Transcript Document
IE5403 Facilities Design and Planning Instructor: Assistant Prof. Dr. Rıfat Gürcan Özdemir http://web.iku.edu.tr/~rgozdemir/IE551/index(IE551).htm Course topics Chapter 1: Forecasting methods Chapter 2: Capacity planning Chapter 3: Facility location Chapter 4: Plant layout Chapter 5: Material handling and storage systems Grading Participation Quizzes Assignment Midterm 1 Final 5% 15% (4 quizzes) 15% (every week) 30% (chapters 1 and 2) 35% (all chapters) 3 IE5403 - Chapter 1 Forecasting methods 4 Forecasting Forecasting is the process of analyzing the past data of a time – dependent variable & predicting its future values by the help of a qualitative or quantitative method 5 Why is forecasting important? Better use of capacity Proper forecasting Reduced inventory costs Lower overall personel costs Increased customer satisfaction Decreased profitability Poor forecasting Collapse of the firm 6 Planning horizon demand actual demand? Forcast demand past demand actual demand? now time planning horizon 7 Designing a forcasting system Forecast need NO YES Collect data? NO NO Data avilable? YES Analyze data Quantitative? YES Causal factors? NO YES Qualitative approach Causal approach Time series 8 Regression Methods Unknown parameters Model xt xt a b t t dependent variable Random error component independent variable xt = a + b t Simple linear model t 9 Estimating a and b parameters e5 xt xt = a + b t e3 e4 e1 Such that sum squares of the errors (SSE) is minimized e2 0 T=5 t forecast error et = ( xt – xt ) 10 Least squares normal equations T T t 1 t 1 SSE ( xt xˆt ) 2 ( xt aˆ bˆt ) 2 T SSE 2 ( xt aˆ bˆt ) 0 aˆ t 1 T SSE 2 ( xt aˆ bˆt )t 0 bˆ t 1 T T T t 1 T t 1 T t 1 t 1 t 1 aˆT bˆ t xt aˆ t bˆ t 2 t xt Least squares normal equations 11 Coefficient of determination (r2) unexplained deviation = (xt – xt)2 xt xt xt (xt (xt – xt)2 = explained – xt)2 = total deviation deviation 2 ˆ ( x x ) explained t t 2 r 2 total ( x x ) t t t , 0 r2 1 12 Coefficient of corelation (r) r = coeff. of determination = r2 r shows the strength of relationship between xt and t Sign of r ,(– / +), shows the direction, ( of the relationship between xt and t r r 2 or ) T txt t xt [T t t ][T x xt ] 2 2 2 t 2 –1r1 13 Example – 3.1 It is assumed that the monthly furniture sales in a city is directly proportional to the establishment of new housing in that month a) Determine regression parameters, a and b b) Determine and interpret r and r2 c) Estimate the furniture sales, when expected establishment of new housing is 250 14 Example – 3.1(continued) New Furniture Month Housing Sales /month /month ($1000) Jan. 100 461 Feb. 110 473 March April May 96 114 120 450 472 481 June 160 538 New Furniture Month Housing Sales /month /month ($1000) July 150 540 Aug. 124 517 Sep. 93 449 Oct. 88 452 Nov. 104 454 Dec. 116 495 15 Example – 3.1(solution to a) t2 t xt txt 100 461 10,000 46,100 110 473 12,100 52,030 96 450 9216 43,200 114 472 12,996 53,808 120 481 14,400 57,720 160 538 25,600 86,080 150 540 22,500 81,000 124 517 15,376 64,108 93 449 8649 41,757 88 452 7744 39,776 104 454 10,816 47,216 116 495 13,456 57,420 T = 12 t = 1375 xt = 