Improving accuracy of
statistical forecast for products
with high variability and slow
moving demand in Cordis
Coorporation
Preface
I would like to thank to the persons that have closely accompanied me during my studies and for all the support received from them. This work is especially dedicated to Karina, who has unconditionally supported me along several months of study and research. To my parents, who always taught me the love for work and encouraged me to climb high and to my aunt Tina and my uncle Adan, because they also gave support to become this dream true. I also thank to Rolph de Groot for his valuable time and supervision in the company and to Dr. Sebastiaan Brinkman who shared with me his valuable knowledge and experience to contribute with this research.
The next research is the result of nearly 6 months of investigation in the European headquarters of Cordis Corporation. During the past months SC management has been striving to find ways to improve forecast accuracy for products that present low forecast performance. The forecast accuracy is one of the main KPI’s in SC. However, causes of forecast inaccuracy were not fully understood creating a gap for solving the issue. When it is large enough the forecast error creates back orders for some products, reducing customer service levels. In the other hand, over forecast may increase holding costs, especially for slow moving products. These issues reduced the operational performance along supply chain.
The main objective of this research was to understand the causes existing behind the forecast error and increase the accuracy of statistical forecast by deeper insights in time series existing for products sold by Cordis. The company produces nearly 5500 medical devices specialized in the cardiovascular and endovascular treatment, these products are distributed in 45 countries. The demand in each market may differ strongly from one product to another; therefore
classification of demand patterns is critical to understand variation existing in the demand. Furthermore, it is also well known that forecast methods should be selected based on time series characteristics
Afterwards, we classified demand and centered this research in those items with lumpy and infrequent demand. We tested alternative forecasting methods and the current forecasting method in order to measure which one provides best accuracy. Furthermore, different forecasting scenarios were created by setting the parameters at different levels for each method.
The accuracy between forecast models was compared using confidence intervals at 95% and models with smaller errors and dispersion were proposed as best models for each category of demand. Finally, the parameter’s contributions to forecast error were analyzed using DOE (Design of experiments) approach. This analysis allowed optimizing the parameter setting for each model. The result was the formulation of an integral model that enabled to select best forecast model depending on variability observed in demand (for slow moving and lumpy demand), improving forecast accuracy (in sample).
TABLE OF CONTENTS
GLOSSARY OF TERMS………...5
CHAPTER 1. INTRODUCTION………7
1.1 The company………7
1.2 The products and the markets………..7
1.3 The SC department………..10
1.4 The SC network……….10
1.5 Sales Inventory and Operations Management (SIOP) & JNJ month………11
1.6 The forecasting process: Demand SIOP………...12
1.7 Problem outline……….…13
CHAPTER 2. PROBLEM EXPLORATION & THEORETICAL FRAMEWORK……….15
2.1 Managerial question and Voice of customer……….15
2.2 Why forecast? A theoretical framework……….15
2.3 Planning horizons and forecast impact………..16
2.3.1 The forecast at operational level……….16
2.3.2 The forecast at tactical level………17
2.3.3 The forecast at strategic level……….17
2.4 Improving the forecast accuracy (the theory)………...17
2.4.1 Bottom up or top down approach ………..17
2.4.2 Quality of input data………..18
2.4.2.1 Outliers…..……….18
2.4.2.2 Marketing research & internal factors………18
2.4.2.3 Leading indicators & external factors………18
2.4.3 Combining techniques………..18
2.4.4 Adding expert judgment………19
2.4.5 Training forecasters and implementing forecast packages………19
