 19 Forecasting Summary Notes Choosing a Forecasting Method General considerations: Method Pros Judgment Can be used in the absence of historical data (e.g. new product) Helpful in identifying turning points and preparing medium- and long-term forecasts Causal Most sophisticated method Very good for predicting turning points and preparing medium- and long-term forecasts Time series Easy to implement Work well when demand relatively stable is Cons Subjective estimates are subject to the biases and motives of the estimators Must have historical data on independent and dependent variables Relationships can be difficult to specify Rely exclusively on past demand data Only useful for short-term estimates Specific considerations for time series methods: Method Pros Naive forecast Easiest method, low cost Works well when random errors are small Simple moving average Easiest moving average method To some extent, controls for random error Weighted moving average Weights Cons Results in highly variable forecasts if the random errors are large Data must be retained for n periods Forecast lags changes in the underlying average of demand Data must be retained for n periods Forecast lags changes in the underlying average of demand Forecast lags changes in the underlying average of demand can be varied to be responsive to demand pattern To some extent, controls for random error Exponential smoothing Requires little data can be varied to be responsive to demand pattern To some extent, controls for random error In general, emphasize recent demand (i.e. small n, large weights for recent observations, large ) for dynamic (i.e. uncertain) demand patterns. Emphasize historical experience for stable demand patterns. If a trend is present, simple moving average, weighted moving average, and exponential smoothing estimates will always lag actual demand. 20 Forecasting Notes Choosing a Time Series Forecasting Method Evaluating forecast performance: Forecast errors can be classified as either bias errors or random errors. Bias errors are the results of systematic over- or underestimation. Random errors are unpredictable. Ideally, a forecast should minimize both bias and random errors. Method Purpose Mean forecast Measures bias errors Mean squared error (MSE) Measures the dispersion of forecast errors; large errors get more weight than when using MAD Mean absolute deviation Measures the dispersion of forecast errors; method is (MAD) intuitive Mean absolute percent Measures the dispersion of forecast errors relative to the error (MAPE) level of demand Forecast error control chart Determines whether the method of forecasting is accurately predicting actual changes in demand 21