Strategic Safety Stocks in Supply Chains: Update on Recent Work Stephen C. Graves MIT, E53-347 sgraves@mit.edu http://web.mit.edu/sgraves/www/ Joint work with Sean Willems, Boston University, Katerina Lesnaia, Oracle, Tor Schoenmeyr, FirstSolar Page 1 Stephen C. Graves Copyright 2010 All Rights Reserved Overview • Motivation and assumptions for SIP model • Prior work – review of base model & example (joint with Willems) • Recent work – extend to account for capacity (joint with Schoenmeyr) • Recent work – extend to include evolving forecasts (joint with Schoenmeyr) • Summary Papers available on request! Page 2 Stephen C. Graves Copyright 2010 All Rights Reserved Strategic Safety Stock Model: Intent • Tactical model to determine the amount and positioning of safety stocks in supply chains • Tactical model to support supply chain improvement teams • Simple model, easily accessible, runs on PC, understandable inputs/outputs; academic version available from www.sipmodel.com • Commercialized by Optiant; applications support both tactical and operational decisions Page 3 Stephen C. Graves Copyright 2010 All Rights Reserved Assumptions • Supply chain modeled by an acyclic graph • Deterministic processing time for each stage • No capacity constraints • Deterministic yield • Periodic review, base stock control for each stage; common review period and no lot sizing Page 4 Stephen C. Graves Copyright 2010 All Rights Reserved Assumptions • Fixed service time between stages where service time is the decision variable • Each stage quotes same service time to all adjacent downstream stages • Stationary, bounded demand process for each end item • Each stage provides 100% service: “Guaranteed service model” Page 5 Stephen C. Graves Copyright 2010 All Rights Reserved Review of guaranteed-service base-stock problem Orders d(t) Orders d(t) Stage k Processing Inventory Service time Sk 1 Processing time Tk Service time Sk At time t S k node k must deliver d(t) to the downstream node from its inventory At time t Sk 1, d(t) units are delivered as raw material from node k + 1, and at time t Tk Sk 1 the d(t) units are ready as inventory at node k Stage k must have a base stock level equal to max demand over the net replenishment time Sk 1 Tk Sk k Page 6 Stephen C. Graves Copyright 2010 All Rights Reserved Base stock mechanics B is base stock level. SI S k 1 ; S I (t) is inventory at end of time t. Sk Demand arrives: received d0 … dt-SI-T-1 … dt-SI-T dt-S … dt shipped d0 … dt-SI-T received shipped I Page 7 d0 … dt-S Stephen C. Graves Copyright 2010 All Rights Reserved Key results for guaranteed service, bounded demand I k t Bk d k t Sk 1 Tk 1 d k t Sk 1 Tk 2 I k t 0 Bk Max d k t Sk 1 Tk 1 dk t Sk d k t S k Dk S k 1 Tk S k For Dk z E I k t Bk Sk 1 Tk Sk z Sk 1 Tk Sk Page 8 Stephen C. Graves Copyright 2010 All Rights Reserved Review of guaranteed-service base-stock problem Stage k+1 Stage k Processing Inventory Tk 1 Safety stock Stage k-1 Processing Inventory Sk 1 Tk z Sk 2 Tk 1 Sk 1 Processing Inventory Sk z Sk 1 Tk Sk Tk 1 Sk 1 z Sk 1 Tk Sk If we view the service times as decision variables we get a global optimization problem: N min hk z S k 1 Tk S k Sk k 1 S k 0 k S1 s1 S k 1 Tk S k 0 k Page 9 Stephen C. Graves Copyright 2010 All Rights Reserved Review of guaranteed-service base-stock problem N min hk S k 1 Tk S k Sk k 1 S k 0 k S1 s1 S k 1 Tk S k 0 k Simpson (1958): Solve serial system through enumeration. “All-or-nothing” property of optimal solution (i.e., either Sk 0 or Sk Sk 1 Tk ) Graves and Willems (2000): Solve spanning tree system through polynomial-time dynamic programming (Lesnaia, 2004). Fast enough for large, real-life applications. Page 10 Stephen C. Graves Copyright 2010 All Rights Reserved Algorithmic Results • For serial systems, Simpson (1958) showed