Risk-based Planning in Smart Supply Networks: The Merit of Multi-model Analytics Gerd J. Hahn German Graduate School of Management and Law, Bildungscampus 2, 74076 Heilbronn, Germany gerd.hahn@ggs.de Abstract. Extant approaches to supply network planning (SNP) are not capable of dealing with increasing requirements due to more volatile markets and the digital transformation of business. This paper proposes multi-model-based analytics approaches as a beneficial and promising avenue for further research in this field. Risk-based planning provides a suitable framework to this end. Keywords: Risk-based Planning · Stochastic Models · Multi-model Ap- proaches. 1 Introduction Supply networks aim at providing superior value to the ultimate customer by integrating business processes across the boundaries of single organizational en- tities [1]. Advanced Planning and Scheduling (APS) systems are prevalent in business practice to support decision-making concerning the design and opera- tion of such global networks [2]. For these purposes, APS systems use predictive and prescriptive analytics approaches to forecast, plan, and optimize integrated production and logistics systems. However, the business environment has changed after the 2008/2009 finan- cial crisis including persistent volatility and uncertainty of markets as well as increased competitive pressure in the course of the ongoing digital transformation [3]. This basically involves more demanding customer requirements concerning availability and customization of value offerings [4]. But high service levels, short lead-times, and product variety typically involve high costs to match supply net- work capabilities with customer demand. At the same time, supply networks become more complex and interconnected driving exposure to various risks such as supplier failure or equipment breakdown [5]. As a consequence, operating models of supply networks need to become more responsive and need to manage the adverse effects of variability proactively while keeping operations cost-efficient. Ultimately, this mandates novel approaches to risk-based decision support in this field. Due to their deterministic and simplified planning approach, APS systems are not qualified to support these aforementioned requirements. Integrating stochas- tic modeling into supply network planning (SNP) could be one way to overcome 15 these deficiencies [6]. Most recently, multi-model approaches and predictive ana- lytics are intensively discussed (or even hyped) for advanced decision support in this context [7]. This paper thus investigates the merit of multi-model-based an- alytics approaches using a risk-based planning perspective and outlines avenues for further research in this field The paper is structured as follows: section 2 provides the research background and summarizes conceptual foundations. In Sections 3 and 4, a research frame- work for multi-model SNP approaches is developed and open research topics are discussed. The paper concludes in Section 4 with a summary of the findings. 2 Research background and conceptual foundations Planning-related tasks within supply networks can be structured along the four value chain stages procurement, manufacturing, distribution, and sales and most fundamentally involve two hierarchical levels: (i) strategic design of the supply network in the long term, and (ii) mid- to short-term operations of the supply network [8] which is in focus of this paper. Predictive and prescriptive analytics provide necessary approaches for deci- sion support in supply networks. Prescriptive analytics approaches (esp. math- ematical programming) have mostly dominated the discussion around APS sys- tems [1] answering the questions ‘what shall we do?’ and ‘why shall we do it?’ [6]. In contrast, the questions ‘what will happen?’ and ‘why will it happen?’ fall within the scope of predictive analytics. Predictive analytics methods are mostly applied for time series analysis in the domain of demand forecasting [1]. The key deficit of APS systems refers to lacking decision support for manag- ing variability [6]. Variability results from random demand and stochastic events in supply network processes such as supplier/equipment failure or variation in operating times. APS systems accommodate variability indirectly by implement- ing exogenous static buffers, i.e., safety stocks and/or safety capacity, or simply adjust parameters to arrive at more conservative plans [9]. However, inadequate buffers involve substantial potential for improvement. Corresponding to the two fundamental principles of production planning and control (push vs. pull), there are two distinct approaches to manage the impli- cations of variability directly: (i) lean planning [10] or demand-driven material requirements planning [11] that follow the pull paradigm aim at avoiding the adverse effects of variability, and (ii) push-oriented stochastic planning that uses predictive and/or prescriptive analytics methods such as discrete-event simu- lation (DES) or queuing models to anticipate variability [12]. Subsequently, this paper focuses on queuing model-related approaches. For an overview of simulation-based optimization, the reader is referred to [13]. Stochastic planning approaches also allow capturing non-linear system be- havior which is mostly omitted in traditional SNP approaches [14]. This relates to the simplistic assumptions of APS systems that order lead times are inde- pendent of the capacity utilization. However, increasing the capacity utilization, i.e., releasing additional orders into the supply network, drives ‘congestion‘ in 16 the system and thus leads to longer lead times [12]. Early papers of Graves [16] and Karmarkar [17] have initiated the literature stream around clearing func- tions that aim at capturing this non-linear relationship within mathematical programming models. For these purposes, clearing functions can be derived from both theoretical queuing models or empirical shop floor data [14]. Summaries of related work can be found in [14] and [18]. Given that clearing functions need to be integrated into discrete-time (linear) mathematical programming models, this implies substantial modeling restric- tions. Multi-model approaches that combine mathematical programming with DES have thus been proposed and are widely discussed [12]. DES provides ex- tensive modeling flexibility with the drawback that numerical effort rises or even gets intractable. Keeping the concept of multi-model approaches that allows for additional modeling flexibility, this research aims at investigating alternative multi-model SNP approaches in this context. 3 Multi-model SNP approaches 3.1 Planning scope The planning problem in focus is located at the mid- to short-term level of sales and operations planning (S&OP) [19]. Detailed characteristics and trade-offs in S&OP are summarized in Figure 1. The objective function involves maximizing (economic) profit for a planning horizon of 6 to 18 months [20]. S&OP cov- ers the three perspectives of sales, operations, and finance and aims at aligning respective plans cross-functionally. These plans cover price ranges and sales vol- umes (sales), material flows, capacities, and production volumes (operations), and current assets and liabilities esp. accounts payable and receivable (finance). Characteristics of Sales & Operations Planning (S&OP) Typical trade-off discussions in S&OP meetings ▪ Objective: Maximize Economic Profit Sales: Customer Value ▪ Decision Variables: Price ranges, sales volumes; capacities, production high flexible due dates volumes, material flows; service level vs. current assets/liabilities vs. leveled low WIP/FGI1 production ▪ Focus: Customer segments, product families, supply chain segments, balance Finance: low WIP/FGI1 Operations: sheet positions Net Working vs. Utilization Capital high inventory ▪ Participants: Sales, buffers Operations, Finance 1 WIP = work in progress (inventory); FGI = Finished goods inventory Fig. 1. Key characteristics and trade-offs in S&OP 17 While the sales function aims at improving customer value, operations is interested in maximizing utilization to reduce capacity-related costs. Net working capital and the reduction of costs of capital are in focus of the finance function. Consequently, three mutual trade-offs emerge that boil down to the strategic question of efficiency vs. responsiveness [21]. Notably, the trade-offs related to the operations function involve the non-linear clearing function relationship. More specifically, leveled production allows for higher utilization at the cost of longer lead times and increased inventory buffers due to larger batches. There are two risk mitigation approaches that also foster responsiveness of supply networks: flexibility and redundancy [22]. Measures of flexibility can re- fer to the supply side (e.g. multi sourcing, flexible supply contracts), internal operations (e.g. multi-process equipment, postponement), or the demand side (responsive pricing). Responsive pricing represents the key lever at the opera- tional level. Redundancy aims at finding the ‘right’ amount of buffer with respect to safety time, capacity, and inventories. Consequently, there are two major op- erational levers to manage the responsiveness of the supply network using a risk-based planning approach: responsive pricing and buffer management. 