2.4.6 Applying more sophisticated methods………...19
2.4.7 Collaborative forecasting………..19
2.4.8 Product categories………20
CHAPTER 3. PROBLEM ANALYSIS………...21
3.1 Causes of forecast inaccuracy………21
3.2 Problem statement………23
3.3 Research objectives……….23
3.4 Preliminary research questions………..23
3.5 Research question………23
3.6 Conceptual model……….24
3.7 DMAI²C………...25
3.9 Scope………..27
CHAPTER 4. ANALYSIS OF CURRENT SITUATION………..29
4.1 Planning hierarchies in SIOP process………...29
4.2 Forecast process ……….……….31
4.2.1 Retrieval of historical information and parameters adjustment…………..31
4.2.2 Forecast calculation………..33
4.2.3 Review and adjust statistical forecast at Control Level………...34
4.2.4 Incorporate regional information……….36
4.2.5 Reconciliation process (Statistical forecast at DFU level)………..36
4.2.6 Publish final version of forecast………..36
4.3 Forecast error………37
4.3.1 PE and APE performance from top-down perspective………38
4.3.2 PE and APE performance from bottom-up perspective………..39
4.4 Demand variability……….40
4.4.1 Volume variability………..40
4.4.2 Frequency variability……….40
4.4.3 Data collection………41
4.4.4 Categories of demand………..………41
CHAPTER 5. FORECASTING AND IMPROVEMENT MODEL………...45
5.1 Forecast classification………..45
5.2 Statistical forecasting………46
5.2.1 Time series methods……….46
5.2.1.1 Moving averages………..46
5.2.1.2 Weighted moving average………..46
5.2.1.3 Single Exponential smoothing………46
5.2.1.4 Simple regression……….47
5.2.1.5 Brown (double exponential smoothing)………47
5.2.1.6 Croston………..48
5.2.1.7 Holt winters………48
5.2.1.8 Fourier………48
5.2.1.9 Lewandowski (FORSYS)……….49
5.2.1.10 Box Jenkins………..50
5.3 Traditional forecasting approach vs forecast per categories……….51
5.4 Focus on categories model……….51
CHAPTER 6. FORECASTING SCENARIOS / PERFORMANCE EVALUATION AND MODEL SELECTION………..53
6.1 Sample size definition………..53
6.2 Selecting alternative forecast methods………..54
6.2.2 Running forecasting scenarios………57
6.2.3 Measuring accuracy error………57
6.2.4 Comparing forecast errors………...58
6.3 Forecasting scenarios with Lewandowski method………..59
6.3.1 Selecting parameters for Lewandowski method………..59
6.3.2 Running Lewandowski scenarios………...60
6.4 Results & Model selection………...61
6.5 Forecasting very slow moving items……….….…63
6.6 Testing significance of parameters in Lewandowski models……….63
6.7 New model for forecast selection………...65
CHAPTER 7. GUIDELINE FOR FUTURE IMPLEMENTATION………...………..67
7.1 Implementation model………..67
7.2 Operational performance improvement……….………70
CHAPTER 8. CONCLUSIONS ……….73
8.1 Conclusions, limitations of this study and further research………73
GLOSSARY OF TERMS
The following terms and acronyms were used in this research:
Absolute Percentage Error (APE): Measures the relative error of absolute forecast deviation. Advanced Planning System (APS): Information systems used to support the planning process. Control Level (CL): Parent‐child model used to create forecasts of demand.
Demand Forecasting Units (DFU): Relationship used in the planning system to specify item selling in
market ‐i.e. country of demand‐. The DFU’s describe demand for SKU’s.
DFU MAP: Model that defines parent‐child relations used in the planning process.
FCST_DRAFT: The forecast resulting from the runs of statistical model previous to the final forecast
version.
Manugistics: APS used in Cordis to support planning process and the forecasting process.
SC: Acronym used for the word Supply Chain and it refers to the supply network but also to the
department.
Seasonality: Is the resulting pattern for demand that is influenced by seasonal factors occurring in
CHAPTER 1 INTRODUCTION 1.1 The company: Cordis group1
For 50 years, Cordis Corporation, a Johnson & Johnson company (JNJ), has pioneered less invasive treatments for vascular disease. Technological innovation and a deep understanding of the medical marketplace and the needs of patients have made Cordis the world's leading
developer and manufacturer of breakthrough stents, catheters and guidewires for interventional medicine, minimally invasive computer-based imaging, and electrophysiology.