the all or nothing property for solution • Graves and Willems (2000) developed a pseudopolynomial DP for spanning trees; also Graves (1988), Inderfurth (1991) and Inderfurth and Minner (1998) • Lesnaia (2004) provides polynomial DP for spanning tree and specialized algorithm for any two layer network • General network is NP hard (Lesnaia, 2004); optimum occurs at an extreme point for concave bound function • Several exact and heuristic algorithms for general networks: Humair and Willems (2006, 2008); Lesnaia (2004); Minner (2000); Magnanti et al. (2006) Page 11 Stephen C. Graves Copyright 2010 All Rights Reserved KIMES 100 Page 12 Stephen C. Graves Copyright 2010 All Rights Reserved Supply Chain: Before Page 13 Stephen C. Graves Copyright 2010 All Rights Reserved Supply Chain: Lead Times Page 14 Stephen C. Graves Copyright 2010 All Rights Reserved Supply Chain: Costs Page 15 Stephen C. Graves Copyright 2010 All Rights Reserved Supply Chain: Optimized Page 16 Stephen C. Graves Copyright 2010 All Rights Reserved Supply Chain: Implemented Page 17 Stephen C. Graves Copyright 2010 All Rights Reserved Supply-Chain Page 18 Stephen C. Graves Copyright 2010 All Rights Reserved KIMES 100 • Project results –Sizing finished goods inventory –Assess where to target lead-time reduction efforts –Framework to work with suppliers on purchasing long lead-time parts Page 19 Stephen C. Graves Copyright 2010 All Rights Reserved Key Benefits & Learning • Shows value from “holistic” perspective • Formalizes inventory-related supply chain costs, and provides an optimal benchmark • Provides framework and standard terminology for crossfunctional debate • Shows the effectiveness of inventory, strategically positioned in a few places to de-couple the supply chain • De-couple supply chain prior to a high-cost added stage; and prior to product explosion • Most leverage from lead time reduction Page 20 Stephen C. Graves Copyright 2010 All Rights Reserved Battery Supply Chain Page 21 Stephen C. Graves Copyright 2010 All Rights Reserved HP Supplies Inventory Modeling Project Suppliers Factories Regions Customers Answer the bulk pen inventory question… … in the context of what is best for the system Page 22 Stephen C. Graves Copyright 2010 All Rights Reserved # Models Built 8000 7000 6000 5000 4000 3000 IPG Builds Roughly 1000 Network Models per Year standard training class and tier based support 2000 1000 0 2001 2002 2003 2004 2005 2006 2007 2008 2009 Key Limitations, circa 2000 • • • • • • Stationary demand assumptions No capacity constraints DP algorithm for spanning tree only Deterministic lead times Common review period Common service time to all downstream customers Page 24 Stephen C. Graves Copyright 2010 All Rights Reserved Strategic safety stocks in supply chains with evolving forecasts • Same assumptions as for base case, but now there is an evolving forecast for the end item demand • Guaranteed service model – each stage commits to a guaranteed service but now for a bound on forecast errors • Each stage uses a forecast-based ordering policy, rather than a base stock policy Page 25 Stephen C. Graves Copyright 2010 All Rights Reserved Forecast evolution model Graves et al. (1986), Heath and Jackson (1994) ft (t i) is our forecast, in period t, for demand in period t + i The forecast is initialized as ft (t H ) at the horizon Each period the forecasts are revised Current forecast is demand ft (t ) d (t ) ft (t i) ft 1 (t i) ft (t i) forecast revision Assumptions (“rational forecasts”): (A1) ft (t i) are i.i.d. R.V. (A2) E[ft ( j )] 0 Unlike previous authors, we make no assumptions about ft (t H ) We show an equivalence with this model and general, state-space models of demand (e.g., ARIMA) Page 26 Stephen C. Graves Copyright 2010 All Rights Reserved Lk Sk 1 j 1T