CTL.SC1x -Supply Chai & Logistics Fudametals Itroductio to Demad Plaig & Forecastig MIT Ceter for Trasportatio & Logistics
Demad Process Three Key Questios What should we do to shape ad create demad for our product? Demad Plaig Product & Packagig Promotios Pricig Place What should we expect demad to be give the demad pla i place? Demad Forecastig Strategic, Tactical, Operatioal Cosiders iteral & exteral factors Baselie, ubiased, & ucostraied How do we prepare for ad act o demad whe it materializes? Demad Maagemet Balaces demad & supply Sales & Operatios Plaig (S&OP) Bridges both sides of a firm Material adapted from Lapide, L. (2006) Course Notes, ESD.260 Logistics Systems. 2
Forecastig Levels Level Horizo Purposes Strategic Tactical Operatioal Year/Years Quarterly Moths/Weeks Days/Hours Busiess Plaig Capacity Plaig Ivestmet Strategies Brad Plas Budgetig Sales Plaig Mapower Plaig Short-term Capacity Plaig Master Plaig Ivetory Plaig Trasportatio Plaig Productio Plaig Ivetory Deploymet Material adapted from Lapide, L. (2006) Course Notes, ESD.260 Logistics Systems. 3
Ageda Forecastig Truisms Subjective vs. Objective Approaches Forecast Quality Forecastig Metrics
Forecastig Truisms 1: Forecasts are always wrog 5
1. Forecasts are always wrog Why? Demad is essetially a cotiuous variable Every estimate has a error bad Forecasts are highly disaggregated w Typically SKU-Locatio-Time forecasts Thigs happe... OK, so what ca we do? Do t fixate o the poit value Use rage forecasts Capture error of forecasts Use buffer capacity or stock 6
Forecastig Truisms 2: Aggregated forecasts are more accurate 7
2. Aggregated forecasts are more accurate Aggregatio by SKU, Time, Locatio, etc. Coefficiet of Variatio (CV) Defiitio: Stadard Deviatio / Mea = σ/µ Provides a relative measure of volatility or ucertaity CV is o-egative ad higher CV idicates higher volatility 200 Red: µ=100, σ=45, CV=0.45 Blue: µ=100, σ=1, CV=0.01 180 160 Daily Demad 140 120 100 80 60 40 20-2/26/11 3/28/11 4/27/11 5/27/11 6/26/11 7/26/11 8/25/11 8
Aggregatig by SKU Coffee Cups ad Lids @ the Sadwich Shop Large ~N(80, 30) CV = 0.38 Medium ~N(450, 210) CV = 0.47 Small ~N(250, 110) CV = 0.44 800 700 600 500 400 300 Small Medium Large 200 100-8/14/13 9/13/13 10/13/13 11/12/13 12/12/13 1/11/14 2/10/14 3/12/14 4/11/14 5/11/14 6/10/14 9
Aggregatig by SKU What if I desig cups with a commo lid? Commo Lid ~N(780, 239) CV = 0.31 µ = (80 + 450 + 250) = 780 uits/day σ = sqrt(302 + 210 2 + 110 2 ) = 239 uits/day Large ~N(80, 30) CV=0.38 Med. ~N(450, 210) CV=0.47 Small ~N(250, 110) CV=0.44 Lids ~N(780, 239) CV=0.31 1,600 1,400 1,200 1,000 800 600 400 200-8/14/13 9/13/13 10/13/13 11/12/13 12/12/13 1/11/14 2/10/14 3/12/14 4/11/14 5/11/14 6/10/14 Example of Modularity or Parts Commoality Reduces the relative variability Icreases forecastig accuracy Lowers safety stock requiremets 10
1,600 1,200 Aggregatig by Time Daily Demad for Lids ~N(780, 239) CV=0.31 Forecasts with loger time buckets have better forecast accuracy. The time bucket used should match the situatio. 800 400-8/14/13 9/13/13 10/13/13 11/12/13 12/12/13 1/11/14 2/10/14 3/12/14 4/11/14 5/11/14 6/10/14 8,000 Weekly Demad for Lids ~N(5458, 632) CV=0.12 6,000 4,000 2,000-1 5 9 13 17 21 25 29 33 37 41 45 49 30,000 Mothly Demad for Lids ~N(21840, 1264) CV=0.06 25,000 20,000 15,000 10,000 5,000-1 2 3 4 5 6 7 8 9 10 11 12 11
Aggregatig by Locatios Suppose we have three sadwich shops Weekly lid demad at each ~N(5458, 632) CV=0.12 CV reduces as we aggregate over SKUs, time, or locatios. ~N(5458, 632) ~N(5458, 632) ~N(5458, 632) ~N(16374, 1095) What if demad is pooled at a commo Distributio Ceter? Weekly lid demad at DC ~N(16374, 1095) CV=0.07 CV id = σ µ CV agg = σ µ = σ µ = CV id 12