Approximately 5,000 employees worldwide share a strong commitment to continue the Company's groundbreaking work in the fight against vascular disease. Cordis business are:
1. Cordis Cardiology for cardiovascular disease management;
2. Cordis Endovascular for the treatment of peripheral vascular and biliary obstructive diseases;
3. Conor Medsystems LLC for controlled vascular drug delivery technologies; 4. Biosense Webster, Inc. for electrophysiology and medical sensor technology in
cardiovascular procedures; and
5. Cordis Biologics Delivery Systems Group in the emerging field of biologics delivery. Business is also divided into 5 main affiliates. Main purpose of affiliates is to provide all the customers with a service of excellence in the different markets around the world. The affiliates conforming Cordis group are:
1. Cordis Corporation - Europe 2. Cordis Corporation - Canada 3. Cordis Corporation - Japan 4. Cordis Corporation - Asian Pacific
5. Cordis Corporation - Latin America 1.2 The products and the markets
The product portfolio existing at Cordis Europe is constituted by three main families of products. This classification is based primarily in functionality of the products.
• Cardiology products: As its name explains, these products are designed for interventional therapies in coronary diseases. Include pre-dilatation catheters, post-dilatation catheters and drug-eluting stents.
.
Figure 1. Drug eluting stent
• Accessories: These products are designed for diagnostic of coronary diseases. The family includes diagnostic catheters, guiding catheters (guide wires) and sheath introducers.
Figure 2. Diagnostic catheter • Endovascular products: This product family includes
minimally invasive endovascular stents, catheters, guidewires, and accessories that help physicians provide alternatives to traditional surgery in the diagnosis and treatment of vascular conditions.
Figure 3. Balloon dilatation catheter In total Cordis produces N different products (items) grouped as follows. The next graph shows diversity per group:
Figure 4. Product diversity per group
Figure 5. Turnover per family of products2
As explained before, Cordis produces, distributes and sells approximately 5417 different
products. These products are sold in 47 countries (excluding USA for OUS market). The size of market and the turnover generated varies from country to country. The next chart depicts the estimated turnover per country or cardiology, accessories and endovascular products for the last 6 months:
Figure 6. Turnover per country
The previous chart is very important because gives us a rough idea about different needs and demand volumes between countries. Typically, demand from EMEA (Europa Midle East and
Africa) countries and Canada is bigger than demand from countries located in other regions. A large part of this study will be a deep analysis of the demand patterns for product sold in specific countries; therefore, it is important to understand (as initial point) that markets are very different between themselves.
1.3 The supply chain department
Within Cordis Corporation the Supply Chain group is responsible for planning or coordinating the different components of the supply chain. Its main goal is to provide customers with the right product in the right time with the minimum cost.
Cordis supply Chain in Europe is known as supply chain OUS (Outside United States). The core processes inside of supply chain OUS are: the SIOP process (Sales Inventory and Operations management), life cycle management, inventory management and order fulfillment. All these activities have as main objective to balance demand and supply.
The supply chain department also cooperates and coordinates the introduction of new products and ensures that these products become globally available for the customers. Supply chain activities for new projects are leaded and coordinated from supply chain OUS.
Another core activity is Inventory management, which ensures to keep optimal levels of inventory along the SC. This is to have the right inventory in order to cope with demand fluctuations, deployment functions are closely related to this activity. Finally, but not least important, order fulfillment takes care of product flow from production plants to end customers.
1.4 The supply chain network
From network perspective, global supply chain is organized in two groups, SC-US (United States) and SC-OUS, each group is dedicated to serve the demand in specific markets. US supply chain is mainly focused to serve demand from USA but it also fulfills part of the demand for some countries located in Asia Pacific, Latin America or Canada. SC-OUS is responsible for planning the demand in the rest of the world with exception of USA. In total, SC-OUS is fulfilling demand of 47 countries. The worldwide demand is split into 5 principal regions; these regions are APAC (Asia and Pacific), Canada, EMEA (Europe Middle East and Africa), Japan, and LATAM (Latin America).
