Forecasting for Operations Everette S. Gardner, Jr. 1 Forecasting for operations Research themes The damped trend Case studies 1. Supply chain costs: Specialty chemicals 2. Manufacturing inventory investment: Snack foods 3. Purchasing workload: Water treatment systems Consequences of forecast errors How to evaluate forecast performance 2 Research themes Intermittent demand Distribution inventory management Biased forecasting Bullwhip effect Sensitivity of costs to forecast errors 3 Intermittent demand Empirical research is mixed - not clear that intermittent methods can beat SES No underlying model exists for the Croston method or any of its variants (Shenstone & Hyndman, IJF, 2005) Why not remove zeroes by aggregation? (Nikolopoulos et al.,JORS, 2011) 4 Distribution inventory management The damped trend gives better inventory performance than other exponential smoothing methods (Gardner, MS, 1990) Marginal improvements in forecast accuracy produce much larger improvements in inventory costs (Syntetos et al., IJF, 2010) 5 Biased forecasting Effects (Sanders & Graman, Omega,2009) Costs are more sensitive to bias than variance Over-forecasting produces lower costs than unbiased forecasting in an MRP environment Objections Conclusions depend on assumptions Safety stock is always a better option than adding bias to the forecasts 6 The bullwhip effect Definition Tendency of demand variability to increase as one moves up a supply chain Caused by lead times and forecast errors Is the bullwhip effect inevitable? Yes – But it can be reduced with centralized demand information (Chen et al., MS, 2000) No – Bullwhip effect is due to poor research design (Fildes & Kingsman, JORS, 2010) 7 Sensitivity of costs to forecast error Fildes and Kingsman (JORS, 2011) Research design Conclusions 8 MRP simulation Distinguishes between noise and specification error Demand processes are experimental factors Cost increases exponentially with demand uncertainty Cost benefits of improved forecasting are greater than the effects of choosing inventory decision rules Performance of the damped trend “The damped trend is a well established forecasting method that should improve accuracy in practical applications.” (Armstrong, IJF, 2006) “The damped trend can reasonably claim to be a benchmark forecasting method for all others to beat.” (Fildes et al., JORS, 2008) 9 Why the damped trend works Rationale The damped trend has an underlying random coefficient state space (RCSS) model that adapts to changes in trend (McKenzie & Gardner, IJF, 2011) Practice Fitting the damped trend is a means of automatic method selection from numerous special cases (Gardner & McKenzie, JORS, 2011) 10 yt t 1 At bt 1 vt SSOE state space models Constant coefficient yt t 1 bt 1 t Random coefficient yt t 1 At bt 1 vt t t 1 bt 1 h1 t t t 1 At bt 1 h*1 vt bt bt 1 h2 t bt At bt 1 h*2 vt {At} are i.i.d. binary random variates White noise innovation processes ε and v are different Parameters h and h* are related but usually different 11 Runs of linear trends in the RCSS model bt At bt 1 h*2 vt With a strong trend, {At } will consist of long runs of 1s with occasional 0s. With a weak trend, {At } will consist of long runs of 0s with occasional 1s. In between, we get a mixture of models on shorter time scales, i.e. damping. 12 Advantages of the RCSS model Allows both smooth and sudden changes in trend. is a measure of the persistence of the linear trend. The mean run length is thus /(1 ) and P( At 1) RCSS prediction intervals are much wider than those of constant coefficient models. 