5782 t2 = 162,853 txt = 670,215 a T + b t = xt a t + b t2 = txt 16 Example – 3.1(solution to a) 1375 x 12 a + 1375 b = 5782 – 12 x a 1375 + b 162,853 = 670,215 b (1375 2 – 12 x 162,853) = (1375 x 5782 – 12 x 670,215) (1375 x 5782 – 12 x 670,215) b = (1375 2 – 12 x 162,853) = 1.45 17 Example – 3.1(solution to a) b = 1.45 12 a + 1375 b = 5782 a 1375 + b 162,853 = 670,215 a= (5782 – 1375 x 1.45) 12 = 315.5 xt = a + b t xt = 315.5 + 1.45 t 18 Example – 3.1(solution to b) t xt xt 100 461 461 110 473 475 96 450 455 114 472 481 120 481 490 160 538 548 150 540 533 124 517 495 93 449 450 88 452 443 104 454 466 116 495 484 xt = 315.5 + 1.45 (100) = 461 xt = xt T = 5782 12 = 482 19 Example – 3.1(solution to b) xt t xt xt xt xt xt xt 100 461 461 -21 -21 110 473 475 -7 -9 96 450 455 -27 -32 114 472 481 -1 -10 120 481 490 8 -1 160 538 548 66 56 150 540 533 51 58 124 517 495 13 35 93 449 450 -32 -33 88 452 443 -39 -30 104 454 466 -16 -28 116 495 484 2 13 xt x t t Explained deviation xt xt x t t Total deviation 20 Example – 3.1(solution to b) t xt xt xt xt xt xt ( xt xt )2 ( xt xt )2 100 461 461 -21 -21 441 441 110 473 475 -7 -9 49 81 96 450 455 -27 -32 729 1.024 114 472 481 -1 -10 1 100 120 481 490 8 -1 64 1 160 538 548 66 56 4.356 3.136 150 540 533 51 58 2.601 3.364 124 517 495 13 35 169 1.225 93 449 450 -32 -33 1.024 1.089 88 452 443 -39 -30 1.521 900 104 454 466 -16 -28 256 784 116 495 484 2 13 4 169 21 Example – 3.1(solution to b) Coefficient of determination: ( xt r2 = ( xt xt )2 11.215 = xt )2 12.314 = 0.91 91% of the deviation in the furniture sales can be explained by the establishment of new housing in the city 22 Example – 3.1(solution to b) Coefficient of corelation: r = r2 = 0.91 = 0.95 a very strong (+) relationship (highly corelated) 23 Example – 3.1(solution to b) r r T txt t xt 2 [T t t 2 2 2 ][T xt xt ] 2 xt2 = 2.798.274 12 x 670,215 – 1375 x 5782 r= = 0.95 [12 x 162,853 – (1375)2 ][12 x 2,798,274 – (5782)2] r2 = (0.95)2 = 0.91 24 Example – 3.1(solution to c) xt = a + b t xt = 315.5 + 1.45 t t = 250 xt = 315.5 + 1.45 (250) = 678 xt = $ 678,000 x $1000 25 Components of a time series 1. 2. 3. 4. Trend ( a continious long term directional movement, indicating growth or decline, in the data) Seasonal variation ( a decrease or increase in the data during certain time intervals, due to calendar or climatic changes. May contain yearly, monthly or weekly cycles) Cyclical variation (a temporary upturn or downturn that seems to follow no observable pattern. Usually results from changes in economic conditions such as inflation, stagnation) Random effects (occasional and unpredictable effects due to chance and unusual occurances. They are the residual after the trend, seasonali 26 and cyclical variations are removed) Components of a time series xt seasonal variation a2 trend slope a1 random effect 0 1 2 Year 1 3 4 5 6 7 t 8 Year 2 27 Simple Moving Average Model xt = a + t xt Constant process Forecast error xt = a a t 28 Simple Moving Average ˆ Forecast X T 1 is average of N previous observations or actuals Xt : 1 ˆ X T 1 ( X T XTT 1 X T 1 N ) N T 1 Xˆ t 1 Xt N t T 1 N Note that the N past observations are equally