The products are manufactured in 3 different locations; the main production facility is located in Cd. Juarez, in Mexico. This facility is currently manufacturing the great majority of the products. A second manufacturing facility is located in Puerto Rico which is exclusively dedicated to manufacturing of Star products. This family of successful products includes fast moving items that have demand in many countries. Finally, a third manufacturing source corresponds to OEMs (Original Equipment Manufacturers) which are dedicated to produce standard products within the product portfolio.
Non-DRP countries and direct customers. The supply chain network at high level is depicted in the next figure:
Figure 7. Supply chain network
1.5 Sales Inventory and Operations Management (SIOP) & JNJ month
SIOP process is the core processes for planning at Cordis supply chain. The complete cycle of SIOP process is fulfilled monthly basis3 and is split into three main steps. First step is referred
as Demand SIOP, where statistical forecasts for countries demand are generated based on historical demand. Here historical information and some qualitative inputs are considered for the creation of statistical forecast). These activities are carried during first week of JNJ planning month. Demand SIOP step will be described in more detail in the next section.
The second step of SIOP process is referred as SIOP supply. In this part of the process, forecasted demand is translated firstly into an unconstrained plan and afterwards, capacity and other constrains are taken into consideration in order to create the constrained plan. This last plan is used to set the material planning and the production schedules for the manufacturing facilities. The last phase in SIOP supply is known as deployment process.
The third step is called executive SIOP process. This step of the SIOP is related to decision making concerning the outputs of demand and supply SIOP processes. Decision making is done considering all the factors intervening in planning process and available information in order to allocate the resources in such a way that serve better to customer and minimizing costs involved. The figure in next page shows steps of SIOP process:
Figure 8. SIOP process4
The forecast is created and published as part of the first process; therefore this research will be focused on the first step of SIOP process.
1.6 The forecasting process: Demand SIOP
Future demand for Cordis products is estimated based on forecast (forecast). This forecast is calculated based on historical demand and the final version of the published each end of JNJ month. The forecast is created in lag 2, this means that forecast is published two months in advance from the period of demand. The main reason for creating the forecast in lag two is because this time corresponds to average lead time required for production of Cordis products. At the beginning of each JNJ month, the planner creates the forecast at certain level of
aggregation based on historical demand; this first version of the forecast is known as
FCST_DRAFT. The final forecast corresponds to the last version modified of the forecast which is published and used for the MPS.
The main steps in the forecasting process are: Retrieval of historical data, creation of statistical forecast, revision and adjustments, incorporate regional information, reconciliation and
publication of final forecast. The next flow chart depicts these steps.
Figure 9. The forecasting process
The forecasting process will be explained in more detail in chapter 4.
1.7 Problem outline
In Cordis, the forecast performance is measured with two indicators. The Net Deviation and the MAPE, both metrics are used to calculate the percent of forecast error. The forecast
performance is reported and reviewed monthly basis. This report rolls out the forecast performance for aggregated information and pinpoints the areas for improvement.
Historically, SC team has observed that forecast error tends to be greater in some specific product (especially for slow movers). The error is reducing the forecast performance for those families with products of erratic or infrequent demand. Another important consequence of forecast error is the potential risk of back orders when demand is under forecasted considerably. On the other hand, increased holding costs occur when over forecasting is continuous.
CHAPTER 2
PROBLEM EXPLORATION AND THEORETICAL FRAMEWORK 2.1 Managerial question and Voice of customer
Due to practical nature of the problem, specific managerial question should be outlined based on information gathered from main stakeholders in the company. The managerial question can be outlined as follows:
“Historically, some products (especially slow movers) present greater forecast error. Forecast error is later translated into backorders or excessive inventories for these items. This project has the objective to classify the products based on demand characteristics and suggest the forecast model that better fits to these characteristics in order to reduce current MAPE values”.
Previous problem description, information gathered by interviewing stakeholders, and problem exploration leaded us to the conclusion that main objective of this study will be to tackle the forecast error for products with historical higher forecast error.