13 Methods automatically identified in the M3 time series Method 14 % Damped trend 43.0 Holt 10.0 SES w/ damped drift 24.8 SES w/ drift 2.4 SES 0.8 RW w/ damped drift 7.8 RW w/ drift 2.5 RW 0.0 Modified exp. trend 8.3 Linear trend 0.1 Simple average 0.3 Case 1: Chemicals supply chain Scope 4 plants: N. and S. America, Europe, Asia 10 component chemicals, 25 products 400 customers, 250,000 tons of annual production Production and transportation plans based on Damped trend Optimization Simulation 15 Examples of chemicals demand series 1 3 16 2 4 Scaled errors Average forecast error measures are misleading Drastic changes in scale Some observations near zero Alternative - Scaled errors (Hyndman & Koehler, 2006) 17 Based on in-sample, one-step errors from the naïve method If scaled error is less than 1, we beat the naïve method Mean absolute scaled error Horizons 1-6 All products Critical products 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 Holt 18 SES Damped Proportions of total demand for 25 time series 4% 4% 20% 23% 26% 19 Mean absolute scaled error Horizons 1-6 All products Critical products 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 Holt 20 SES Damped Supply chain model Damped trend Monthly demand forecasts Actual demand MIP: Minimize total supply chain cost Monthly production schedule 21 Inv. on hand Inv. in transit Backorders MIP: Disaggregate monthly schedule Simulation: daily mfg. & shipments Detailed weekly schedule Top-level mixed integer program (MIP) Objective: Minimize total supply chain costs, including 22 Inventory carrying Production Transportation Import tariffs Top-level MIP continued Data requirements Demand forecasts Pending orders Shipments in transit Inventory levels Machine and storage capacity Business rules for 23 Production run lengths Transportation modes Supply chain model Damped trend Monthly demand forecasts Actual demand MIP: Minimize total supply chain cost Monthly production schedule 24 Inv. on hand Inv. in transit Backorders MIP: Disaggregate monthly schedule Simulation: daily mfg. & shipments Detailed weekly schedule Second-level MIP Disaggregates top-level schedule Weekly schedule for each machine at each plant 12-week horizon Data requirements Forecasts Week-ending inventories Pending orders Scheduled in and out bound shipments Bootstrap safety stocks (Snyder et al., IJF, 2002) 25 Supply chain model Damped trend Monthly demand forecasts Actual demand MIP: Minimize total supply chain cost Monthly production schedule 26 Inv. on hand Inv. in transit Backorders MIP: Disaggregate monthly schedule Simulation: daily mfg. & shipments Detailed weekly schedule Simulation model Executes manufacturing plans on a daily basis using actual demand history Feeds production, inventories, backorders, and shipments to the MIP models Sources of uncertainty Demand Transportation lead times Machine breakdowns 27 Cost vs. weighted lateness (tons x days) 250,000 Weighted lateness Holt 200,000 SES 150,000 100,000 50,000 0 114 Damped trend 115 116 117 118 119 Total supply chain cost (Millions of $) 28 120 121 Cost vs. percentage of backorders Percentage of backorders 30% Holt 25% 20% SES 15% 10% Damped trend 5% 0% 114 115 116 117 118 119 Total supply chain cost (millions of $) 29 120 121 Case 2: Snack-food manufacturer Scope 82 snack foods Food stocks managed by commodity traders Packaging materials managed with subjective forecasts and EOQ/safety stock inventory rules Problems Excess stocks of perishable packaging materials Difficult to predict inventory on the balance sheet 30 11-Oz. corn chips Monthly packaging inventory and usage $2,500,000 $2,000,000 Actual Inventory from subjective forecasts $1,500,000 $1,000,000 $500,000 $0 Month 31 Monthly Usage Snack-food manufacturer Solution Automatic forecasting with the damped trend Retain EOQ/safety stock inventory rules 32 Damped-trend performance 11-oz. corn chips $500,000 Outlier $450,000 $400,000 $350,000 $300,000 $250,000 $200,000 33 Actual Forecast Investment analysis: 11-oz. corn chips Forecast annual usage (000s) Economic order quantity Standard deviation of forecast errors Probability of shortage 0.1 0.05 0.001 0.00001 0.0000001 34 $4,138 $318 $34 Safety Order Maximum stock quantity investment $44 $318 $362 $56 $318 $375 $106 $318 $424 $146 $318 $464 $177 $318 $496 Safety stocks vs. shortages 11-oz. corn