1、预测供应链需求预测供应链需求 CR (2004) Prentice Hall, Inc.Chapter 8I hope youll keep in mind that economic forecasting is far from a perfect science. If recent historys any guide, the experts have some explaining to do about what they told us had to happen but never did.Ronald Reagan, 19841产品计划三角形产品计划三角形Product i
2、n the Planning TriangleCR (2004) Prentice Hall, Inc.PLANNINGORGANIZINGCONTROLLINGTransport Strategy Transport fundamentals Transport decisionsCustomer service goals The product Logistics service Ord . proc. & info. sys.Inventory Strategy Forecasting Inventory decisions Purchasing and supply scheduli
3、ng decisions Storage fundamentals Storage decisionsLocation Strategy Location decisions The network planning process 计划计划 组织组织 控制控制Transport Strategy Transport fundamentals Transport decisionsCustomer service goalsThe product Logistics service Ord . proc. & info. sys.Inventory Strategy Forecasting I
4、nventory decisions Purchasing and supply scheduling decisions Storage fundamentals Storage decisionsLocation Strategy Location decisions The network planning process 库存战略 预测客户服务目标采购和供应时间决策存储基础知识存储决策产品物流服务订单管理和信息系统 库存决策 运输战略 运输基础知识 运输决策 选址战略 选址决策 网络规划流程2Forecasting in Inventory StrategyCR (2004) Pren
5、tice Hall, Inc.PLANNINGORGANIZINGCONTROLLINGTransport Strategy Transport fundamentals Transport decisionsCustomer service goals The product Logistics service Ord. proc. & info. sys.Inventory Strategy Forecasting Inventory decisions Purchasing and supply scheduling decisions Storage fundamentals Stor
6、age decisionsLocation Strategy Location decisions The network planning processPLANNINGORGANIZINGCONTROLLINGTransport Strategy Transport fundamentals Transport decisionsCustomer service goals The product Logistics service Ord. proc. & info. sys.Inventory Strategy Forecasting Inventory decisions Purch
7、asing and supply scheduling decisions Storage fundamentals Storage decisionsLocation Strategy Location decisions The network planning process3供应链预测什么供应链预测什么Demand, sales or requirements需求，销售或请求Purchase prices购买价格Replenishment and delivery times补给和交货时间CR (2004) Prentice Hall, Inc.48.1需求预测n1.需求的时间和空间特
8、征（Spatial versus Temporal Demand）n2.尖峰需求和规律性的需求（Lumpy versus Regular Demand）n3.派生需求和独立需求（Derived versus Independent Demand）5CR (2004) Prentice Hall, Inc.典型时间序列模式典型时间序列模式Typical Time Series Patterns:随机随机Random0501001502002500510152025TimeSalesActual salesAverage sales随机性或水平发展的需求，无趋势或季节性因素6CR (2004) P
9、rentice Hall, Inc.典型时间序列模式典型时间序列模式Typical Time Series Patterns:随机有趋势随机有趋势Random with Trend0501001502002500510152025TimeSalesActual salesAverage sales随机性需求，上升趋势，无季节性因素7CR (2004) Prentice Hall, Inc.Typical Time Series Patterns:Random with Trend & Seasonal0100200300400500600700800010203040Tim eSalesAct
10、ual salesTrend in salesSm oothed trend and seasonal sales随机性需求，有趋势，季节性因素8CR (2004) Prentice Hall, Inc.Typical Time Series Patterns:LumpyTimeSales尖峰需求模式9CR (2004) Prentice Hall, Inc.8.2预测方法预测方法1.定性方法Qualitative 调查法Surveys 专家系统Expert systems or rule-based2.历史映射法（时间序列分析Historical projection）移动平均Moving
11、average指数平滑Exponential smoothing3.因果或联想法Causal or associative回归分析Regression analysis4.协同Collaborative108.3 对物流管理者有用的方法对物流管理者有用的方法8.3.1.移动平均法移动平均法Moving AverageBasic formulatntiiAnMA11wherei = time periodt = current time periodn = length of moving average in periods Ai = demand in period iCR (2004) P
12、rentice Hall, Inc.11Example 3-Month Moving Average ForecastingMonth, iDemand formonth, iTotal demandduring past 3months3-monthmovingaverage.20120.21130360/312022110380/3126.6723140 360/312024110380/3126.672513026?CR (2004) Prentice Hall, Inc.12 加权移动平均加权移动平均Weighted Moving Averageperiod current in fo
13、recast period current in demand actual period next for forecast 0.30 to 0.01 usually constant smoothing where)1(formula smoothing exponential only, level basic, the to reduces which)1(.)1()1()1(then form, in exponential are )( weightsIf1.1133221112211ttttttntnttttniinnFAFFAFMAAAAAAMAwwwhereAwAwAwMAa