weighted. Issues with moving average forecasts: All N past observations treated equally; Observations older than N are not included at all; Requires that N past observations be retained. Simple Moving Average Include N most recent observations Weight equally Ignore older observations weight 1/N T+1-N ... T-2 T-1 T today Parameter N for Moving Average If the process is relatively stable choose a large N If the process is changing choose a small N 31 Example 3.2 Week 1 2 3 4 5 6 7 8 9 10 11 12 Demand 650 678 720 785 859 920 850 758 892 920 789 844 What are the 3-week and 6-week Moving Average Forecasts for demand of periods 11, 12 and 13? 32 Weighted Moving Average Include N most recent observations Weight decreases linearly when age of demand increases Weighted Moving Average T wx t t wt = weight value for xt t=T-N+1 WMT = The value of wt is higher T wt for more recent data t=T-N+1 34 Example 3.3 Month Jan. Feb. March April May Sales 10 12 13 ? ? a) Use 3-month weighted moving average with the following weight values to predict the demand of april wT = 3 b) Assume demand of april is realized as 16, what is the demand of may? wT-1 = 2 wT-2 = 1 35 Exponential Smoothing Method A moving average technique which places weights on past observations exponentially Realized demand at period T ST = a Smoothed value xT + 1a ST-1 Smoothing constant 36 Exponential Smoothing Include all past observations Weight recent observations much more heavily than very old observations: weight Decreasing weight given to older observations today Exponential Smoothing Include all past observations Weight recent observations much more heavily than very old observations: 0 a 1 a weight Decreasing weight given to older observations today Exponential Smoothing Include all past observations Weight recent observations much more heavily than very old observations: weight 0 a 1 a a (1 a ) Decreasing weight given to older observations today Exponential Smoothing Include all past observations Weight recent observations much more heavily than very old observations: weight 0 a 1 a a (1 a ) Decreasing weight given to older observations a (1 a ) 2 today Exponential Smoothing Include all past observations Weight recent observations much more heavily than very old observations: 0 a 1 weight a a (1 a ) Decreasing weight given to older observations a (1 a ) 2 a (1 a ) today 3 Exponential Smoothing 2 ˆ X T 1 aX t a (1 a ) X t 1 a (1 a ) X t 2 Xˆ T 1 aX t (1 a )aX t 1 a (1 a ) X t 2 Exponential Smoothing 2 ˆ X T 1 a X t a (1 a ) X t 1 a (1 a ) X t 2 Xˆ T 1 a X t (1 a )aX t 1 a (1 a ) X t 2 Xˆ T 1 aX T (1 a ) Xˆ T ST aX T (1 a) ST 1 The meaning of smoothing equation ST = a xT ST = a + 1a ST-1 xT + ST-1 a ST-1 ST = ST-1 + a ( ST = xT+t New forecast for future periods ST-1 = xT ST-1 ) xT Old forecast for the most recent period eT = xT – xT Forecast error 44 Exponential Smoothing Thus, new forecast is weighted sum of old forecast and actual demand Notes: 2 values ( X T and Xˆ T ) are required, compared with N for moving average Parameter a determined empirically (whatever works best) Rule of thumb: a < 0.5 Typically, a = 0.2 or a = 0.3 work well Only Choice