j cumulative lead time k Order placed by stage k Forecast-based order policy: Pk (t ) f t (t Lk ) Scheduled future demand Lk 1 f (t i) i 0 t . Downstream schedule changes For zero service times, this corresponds to orders in a simple (no lot sizing, etc) MRP system. For zero service times, the forecast-based orders have some local optimality properties (Aviv, 2003) Both the forecast evolution model, and similar order mechanisms have been considered before; new contribution is to consider non-zero service times in a global optimization problem Page 27 Stephen C. Graves Copyright 2010 All Rights Reserved We can use the equations for the evolving forecast and the order policy to derive the inventory t k t Lk I k (t Sk 1 Tk ) I fi ( j ) 0 k i t 1 j i Safety stock constant Forecast revisions t k t Lk If we can find a bound on the sum: f ( j ) i t 1 j i and set the safety stock level to i t t k t Lk 0 I k max fi ( j ) i t 1 j i then the stage can guarantee service; i.e., Page 28 Ik t 0 Stephen C. Graves Copyright 2010 All Rights Reserved How might we set the bound? tL D For the cumulative forecast error: j t 1 Define j ft ( j ) tL F ( L) D j f t ( j ) j t 1 D is demand RV; σ() is now a function too! F ( L) can be calculated from historical data on demand and forecasts We find that t k t Lk fi ( j ) F 2 Lk F 2 Lk 1 i t 1 j i which we propose as a bound. By setting the safety stock level I k0 z F 2 Lk F 2 Lk 1 we have a valid (probabilistic) bound. Loosely speaking, the stages will provide guaranteed service as long as the cumulative forecast errors are smaller than zF ( L ) Page 29 Stephen C. Graves Copyright 2010 All Rights Reserved L2 L3 Forecast error L1 1 2 3 4 5 6 7 8 9 10 j Time into the future Optimization problem: how do we find the least cost safety stock configuration that maintains guaranteed service for any forecast/demand realization within the bounds N min hk F 2 Lk F 2 Lk 1 Sk k 1 Lk Sk 1 j 1 T j k k Sk 0 k S1 s1 Sk 1 Tk Sk 0 k Page 30 Stephen C. Graves Copyright 2010 All Rights Reserved Forecast problem Base stock problem N min hk F 2 Lk F 2 Lk 1 Sk N min hk z S k 1 Tk S k k 1 Lk Sk 1 j 1 T j k Sk k k 1 S k 0 k Sk 0 k S1 s1 S1 s1 S k 1 Tk S k 0 k Sk 1 Tk Sk 0 k The problem is very similar to the base stock problem solved by Simpson (1958), and extended by Graves and Willems (2000) and others. Under some mild assumptions about the forecasts, we show that the allor-nothing property holds We can use existing, effective algorithms to find optimal service times, after modifying the bound function Page 31 Stephen C. Graves Copyright 2010 All Rights Reserved Managerial insight nugget • Assuming forecast revisions at time t are independent and with proposed bound on forecast errors: I z S k 1 Tk S k 0 k j Lk j Lk 1 1 2 D t j ; ft t j Page 32 Stephen C. Graves Copyright 2010 All Rights Reserved • Electronic test system manufactured by Teradyne, Inc. • 3,866 part/locations • Used real data on supply chain topology, lead times, costs of parts Page 33 Stephen C. Graves Copyright 2010 All Rights Reserved Correlation forecast - demand 1.2 1 0.8 0.6 0.4 0.2 0 -0.2 0 1 2 3 4 5 6 7 8 9 10 Weeks into the future • Schedule contained booked and “preliminary” orders, and got increasingly locked down as the date of delivery approach • The schedule was effectively a forecast, and we used data on past schedule changes to calculate F(L) • As a forecast of actual demand, it was fairly accurate in the short term but useless >10 weeks out Page 34 Stephen C. Graves Copyright 2010 All Rights Reserved Total cost ? ? ??? 