14、aaaaaaaaaaa13 I. Level only Ft+1= aAt + (1-a)Ft II. Level and trend St= aAt + (1-a)(St-1 + Tt-1) Tt= (St - St-1) + (1-)Tt-1 Ft+1= St + TtIII. Level, trend, and seasonality St= a(At/It-L) + (1-a)(St-1 + Tt-1) It= g(At/St) + (1-g)It-L Tt= (St - St-1) + (1-)Tt-1 Ft+1= (St + Tt)It-L+1where L is the time
15、 period of one full seasonal cycle. IV. Forecast errorMAD =|AtFNttN|1orS(AF )NFtt2t 1Nand SF 1.25MAD.8.3.2.指数平滑公式指数平滑公式Exponential Smoothing FormulasCR (2004) Prentice Hall, Inc.14CR (2004) Prentice Hall, Inc.Example Exponential Smoothing ForecastingTime series data1234Last year12007009001100This ye
16、ar14001000?QuarterGetting startedAssume a = 0.2. Average first 4 quarters of data and use for previous forecast, say Fo15CR (2004) Prentice Hall, Inc.Example (Contd)Begin forecasting9754/ )11009007001200(0FFirst quarter of 2nd year1000)975(8.0)1100(2.0)2.01 (2.0001FAFSecond quarter of 2nd year1080)1
17、000(8.0)1400(2.0)2.01 (2.0112FAF16CR (2004) Prentice Hall, Inc.Example (Contd)Third quarter of 2nd year1064)1080(8.0)1000(2.0)2.01 (2.0023FAFSummarizing1234Last year12007009001100This year14001000?Fore- cast100010801064Quarter17CR (2004) Prentice Hall, Inc.Example (Contd)Measuring forecast error as
18、MAD绝对差or RMSE (std. error of forecast) 标准差nFAMADnttt1|1)(12nFASntttF18CR (2004) Prentice Hall, Inc.Example (Contd)Using SF and assuming n=2408121080)(10001000)(140022FSNote To compute a reasonable average for SF, n should range over at least one seasonal cycle in most cases.19SF= 408Example (Contd)R
19、ange of the forecast0BiasnFAnttt1 F3=1064RangeIf forecast errors are normally distributed and the forecast is at the mean of the distribution, i.e., ,a forecast confidence band can be computed. The error distribution for the level-only model results is:Bias should be 0 or close to it in a model of g
20、ood fitCR (2004) Prentice Hall, Inc.8-1920CR (2004) Prentice Hall, Inc.Example (Contd)From a normal distribution table, z95%=1.96. The actual time series value Y for quarter 3 is expected to range between:or264 Y 18648001064)408(96.11064)(3FSzFY21CR (2004) Prentice Hall, Inc.校正趋势校正趋势Correcting for T
21、rend in ESThe trend-corrected model is St = aAt (1 a)(St-1 Tt-1) Tt = (St St-1) (1 )Tt-1Ft+1 = St Ttwhere S is the forecast without trend correction.Assuming a = 0.2, = 0.3, S-1 = 975, and T-1 = 0 Forecast for quarter 1 of this yearS0 = 0.2(1100) 0.8(975 + 0) = 1000T0 = 0.3(1000 975) 0.7(0) = 8F1 =
22、1000 8 = 100822Forecast for quarter 2 of this year S0 T0S1 = 0.2(1400) 0.8(1000 8) = 1086.4T1 = 0.3(1086.4 1000) 0.7(8) = 31.5F2 = 1086.4 31.5 = 1117.9Forecast for quarter 3 of this yearS2 = 0.2(1000) 0.8(1086.4 31.5) = 1094.3T2 = 0.3(1094.3 1086.4) 0.7(31.5) = 24.4F3 = 1094.3 24.4 = 1118.7, or 1119
23、CR (2004) Prentice Hall, Inc.Correcting for Trend in ES (Contd)23CR (2004) Prentice Hall, Inc.Correcting for Trend in ES (Contd)Summarizing with trend correction 1234Last year12007009001100This year14001000?Fore- cast100811181119Quarter24a01Fore-casterrorCR (2004) Prentice Hall, Inc.Optimizing a a f
24、or ESMinimize averageforecast error8-2425CR (2004) Prentice Hall, Inc.Controlling Model Fit in ESMSEFAtt signal TrackingTracking signal monitors the fit of the model to detect when the model no longer accurately represents the datawhere the Mean Squared Error (MSE) isntFtAMSEnt12)(If tracking signal
25、 exceeds a specified value (control limit), revise smoothing constant(s).n is a reasonable numberof past periods dependingon the application8-25268.3.3经典时间序列分解模型经典时间序列分解模型Classic Time Series Decomposition ModelBasic formulation F = T S C Rwhere F = 需求预测forecast T = 趋势水平trend S = 季节指数seasonal index C