of a Small a Slower response Large a Quicker response Equivelance between a and N a = 2 N+ 1 N = 2–a a 46 Example 3.4 Week 1 2 3 Demand 820 775 680 Given the weekly demand data, what are the exponential smoothing forecasts for periods 3 and 4 using a = 0.1 and a = 0.6 ? Assume that S1 = x1 = 820 47 Example 3.4 (solution for a = 0.1) t xt St xt 1 820 820 2 775 815.5 820 3 680 801.95 815.5 4 S1 = x1 = 820 x2 = 820 S2 = a x2 + 1a S1 S2 = 0.1(775) + 0.9(820) = 815.5 801.95 x3 = 815.5 48 Example 3.4 (solution for a = 0.6) t xt St xt 1 820 820 2 775 793.0 820 3 680 725.2 793.0 4 725.2 S1 = x1 = 820 x2 = 820 S2 = a x2 + 1a S1 S2 = 0.6(775) + 0.4(820) = 793.0 x3 = 793.0 49 Winters’ Method for Seasonal Variation Seasonal factor for period t Model xt = ( a + b t ) ct + t xt Constant parameter Trend parameter Random error component t 50 Initial values of a , b and c for one year available demand data L xt t=1 a0 = L b0 = 0 ct = xt a0 L ct t=1 = L YES ct values are valid NO ct values are normalized : ct = ct L L ct t=1 51 Smoothing equations xT ˆ ˆ aˆT a ( 1 a ) a b T 1 T 1 ˆ c T L 0 <a < 1 bˆT aˆT aˆT 1 (1 )[bˆT 1 ] 0<<1 xT cˆT g (1 g )cˆT L aˆT 0 <g < 1 52 Forecast Equation x(T+t) = ( aT + t bT ) c(T+t-L) = the smallest integer ≥ t L 53 Example 3.6 Month(2005) 1 Demand 4 2 2 3 4 5 6 7 8 9 10 11 12 5 8 11 13 18 15 9 6 5 4 a) Forecast the demand of Jan.’06 using Winters method with a = 0.2, = 0.1, g = 0.5 b) Forecast the demand of Feb.’06 when Jan.’06 realizes as 5 using Winters method with a = 0.2, = 0.1, g = 0.5 c) Forecast the demand of Mar.’06 and Mar.’07 when Feb.’06 realizes as 4 using Winters method with a = 0.2, = 0.1, g = 0.5 54 Example 3.6 (solution to a) L a0 = xt t=1 L = 100 12 = 8.3 b0 = 0 c1 = x1 a0 = 4 = 0.48 8.3 x(T+t) = ( a0 +1b0 ) c1 x(12+1)= ( 8.3 + 0 ) 0.48 = 4 t 1 2 3 4 5 6 7 8 9 10 11 12 ct 0.48 0.24 0.60 0.96 1.32 1.56 2.16 1.80 1.08 0.72 0.60 0.48 12 55 Example 3.6 (solution to b) JAN.’06 realizes as: x13 = 5 x13 5 ˆ = 0.2 ˆ aˆ13 a ( 1 a ) a b 0.8 8.3 + 0 = 8.72 + 0 0 0.48 cˆ1312 bˆT aˆT aˆT 1 (1 )[bˆT 1 ] = 0.1 8.72 – 8.3 + 0.9 0 = 0.043 xT cˆT g (1 g )cˆT L = 0.5 5 0.5 0.48 = 0.53 + 8.72 aˆT x(T+t) = ( a13 +1b13 ) c2 x(13+1)= ( 8.72 + 1 x 0.043 ) 0.24 = 2.1 56 Updated seasonal values should be normalized! 57 Example 3.6 (solution to c) FEB.’06 realizes as: x14 = 4 a14 = 0.2 4 + 0.8 8.72 + 0.043 = 10.34 0.24 b14 = 0.1 10.34 – 8.72 + 0.9 0.043 = 0.2 c14 = 0.5 4 + 0.5 0.24 10.34 = 0.31 x(14+1)= ( 10.34 + 1 x 0.2 ) 0.60 = 6.3 x(14+13)= ( 10.34 + 13 x 0.2 ) 0.60 = 7.76 Forecast of Mar.’06 Forecast of Mar.’07 58 Forecast accuracy Forecast accuracy shows the performance of the model for complying with the demand process, and is measured by using forecast error Forecast error is the difference between the actual demand and the forecast et = xt – xt Forecast error Actual demand Forecast 59 Error measures Looking at the error for an isolated period does not provide much useful information Rather we will look at errors over the history of the forecasting system. There are several methods for this process, although each has different meaning 1. Cumulative (sum) error, Et 2. Mean error, ME 3. Mean