chips $200,000 $180,000 Target Safety stock $160,000 $140,000 $120,000 $100,000 $80,000 $60,000 $40,000 $20,000 $0 0 10 20 30 40 50 60 70 Shortages per 1,000 order cycles 35 80 90 100 Safety stock vs. forecast errors 11-oz. corn chips $200,000 Safety stock $150,000 $100,000 $50,000 $0 ($50,000) ($100,000) ($150,000) ($200,000) 36 Forecast errors 11-Oz. corn chips Target vs. actual packaging inventory $2,500,000 $2,000,000 Actual Inventory from subjective Actual Inventory forecasts from subjective forecasts $1,500,000 $1,000,000 $500,000 $0 Target maximum inventory based on damped trend 37 Month Monthly Usage Forecasting regional demand Forecast total unit demand with the damped trend Forecast regional percentages with simple exponential smoothing 38 Regional sales percentages: Corn chips 50% South 40% 30% 20% East North West 10% 0% Mar 39 Jun Sep Dec Mar Jun Sep Dec Packaging inventory (millions of $): 82 products 183 200 180 135 160 140 120 100 80 60 40 20 0 Actual 40 Target Case 3: Water treatment company Scope Assembly of systems and distribution of supplies Annual sales = $16 million Inventory = $6 million (23,000 SKUs) Inventory system Reorder monthly to maintain 3 months of stock Numerous subjective adjustments Forecasting system 6-month weighted moving average Numerous subjective adjustments 41 Problems Forecasts vs. reality Annual forecasts on stock records = $29 million Annual sales = $16 million Purchasing workload 76,000 purchase orders per year Messy stock records 42 Dead stock Substitute items not linked to primary items Water treatment company: Inventory status 7,526 with no hits in 12 months 33% 6,336 active items 27% 43 2,200 obsolete 9% 2,928 substitute items 13% 4,202 with inadequate demand to stock 18% Solutions Forecast demand with the damped trend Develop a decision rule for what to stock Use the forecasts to do an ABC classification Replace the monthly ordering policy with a hybrid inventory control system: 44 Class A Class B Class C JIT EOQ/safety stock Annual buys Water treatment supplies: One-step MASE 1.2 1.0 0.8 0.6 0.4 0.2 0.0 Moving average 45 Moving Damped trend average + subjective adjustments What to stock? Cost to stock Average inventory balance x holding rate + Number of stock orders x transportation cost Cost to not stock Nbr. of customer orders x drop-ship transportation cost 46 ABC classification based on damped-trend forecasts 47 Class Sales forecast System Items Dollars A > $36,000 JIT 3% 75% B $600 - $35,999 EOQ 49% 18% C < $600 Annual buy 48% 7% Annual purchasing workload Total savings = 58,000 orders (76%) 40,000 Monthly ordering ABC system 35,000 30,000 25,000 EOQ 20,000 EOQ 15,000 Annual buys JIT 10,000 JIT 5,000 0 A 48 B C Inventory investment Total savings = $591,000 (15%) Monthly ordering ABC system 3,000,000 2,500,000 JIT 2,000,000 EOQ 1,500,000 EOQ 1,000,000 Annual buys JIT 500,000 0 A 49 B C Consequences of forecast errors Limited capacity creates interactions amongst products: Under-forecasting Over-forecasting 50 Chain reaction of backorders Premium transportation Excess stocks Chain reaction of backorders (limited capacity put to wrong use) Premium transportation Consequences of forecast errors (cont.) Errors often reverse themselves before system has fully responded to 51 Backorders, or Excess stocks How to evaluate forecast performance Operational measures Backorder delay time Percent of time in stock Percent of orders filled immediately Number of purchase orders or production setups Financial measures Manufacturing, distribution, and supply chain costs Value of backorders Inventory investment on the balance sheet 52 Future research Research is needed: In real operating systems Gardner & Makridakis (IJF,1988) On the benefits of improved forecasting Fildes & Kingsman (JORS, 2010) On the relationship between forecast accuracy and operational performance Syntetos et al. (IJF, 2010) 53 Presentation and papers available at www.bauer.uh.edu/gardner 54