26、 = 周期指数cyclical index (usually 1) R = 残差指数residual index (usually 1)Some time series data1234Last year12007009001100This year14001000?QuarterCR (2004) Prentice Hall, Inc.27CR (2004) Prentice Hall, Inc.Classic Time Series Decomposition Model (Contd)Trend estimationUse simple regression analysis to fi
27、nd the trend equation of the form T = a bt. Recall the basic formulas:22)(t nttYntYbandtbYa28CR (2004) Prentice Hall, Inc.Classic Time Series Decomposition Model (Contd)Redisplaying the data for ease of computation.tYYtt2112001200127001400439002700941100440016514007000256 6 1000600036 t=21 Y=6300Yt=
28、22700 t2=9129Classic Time Series Decomposition Model (Contd)Hence,andthen26(21/6)9100/6)6(21/6)(6322700b920.01)37.14(21/666300aT = 920.01 37.14tForecast for 3rd quarter of this year is:T = 920.01 37.14(7) = 1179.99CR (2004) Prentice Hall, Inc.30CR (2004) Prentice Hall, Inc.Classic Time Series Decomp
29、osition Model (Contd)Compute seasonal indicesThe procedure is to form a ratio of actual demand to the estimated demand for a full seasonal cycle (4 quarters). One way is as follows.tYTSeasonalIndex, St11200957.15*1.25*2700994.290.7039001031.430.87411001068.571.03*T=920.01 37.14(1)=957.15*St=1200/957
30、.15=1.2531CR (2004) Prentice Hall, Inc.Classic Time Series Decomposition Model (Contd)Compute seasonal indicesSince C and R index values are usually 1, the adjusted seasonal forecast for the 3rd quarter of this year would be:F7 = 1179.99 x 0.87 = 1026.59 32CR (2004) Prentice Hall, Inc.Classic Time S
31、eries Decomposition Model (Contd)Forecast rangeThe standard error of the forecast is:1)(12nFYSntttFSF预测的标准误差Yt第t期的实际需求Ft第t期的预测值N预测期t的数量33CR (2004) Prentice Hall, Inc.Classic Time Series Decomposition Model (Contd)QtrtYtTtStFt111200957.151.2522700994.290.70339001031.430.874411001068.571.031514001105.
32、711.271404.25*2610001142.850.881005.71*371179.991026.59*1105.71x1.27=1404.25*1142.85x0.88=1005.71Tabled computations34CR (2004) Prentice Hall, Inc.Classic Time Series Decomposition Model (Contd)There is inadequate data to make a meaningful estimate of SF. However, we would proceed as follows:infinit
33、y 121005.71)(10001404.25)(140022FSThen,Ft z(SF) Y Ft z(SF)Normally, a larger sample size would be used giving a positive value for SF35CR (2004) Prentice Hall, Inc.8.3.4回归分析回归分析Regression Analysis基本式Basic formulationF = o 1X1 2X2 nXn ExampleBobbie Brooks, a manufacturer of teenage womens clothes, wa
34、s able to forecast seasonal sales from the following relationshipF = constant 1(no. nonvendor accounts) 2(consumer debt ratio)36CR (2004) Prentice Hall, Inc.Sales period(1)Timeperiod, t(2)Sales (Dt )($000s)(3)Dt t(4)t2(5)Trend value(Tt)(6)=(2)/(5)SeasonalindexForecast($000s)Summer1$9,4589,4581$12,05
35、30.78Trans-season211,54223,084412,5390.92Fall314,48943,467913,0251.11Holiday415,75463,0161613,5121.17Spring517,26986,3452513,9981.23Summer611,51469,0843614,4840.79Trans-season712,62388,3614914,9700.84Fall816,086128,6886415,4561.04Holiday918,098162,8828115,9421.14Spring1021,030210,300 10016,4281.28Su
36、mmer1112,788140,668 12116,9150.76Trans-season1216,072192,864 14417,4010.92Fall13?17,887*$18,602Holiday14 ? 18,373*20,945Totals78176,723 1,218,217 650Regression Forecasting Using Bobbie Brooks Sales DataN = 12 Dt t = 1,218,217 t2 = 650 =(,/),.176 7231214 726 92 =78126 5/.Regression equation is: Tt =
37、11,567.08 + 486.13t *Forecasted valuesDt8-35378.3.5组合模型预测组合模型预测 Combined Model ForecastingCombines the results of several models to improve overall accuracy. Consider the seasonal forecasting problem of Bobbie Brooks. Four models were used. Three of them were two forms of exponential smoothing and a