Forecastig Truisms 3: Shorter horizo forecasts are more accurate 13
3. Shorter horizo forecasts are more accurate???? 21 22 23 24 1 2 3 4 5 6 7 8 9 10 11 12 14
3. Shorter horizo forecasts are more accurate Postpoed fial customizatio to closer time of cosumptio Risk poolig of compoet (e.g., ham) icreases forecast accuracy.? 21 22 23 24 1 2 3 4 5 6 7 8 9 10 11 12 15
Forecastig Truisms Forecasts are always wrog è Use rages & track forecast error Aggregated forecasts are more accurate è Risk poolig reduces CV Shorter time horizo forecasts are more accurate è Postpoe customizatio util as late as possible 16
Subjective & Objective Approaches 17
Fudametal Forecastig Approaches Subjective Judgmetal Sales force surveys Jury of experts Delphi techiques Experimetal Customer surveys Focus group sessios Test marketig Causal / Relatioal Ecoometric Models Leadig Idicators Iput-Output Models Time Series Objective Black Box Approach Past predicts the future Idetify patters Ofte times, you will eed to use a combiatio of approaches
Forecastig Quality 19
Cost of Forecastig vs Iaccuracy ß Overly Naïve Models à ß Good Regio à ß Excessive Causal Models à Total Cost Cost Cost of Errors I Forecast Cost of Forecastig Forecast Accuracy
How do we determie if a forecast is good? What metrics should we use? Example - Which is a better forecast? Squares & triagles are differet forecasts Circles are actual values 1100 1000 900 time
Accuracy versus Bias Accuracy - Closeess to actual observatios Bias - Persistet tedecy to over or uder predict Accurate Not Accurate Biased Not Biased
Forecastig Metrics 23
Forecastig Metrics e t = A t F t Mea Deviatio (MD) Mea Squared Error (MSE) MD MSE = = t= 1 t= 1 e t e 2 t Mea Absolute Deviatio (MAD) Root Mea Squared Error (RMSE) MAD = RMSE = t= 1 t= 1 e e t 2 t Mea Percet Error (MPE) Notatio: MPE = e t t=1 A t Mea Absolute Percet Error (MAPE) MAPE = A t = Actual value for obs. t e t = Error for observatio t F t = Forecasted value for obs. t = Number of observatios t=1 e t A t
Example: Forecastig Bagels For the bagel forecast ad actual values show below, fid the: Mea Absolute Deviatio (MAD) Root Mea Square of Error (RMSE) Mea Absolute Percet Error (MAPE) MAD = t= 1 e t Forecast Actual Moday 50 43 Tuesday 50 42 Wedesday 50 66 RMSE = t= 1 e 2 t Thursday 50 38 Friday 75 86 MAPE = t=1 e t A t 25
Example: Forecastig Bagels Solutio: 1. Graph it. 2. Exted data table: 90 80 Forecast Actual w Error: e t =A t -F t w Abs[error] = e t w Sqr[error] = e 2 w AbsPct[error] = e t /A t 3. Sum ad fid meas Daily Bagel Demad 70 60 50 40 F t A t e t e t e 2 e t /A t Moday 50 43-7 7 49 16.3% 30 Moday Tuesday Wedesday Thursday Friday Tuesday 50 42-8 8 64 19.0% Wedesday 50 66 16 16 256 24.2% Thursday 50 38-12 12 144 31.6% Friday 75 86 11 11 121 12.8% Sum 0 54 634 104% MAD = 54/5 = 10.8 RMSE = sqrt(126.8) = 11.3 MAPE = 104%/5 = 21% Mea 0 10.8 126.8 21% 26
Key Poits from Lesso 27
Key Poits Forecastig is a meas ot a ed Forecastig Truisms Forecasts are always wrog Aggregated forecasts are more accurate Shorter horizo forecasts are more accurate Subjective & Objective Approaches Judgmetal & experimetal Causal & time series Forecastig metrics Capture both bias & accuracy MAD, RMSE, MAPE
CTL.SC1x -Supply Chai & Logistics Fudametals Questios, Commets, Suggestios? Use the Discussio! Jaie Photo courtesy Yakee Golde Retriever Rescue (www.ygrr.org) MIT Ceter for Trasportatio & Logistics caplice@mit.edu