2.2 Why forecasting? A theoretical framework
Makridakis, S. et al (1998) and Schonsleben, P. (2007) state that main reason to forecast the demand is because often lead time exceeds customer tolerance time. Additionally, Sanders, N. (1995) and Makridakis, S. (1998) state that forecasts are necessary for effective planning and decision making. In the medical sector, the products fulfill a need from patients, therefore, in this market products are always expected to be available. The truth is that when product is not available, customer chooses an equivalent product from the competence. This situation results in opportunity costs and negative perception of the brand from customers. As observed in chapter 1, the SC network of Cordis has different types of customers. Each customer represents a different need but also a different source of demand (uncertainty). Main customers are the affiliates and the direct customers. Typically (as the name implies), direct customers are supplied directly from DC’s. In contrast, countries with affiliates fulfill the demand of hospitals and distributors trough regional warehouses. The fact that customers are different implies different levels of response. For example, response for orders supplied to direct customers (such as hospitals located within a reasonable distance from the DC) is different than orders coming from regional warehouses. Furthermore, each customer (hospital in X country) has specific needs, the way to deal with this uncertainty in terms of mix and volume is by forecasting the demand. Another important point mentioned in Simchi-Levi, D. et al (2008) is the amount of integration between SC entities. For example, affiliates are sharing common IT systems with the SC department, improving visibility of demand, but when this is not the case, demand needs to be estimated by forecast.
Besides demand anticipation, effective forecasting can improve competitiveness and
impact different levels in the SC where error tend to be amplified (i.e. bullwhip effect) resulting in poor service levels or excess of inventories.
Another important point to consider is that Cordis is currently performing different projects to improve the operational performance along the supply chain; however, a project focused in forecast accuracy improvement is still missing. It is in this unexplored area, where we see a clear opportunity for further research5.
2.3 Planning horizons and forecast impact
The forecast fulfill important functions within the company. These functions can be classified according heir time scope. The forecast supports decisions in the short, medium and long term, the duration also defines the impact that has across the company. According to APICS CPIM (2006) planning horizons are divided in short, mid and long term. The planning horizon is related to the impact of the decisions associated with the forecast. In Cordis, the forecast influences operational, tactical and strategic decisions. Next picture depicts the planning horizons in the forecasting process.
Figure 10. Planning horizons in forecasting process6
2.3.1 The forecast at operational level
In the short term the forecast is used as an input for deployment function (DRP). The DRP process generates the unconstrained plan taking the forecast and other important inputs such as customer orders, inventories, safety stock, lead time and product in transit. The
unconstrained plan is used later to generate MPS, this plan considers daily production rates and other inputs. When MPS is ready and published it act as trigger for production at the
manufacturing plants.
5 Another SC projects focused to increase the operational performance in Cordis are: SKU rationalization, SS model
upgrade, lot splitting and lot size optimization, postponement, global sourcing and deployment organization.
2.3.2 The forecast at tactical level
At tactical level, forecast numbers are used to support decisions for materials requirement (MRP). Usually, components need to be planned within sort-medium term in advance because lead time of components exceed the expected customer lead time. The forecast allow the suppliers to know in advance what to produce in the next months. In some cases (when demand rise up), suppliers are required to reserve additional capacity to supply the demand, therefore a good forecast can predict with reasonable accuracy this kind of events . The forecast also provides good visibility the capacity that will be required at certain point, this characteristic is important to level the load more evenly in the production plants.
2.3.3 The forecast at strategic level
Strategic decisions have always important financial consequences for the organizations. Examples of these kinds of decision are: capacity management, resources planning and
capacity management for future launch of products (i.e. launch of a product that will share same production line with older products). Usually the horizon of these decisions is from 6 months or longer (1 or two years). Although it is well known that forecasts are less accurate the longer the horizon is predicted (APICS CPIM, 2006, Makridakis, S. 1998 and O’Connor, M. 2000),
companies still need to have visibility in order to plan future events (Sanders, N. 1995). From these chapters we can also determine main business processes in Cordis which associated to the forecast. Creation of constrained and unconstrained plans for demand, creation of MPS, creation of MRP, capacity planning and investments as part of strategic decisions.