38、 regression model. The fourth was managerial judgement used by a vice president of marketing using experience. Each forecast is then weighted according to its respective error as shown below.Calculation of forecast weightsModeltype(1)Forecasterror(2)Percentof totalerror(3)=1.0/(2)Inverse oferrorprop
39、ortion(4)=(3)/48.09ModelweightsMJ9.00.4662.150.04R0.70.03627.770.58ES11.20.06315.870.33ES28.40.4352.300.05 Total19.31.00048.091.00CR (2004) Prentice Hall, Inc.38Combined Model ForecastingCombines the results of several models to improve overall accuracy. Consider the seasonal forecasting problem of
40、Bobbie Brooks. Four models were used. Three of them were two forms of exponential smoothing and a regression model. The fourth was managerial judgement used by a vice president of marketing using experience. Each forecast is then weighted according to its respective error as shown below.Calculation
41、of forecast weightsModeltype(1)Forecasterror(2)Percentof totalerror(3)=1.0/(2)Inverse oferrorproportion(4)=(3)/48.09ModelweightsMJ9.00.4662.150.04R0.70.03627.770.58ES11.20.06315.870.33ES28.40.4352.300.05 Total19.31.00048.091.00CR (2004) Prentice Hall, Inc.39Combined Model Forecasting (Contd)Weighted
42、 Average Fall Season Forecast Using Multiple Forecasting TechniquesForecasttype(1)Modelforecast(2)Weightingfactor(3)=(1) (2)WeightedproportionRegressionmodel (R)$20,367,0000.58$11,813,000ExponentialSmoothingES120,400,0000.336,732,000Combinedexponentialsmoothing-regressionmodel(ES2)17,660,0000.05883,
43、000Managerialjudgment(MJ)19,500,0000.04 780,000 Weighted average forecast $20,208,000CR (2004) Prentice Hall, Inc.40CR (2004) Prentice Hall, Inc.Multiple Model Errors8-3841CR (2004) Prentice Hall, Inc.Actions When Forecasting is Not AppropriateSeek information directly from customersCollaborate with
44、 other channel membersApply forecasting methods with caution (may work where forecast accuracy is not critical)Delay supply response until demand becomes clearShift demand to other periods for better supply responseDevelop quick response and flexible supply systems42CR (2004) Prentice Hall, Inc.8.4
45、物流管理者遇到的特殊预测问题物流管理者遇到的特殊预测问题1.启动2.尖峰需求3.地区性预测4.预测误差43CR (2004) Prentice Hall, Inc.协同预测协同预测Collaborative Forecasting需求是块状或高度不确定Demand is lumpy or highly uncertainInvolves multiple participants each with a unique perspective“two heads are better than one”目标是减少预测误差Goal is to reduce forecast error预测过程本质
46、上是不稳定的The forecasting process is inherently unstable44CR (2004) Prentice Hall, Inc.Collaborative ForecastingDemand is lumpy or highly uncertainInvolves multiple participants each with a unique perspective“two heads are better than one”Goal is to reduce forecast errorThe forecasting process is inhere
47、ntly unstable45CR (2004) Prentice Hall, Inc.协同预测协同预测Collaborative Forecasting: 关键步骤关键步骤Key Steps建立一个主要过程Establish a process champion确定所需信息和收集流程Identify the needed Information and collection processes建立多来源信息和分配多权重的预测方法建立将预测转换成各方所需信息的方法Create methods for translating forecast into form needed by each p
48、arty建立实时预测和修正的过程Establish process for revising and updating forecast in real time创建预测方法Create methods for appraising the forecast协同预测带给各方的益处应该是明确而真实的Show that the benefits of collaborative forecasting are obvious and real46CR (2004) Prentice Hall, Inc.Collaborative Forecasting: Key StepsEstablish a
49、process championIdentify the needed Information and collection processesEstablish methods for processing information from multiple sources and the weights assigned to multiple forecastsCreate methods for translating forecast into form needed by each partyEstablish process for revising and updating f
50、orecast in real timeCreate methods for appraising the forecastShow that the benefits of collaborative forecasting are obvious and real47CR (2004) Prentice Hall, Inc.8.5灵活性和快速响应灵活性和快速响应管理高度不确定的需求管理高度不确定的需求Managing Highly Uncertain Demand尽可能长时间延迟预测根据产品的不确定程度供应（优先供应确定品)将延迟原则应用于最不确定的产品(delay committing