3.2 Conceptual framework Capturing the stochastic and dynamic behavior of a supply network for plan- ning purposes would require a stochastic non-linear programming approach [15]. However, one can use hierarchical decomposition [23] to simplify matters as fol- lows: first, a deterministic linear programming (LP) model serves as the top level determining key sales, operations, and finance decisions given dynamic de- mand. An exemplary LP model for S&OP can be found in [20]. Second, an anticipated base level is implemented below that captures the stochastic and non-linear implications of the top-level decisions. Specific modeling approaches for the anticipated base level are discussed further below. Capacity levels and production volumes serve as the top-down instructions for the anticipated base level. They basically determine inflow into the supply network and thus the workload in the system. Lead times or (WIP inventories) and capacity buffers in turn represent the bottom-up reaction and are incorpo- rated as feedback at the top-level. The feedback can be implemented via ‘hard’ constraints or penalty costs in the objective function. Given that the overall ap- proach aims at perfect anticipation [23] full coordination can be reached when applying an iterative algorithm. The hierarchical planning framework is summa- rized in Figure 2. 3.3 Methodological approaches Modeling the anticipated base level involves two design decisions: scope/number of the model(s) and the analytics approaches. Both approaches of establishing one single model for the entire supply network and dividing the network into several sub-models can be found in the literature [24]. Here, the considerations for hierarchical decomposition also apply concerning the trade-off between modeling 18 Sales & Operations Planning • Aggregate planning, Top Level: Network Flow Model balancing S&OP trade-offs • Deterministic LP model Capacity levels, Lead times, Planning production volumes capacity buffers • Anticipating stochastic and non- linear behavior of supply network Anticipated Base Level: Clearing function-based Model • Mixed-/multi-model analytics approaches Our focus Order acceptance/ State of supply release decisions network Control Supply Network Fig. 2. Hierarchical planning framework (based on [24, 15]) accuracy and numerical tractability. Decomposition can also be required along the time dimension if parameters such as demand are time-dependent and the application of queuing theory requires steady state conditions [15]. Predictive and prescriptive analytics approaches or a combination of both can be applied. Predictive analytics approaches such as data mining have been proposed to estimate lead times in manufacturing systems [25]. [15] use a queu- ing networks-based approach to determine lead-time optimal batch sizes in a stochastic job shop setting (prescriptive analytics). Furthermore, one can find combined approaches: [24] use a moment-iteration algorithm to determine the workload in the system and apply a local smoothing algorithm to find a lead-time feasible schedule of planned order releases. 4 Avenues for further research 4.1 Planning scope There are three avenues for further research concerning the planning problem: first, single-model approaches have been presented for supporting order accep- tance decisions including workload considerations [26]. This could be further de- veloped towards a responsive pricing approach to capture additional value from the customer. Second, the issue of multi-stage supply networks and especially the integration of production and logistics processes has not yet been considered and thus provides opportunities for further research. Third, existing approaches for buffer management could be enhanced to ar- rive at more accurate predictions that would allow for the reduction of costly 19 6 buffers while keeping supply network risk under control. This could involve more sophisticated approaches to modeling failure of technical equipment and corre- sponding repair activities [15] or approaches for predictive maintenance. More- over, levers of manufacturing flexibility based on alternative routings in the job shop could be evaluated. [27] provide a corresponding approach for the process industries which could be extended to further domains of application. 