25.5% improvement Difficult to compare with current situation because no consistent optimization procedure/ service level used Current Optimal base stock Optimal forecast/MRP • In the forecasted case, most savings were far downstream, where forecasts were accurate • Optimization time ~1 minute on a laptop computer • Schoenmeyr thesis discusses generalizations for multi-product networks Page 35 Stephen C. Graves Copyright 2010 All Rights Reserved 2 Base stock policy (Graves-Willems) Total holding costs 5 week forecast 10 week forecast (current) 1 20 week forecast 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Total lead time (weeks) Page 36 Stephen C. Graves Copyright 2010 All Rights Reserved Summary of results • We have shown how to map the optimization method used for base stock systems, so that it can be used for forecast-driven (push) systems • This approach enables optimization of large system with • Evolving schedule in make-to-order context • Evolving demand forecast in make-to-stock context • Benefit relative to base stock case depends on forecast quality; in one case study it was ~25% Page 37 Stephen C. Graves Copyright 2010 All Rights Reserved Strategic safety stocks in supply chains with capacity constraints • Same assumptions as before, but now there can be a capacity constraint at each stage • Guaranteed service model – each stage commits to a guaranteed service for bounded demand • Deterministic production lead time T • Each stage follows a base stock policy, subject to capacity constraint Page 38 Stephen C. Graves Copyright 2010 All Rights Reserved Open problem: what if there are capacity constraints? Orders d(t) Orders d(t) Stage k “stuck” units Service time Sk 1 ck / period Processing Processing time Inventory Tk Service time Sk Now z Sk 1 Tk Sk may not be enough safety stock, because any units that get “stuck” will be delayed. Q: How much extra inventory do we need? Q: How do we optimize a supply chain with one or more capacity constraints? Q: Do the structural results from before (“all or nothing”) hold up? Page 39 Stephen C. Graves Copyright 2010 All Rights Reserved Q: How much extra inventory do we need? A: The base stock level for the node with capacity constraint can be calculated with a functional transformation of the unconstrained base stock level. In general Original base stock level/order bound: B( ) D( ), for Sk 1 Tk Sk Base stock with capacity constraint: B( ) ( D)( ) max D( n) ck n n0 Page 40 Stephen C. Graves Copyright 2010 All Rights Reserved Q: How much extra inventory do we need? A: The base stock level for the node with capacity constraint can be calculated with a functional transformation of the unconstrained base stock level. Common example Original base stock level/order bound: Base stock with capacity constraint: B( ) D( ) 2 2 ck (ck ) B( ) ( D)( ) D 2 for (ck ) 2 2 for (ck )2 The base stock grows hyperbolically as we decrease capacity. On the other hand, if the capacity constraint is large enough it becomes irrelevant. Page 41 Stephen C. Graves Copyright 2010 All Rights Reserved 2 ck (ck ) ( D)( ) D 2 for (ck ) 2 for 2 (ck ) 2 D( ) 2 With a capacity constraint, we permit negative net replenishment time A stage with a capacity constraint needs safety stock even at zero net replenishment time Page 42 For sufficiently large net replenishment times the capacity constraint does not matter Stephen C. Graves Copyright 2010 All Rights Reserved Q: How do we optimize a supply chain with one or more capacity constraints? A: We have shown how to calculate new base stock levels for a single stage. Other stages are not affected (orders placed/delivered as before). Hence after transforming affected bounds, we can use existing optimization procedures. Q: Do the structural results from before (“all or nothing”) hold up? A: The functional transformation ( D)( ) max D( n) ck n n 0 preserves concavity, and hence the “all or nothing” property holds. Page 43 Stephen C. Graves Copyright 2010 All Rights Reserved But can we do better? CensoredOrders orders dd(kt()t ) Orders d k(t1)(t ) Stage k BLk (t ) “stuck” units Service time Sk 1 ck / period Processing Processing time Inventory Tk Service time Sk “Why should we ever order ck units if we cannot process them when they arrive?” Place the censored order dk (t ) min(dk 1 (t ) BLk (t 1), ck ) where we keep a backlog of delayed orders BLk (t ) max BLk (t 1) dk 1 (t ) ck ,0 Page 44 Stephen C. Graves Copyright 2010 All Rights Reserved Q: But how much inventory do we need/ how do we optimize supply chain? A: We find that base stock transformation remains the same but we need another functional transformation to obtain a new bound for orders (demand) placed by a censoring node k 1Dk ( ) min(ck , Dk ( )) Order bound Dk 1 ( ) k 1Dk ( ) Dk ( ) k Dk 1 ( ) Dk 1 ( ) k 1Dk 2 ( ) Base stock Bk 1 ( ) k 1Dk ( ) Bk ( ) k Dk 1 ( ) Bk 1 ( ) k 1Dk 2 ( ) ck 1 ck ck 1 Q: Do the structural results from before (“all or nothing”) hold up? A: Yes. (Φ also preserves concavity) Page 45 Stephen C. Graves Copyright 2010 All Rights Reserved Serial system with 8 nodes and capacity constraint at node 3. Assumed processing time 5 at each node; holding costs increase with 40% per stage. Average Inventory 8 7 6 5 4 3 2 1 Total cost No capacity constraint 0 0 0 17.9 0 0 0 17.9 2,377 Capacity constraint, no censorship 0 0 15.5 0 0 15.5 0 12.6 2,433 Capacity constraint, censorship 5 5 5 5 5 6.5 0 12.6 2,233 • Censorship reduces cost impact of constraint • Censorship cost is sometimes even lower than uncapacitated problem! • “Paradox”: Under censorship, add constraint → better solution • Explanation 1: Censorship smoothes demand and reduces safety stocks upstream • Explanation 2: The (uncensored) local base stock policy is not optimal in a multistage system with guaranteed service • It may be of interest to censor even in the absence of actual capacity constraints Page 46 Stephen C. Graves Copyright 2010 All Rights Reserved Summary of results for capacity constraints • We can generalize the base stock model to incorporate capacity constraints. • For serial systems, we find exact analytical transformations, under which existing algorithms can be used with small modifications • Known structural results (“all-or-nothing”) hold. • These results also hold if we censor orders with the capacity. The necessary safety stocks are reduced. • Censored orders sometimes lead to costs that are even lower than for the same problem without capacity constraints (in many examples 30-40% reductions by censoring the right amount at the right location) • Development is for serial systems, and extends immediately to assembly structures; more general networks require a calculus to combine bounds Page 47 Stephen C. Graves Copyright 2010 All Rights Reserved Overall Summary • Motivation, assumptions and review of guaranteed service supply chain model • Extension for capacity – Requires transformation of base stock and of demand bound – Structural results and algorithms extend directly – Capacity constraint can lead to lower cost solution – Multi-item supply chains requires more work • Extensions for evolving forecast – Requires forecast-based ordering and bound on forecast errors – Structural results and algorithms extend directly – Incorporating forecast can lead to lower costs – Multi-item supply chains requires more work Page 48 Stephen C. Graves Copyright 2010 All Rights Reserved Example demand data 600 500 Maximum Demand Expected Demand 300 Safety Stock 200 100 20 18 16 14 12 10 8 6 4 2 0 0 Units 400 Net Replenishment Time Page 49 Stephen C. Graves Copyright 2010 All Rights Reserved When we have no forecast tL F ( L) z D j ft ( j ) z L j t 1 We then set the safety stock level I k0 F 2 Lk F 2 Lk 1 z Lk Lk 1 z S k 1 Tk S k to get the base-stock model ! Page 50 Stephen C. Graves Copyright 2010 All Rights Reserved