square error, MSE 4. Mean absolute deviation, MAD 5. Mean absolute percentage error, MAPE 60 Cumulative (sum) error, Et T ET = et t=1 Et should be close to zero if the forecast is behaving properly. That is, sometimes it overestimates and sometimes it underestimates, but in the long run these should cancel out 61 Mean error, ME 1 ME = n n et t=1 ME should be interpreted same as sum error, Et , that is, it shows whether the model is biased toward certain direction or not A forecast consistently larger than actual is called biased high A forecast consistently lower than actual is called biased low 62 Example 3.9 t xt 1 2 3 4 5 6 7 8 9 9 12 15 18 21 24 27 30 33 Validate the moving average (N=3) if it is suited to the given past data using ME and Et , and say if it is bias 63 Example 3.9 (solution) t xt MA(t-1) et 1 2 3 4 9 12 15 18 12 6 5 6 7 8 21 24 27 30 15 18 21 24 6 6 6 6 E9 = 6 + 6 +... + 6 = 36.00 9 33 27 6 ME = 1/n (et)= E9 / 6 = 6.00 BIASED LOW ! 64 Mean square error, MSE 1 MSE = n n et t=1 2 MSE is mainly used to counteract the inefficiency in error measuring as negative errors (– et) cancel out the positive error terms (+ et) By squaring the error terms, the “penalty” is increased for large errors. Thus a single large error greatly increases MSE 65 Mean absolute deviation, MAD 1 MAD = n n et t=1 MAD is another error measure for solving the neutralizing problem MAD measures the dispersion of the errors, and if MAD is small the forecast should be close to actual demand 66 Mean absolute percentage error, MAPE 1 MAPE = n n PEt t=1 xt – xt PEt = x (100) t MAPE is mainly used to counteract the inefficiency in error measuring as the previously defined mesaures depend on the magnitude of the numbers being forecast If the numbers are large the error tends to be large. It may more meaningful to look at error relative to the magnitude of the forecasts, which is done by MAPE 67 Example 3.11 t 1 2 3 4 5 6 7 8 9 10 11 12 xt 10 12 15 45 130 180 170 120 125 100 125 135 Compare a 3-period moving average model and a 6-period moving average model using given past data and show which suits better with respect to MSE, MAPE. 68 Example 3.11 (solution) t 1 2 3 4 5 6 7 8 9 10 11 12 xt MA[3]t-1 10 12 15 45 12.33 130 24.00 180 63.33 170 118.33 120 160.00 125 156.67 100 138.33 125 115.00 135 116.67 et 51.67 -40.00 -31.67 -38.33 10.00 18.33 1 12 e 2 MSE= t = 1196.30 6 t=7 69 Example 3.11 (solution) t xt MA[6]t-1 et 1 10 2 12 3 15 4 45 5 130 6 180 7 170 65.33 104.67 8 120 92.00 28.00 9 125 110.00 15.00 10 100 128.33 -28.33 11 125 137.50 -12.50 12 135 136.67 -1.67 1 12 e 2 MSE= t = 2154.32 6 t=7 70 Example 3.11 (solution) MA[3] model results t xt MA[3]t-1 7 170 118.33 8 120 160.00 9 125 156.67 10 100 138.33 11 125 115.00 12 135 116.67 et 51.67 -40.00 -31.67 -38.33 10.00 18.33 MA[6] model results Pet MA[6]t-1 30.39 33.33 25.33 38.33 8.00 13.58 65.33 92.00 110.00 128.33 137.50 136.67 et Pet 104.67 28.00 15.00 -28.33 -12.50 -1.67 61.57 23.33 12.00 28.33 10.00 1.23 1 12Pe 1 12Pe MAPE = 6 t = 24.83 MAPE = t = 22.74 6 t=7 t=7 71 Example 3.11 (solution) Error measures MSE MAPE Forecast models MA[3] 1196.30 24.83 MA[6] 2154.32 22.74 72