4.2 Conceptual framework The digitalization of the industrial sector – also known as Industry 4.0 (i4.0) – envisions intelligent and interconnected ‘things’, i.e. smart products and ma- chines, that operate autonomously and that can form self-coordinating smart supply networks [3]. Literature in this field mostly focuses on issues of oper- ational scheduling and shop floor control [28]. Consequently, further research is warranted to develop novel planning frameworks and approaches that could support operational decision-making in smart supply networks. Although the smart supply network paradigm proposes fully customized man- ufacturing using orders of ‘lotsize 1’ [29], component manufactures will still pro- duce in small batch sizes for economic reasons. Given the volatile market envi- ronment, a dynamic batch planning approach could be one avenue for further research. Batch sizes need to be adjusted dynamically to support more flexible i4.0-like control approaches in supply networks. A corresponding comprehensive framework for integrated planning and control of supply networks should be developed to this end. 4.3 Methodological approaches As outlined above, evaluating further predictive analytics approaches could yield novel approaches to risk-based SNP. There are two avenues that should be fur- ther considered: first, data-driven approaches could be applied extending extant concepts of data mining. This could involve determining relevant queuing-related statistics from empirical shop floor data to derive clearing function relationships. Related approaches are currently scant in the literature [30]. The second avenue relates to making more extensive use of data via predic- tive analytics which can inform prescriptive analytics approaches. [24] provide a corresponding example for the specific context of SNP. Further research in this direction would require investigation of existing methods and development of novel approaches considering both solution quality and numerical performance. Besides, the benefit of using a multi-model approach (compared to a single-model approach) should be evaluated for both aforementioned avenues. 5 Conclusion and outlook Motivated by current developments in global supply networks due to more volatile markets and increasing customer requirements as well as shortcomings 20 of extant APS systems, this paper investigated the merit of multi-model analyt- ics approaches for risk-based SNP. For these purposes, current state of the art and avenues for further research have been identified along three lines: planning scope, conceptual framework, and methodological approaches. From a business application perspective, this research area is closely related to Industry 4.0 and the smart factory paradigm. Especially component manu- facturers that produce to order in small batches might represent an interesting industry for applications and case studies. Consequently, a design-oriented re- search approach should be pursued which could also yield managerial insights with respect to general rules for planning and decision support. References 1. Stadtler, H.: Supply chain management and advanced planning – Basics, overview and challenges. European Journal of Operational Research 163(3), pp. 575–588 (2005). 2. Lebreton, B., Meyr, H., Rosič, H., Seipl, C., Wetterauer, U.: Architecture of selected APS. In: Stadlter, H., Kilger, C., Meyr., H. (Eds.) Supply Chain Management and Advanced Planning, pp. 341–361. 5th ed. Springer, Berlin (2015). 3. McKinsey Digital: Industry 4.0. How to navigate digitization of the manufacturing sector, McKinsey & Company (2015). 4. Obermaier, R.: Industrie 4.0 als unternehmerische Gestaltungsaufgabe. Strategische und operative Handlungsfelder für Industriebetriebe. In: Obermaier, R. (Ed.) Indus- trie 4.0 als unternehmerische Gestaltungsaufgabe, pp. 3–34. Springer Fachmedien, Wiesbaden (2016). 5. Ho, W., Zheng, T., Yildiz, H., Talluri, S.: Supply chain risk management. A liter- ature review. International Journal of Production Research 53(16), pp. 5031–5069 (2015). 6. Hahn, G.J., Packowski, J.: A perspective on applications of in-memory analytics in supply chain management. Decision Support Systems 76, pp. 45–52 (2015). 7. Waller, M.A., Fawcett, S.E.: Data science, predictive analytics, and big data. A revolution that will transform supply chain design and management. Journal of Business Logistics 34(2), pp. 77–84 (2013). 8. Fleischmann, B., Meyr, H., Wagner, M.: Advanced planning. In: Stadlter, H., Kilger, C., Meyr, H. (Eds.) Supply Chain Management and Advanced Planning, pp. 71–95. 5th ed. Springer, Berlin (2015). 9. Gebhard, M., Kuhn, H.: Robuste hierarchische Produktionsplanung mit Bedarf- sszenarien. In: Otto, A., Obermaier, R. (Eds.) Logistikmanagement – Analyse, Be- wertung und Gestaltung logistischer Systeme, pp. 161–183. Dt. Univ.-Verl., Wies- baden (2007). 10. Pool, A., Wijngaard, J., van der Zee, D.-J.: Lean planning in the semi-process industry, a case study. International Journal of Production Economics 131(1), pp. 194–203 (2011). 11. Miclo, R., Lauras, M., Fontanili, F., Lamothe, J., Melnyk, S.A.: Demand driven MRP: Assessment of a new approach to materials management. International Jour- nal of Production Research 49(4), pp. 1–16 (2018). 12. Armbruster, D., Uzsoy, R.: Continuous dynamic models, clearing functions, and discrete-event simulation in aggregate production planning. In: Smith, J.C., Mir- chandani, P.B. (Eds.) 2012 Tutorials in Operations Research – New Directions 21 in Informatics, Optimization, Logistics, and Production, pp. 103–126. INFORMS, Hanover (2012). 13. Pasupathy, R., Ghosh, S.: Simulation optimization. A concise overview and imple- mentation guide. In: Smith, J.C., Topaloglu, H. (Eds.) 2013 Tutorials in Operations Research – Theory Driven by Influential Applications, pp. 122–150. INFORMS, Hanover (2013). 14. Pahl, J., Voß, S., Woodruff, D.L.: Production planning with load dependent lead times. An update of research. Annals of Operations Research 153(1), pp. 297–345 (2007). 15. Hahn, G.J., Sens, T., Decouttere, C., Vandaele, N.J.: A multi-criteria approach to robust outsourcing decision-making in stochastic manufacturing systems. Comput- ers & Industrial Engineering 98, pp. 275–288 (2016). 16. Graves, S.C.: A tactical planning model for a job shop. Operations Research 34(4), pp. 522-533 (1986). 17. Karmarkar, U.S.: Lot sizes, lead times and in-process inventories. Management Science 33(3), pp. 409–418 (1987). 18. Missbauer, H., Uzsoy, R.: Optimization models of production planning problems. In: Kempf, K.G., Keskinocak, P., Uzsoy, R. (Eds.) Planning Production and Inven- tories in the Extended Enterprise. A State of the Art Handbook, Volume 1, pp. 437–507. Springer, New York (2011). 19. Tavares Thomé, A.M., Scavarda, L.F., Fernandez, N.S., Scavarda, A.J.: Sales and operations planning. A research synthesis. International Journal of Production Eco- nomics 138(1), pp. 1–13 (2012). 20. Hahn, G.J., Kuhn, H.: Optimising a value-based performance indicator in mid- term sales and operations planning. Journal of the Operational Research Society 62(3), pp. 515–525 (2011). 21. Fisher, M.L.: What is the right supply chain for your product? Harvard Business Review 75, pp. 105-117 (1997). 22. Sodhi, M.S., Tang, C.S.: Managing Supply Chain Risk. Springer, Boston (2012). 23. Schneeweiss, C.: Distributed decision making. A unified approach. European Jour- nal of Operational Research 150(2), pp. 237–252. 24. Jansen, M.M., De Kok, T.G., Fransoo, J.C.: Lead time anticipation in supply chain operations planning. OR Spectrum 35(1), pp. 251–290 (2013). 25. Öztürk, A., Kayalıgil, S., Özdemirel, N.E.: Manufacturing lead time estimation using data mining. European Journal of Operational Research 173(2), pp. 683-700 (2006). 26. Brahimi, N., Aouam, T., Aghezzaf, E.H.: Integrating order acceptance decisions with flexible due dates in a production planning model with load-dependent lead times. International Journal of Production Research 53(12), pp. 3810–3822 (2015). 27. Hahn, G.J., Brandenburg, M.: A sustainable aggregate production planning model for the chemical process industry. Computers & Operations Research 94, pp. 154– 168 (2018). 28. Hahn, G.J.: Industry 4.0. A Supply Chain Innovation Perspective. Working Paper, GGS Heilbronn/Germany (2019). 29. Kagermann, H., Wahlster, W., Helbig, J.: Recommendations for implementing the strategic initiative INDUSTRIE 4.0. Final report of the Industrie 4.0 Work- ing Group. Forschungsunion and Acatech, Frankfurt (2013). 30. Simchi-Levi, D.: OM forum: OM research. From problem-driven to data-driven research. Manufacturing & Service Operations Management 16(1), pp. 2-10 (2013). 22