2.4 Improving the forecast accuracy
In previous paragraph we explain the importance of accurate forecast and its impact at different levels. In the following paragraphs we explore the alternatives to improve the forecast accuracy based on theory. For this purpose, a broad range of literature was reviewed.
2.4.1 Bottom up or top down approach
In Geurts, M. et al (1999) it is mentioned that one of the ways to improve forecast accuracy is by understanding the causes of variance in the sales pattern. This can be done by forecasting the components instead forecasting the aggregated or “pooled” time series. By disaggregating the time series, the variance generators (such as trend, cycles, seasonality, atypical events, marketing effects and noise) are visualized and their effects over the aggregated forecast are understood more clearly. For example, one national forecast may be more accurate when it is the result of adding regional individual forecast rather than produce an aggregated forecast. Another approach mentioned for the authors consist into analyze the demand in shorter periods rather than large an aggregated periods. For example, monthly forecast should be more
accurate than quarterly forecasts. A similar study conducted by Gordon, T. Morris, J. and
2.4.2 Quality of input data
Several authors also mention the importance of quality data used to create the forecasts. This information can be classified within next groups.
2.4.2.1 Outliers
According Geurts, M. and Whitlark (1996) besides to traditional elements of any time series (seasonal, cycles, etc.) the outliers may be part of the time series. According their study, outliers may be wrongly interpreted as other component in the time series (i.e trend). Therefore, they recommend to replace them with simple techniques, for instance using a smoothing techniques or calculating the average value before running the model. Makridakis, S. et al. (1998) mentions some effects of outliers in model fitting, the author recommend to remove them for time series decomposition and mention the limitations of the linear models for these observations.
According to CPIM (APICS, 2000) an outlier is a data point that significantly differ from other data describing same phenomenon and that is unlikely to be repeated any time in the future. Therefore, it is considered important pinpoint these points and determine the root cause of them. Sometimes removal of outlier is preferred instead adjusting forecasting model, although this approach is easier, removed outliers and the circumstances associated with them should be recorded, otherwise capability to anticipate similar events will be lost.
2.4.2.2 Marketing research & internal factors
In Geurts, M. (1999) it is mentioned that another approach to improve accuracy consists to use the marketing research information as input for the base forecast. With this information
forecaster can re-calibrate and modify the sales forecast. Other internal factors that may impact sales are: Planed changes in prices, changes in sales force, advertising, therefore considering them can be positive in terms of accuracy.
2.4.2.3 Leading indicators & external factors
In econometric (or explanatory) forecasting, leading indicators are use as inputs to create a forecasting function output (Makridakis, S. 1998). APICS mentions several leading indicators (external factors). Typical factors such as competition information, customers, plans of mayor customers, government policies, regulatory concerns, economic conditions, environmental issues, weather conditions and market trends may influence demand. Typically none of these factors is controlled by the company.
2.4.3 Combining techniques
In the large study presented in Makridakis, et al (1982) several methods for time series were tested and authors concluded that accuracy is clearly improved especially when combining specific methods. This accuracy is also better when comparing accuracy of single methods used in the combination.
In their study M. Hibon, T. Evgeniou (2005) conclude after testing several hypothesis that combination is not always more accurate than applying individual methods, especially when individual method is the best fitting for an specific time series. In contrast, combination of
2.4.4 Adding expert judgment
Lawrence, M. Goodwin, P. O’Connor, M. and Onkal, D. (2006) mention the factors intervening in judgmental adjustments: feedback, decomposition, forecast combinations, judgmental
adjustments and technical knowledge of forecasters. They conclude that adjustments may positive under certain circumstances but undesirable in others situations. For example, if the forecaster makes an adjustment to omit one variable in the forecast model, adjustment acts in detriment of the model when same variable results to be collinear with any variable included in the model. Authors also mention that training interacts with way that data is presented. For example, if data presented in tabular way is combined with technical training, judgmental adjustments improve accuracy always assuming that all time series data available.
In contranst, R. Fildes et al. (2009) mention that judgmental adjustment is a complex task that requires the existence of several assumptions such as, previous accurate statistical forecast, collaboration between supplier and retailers and forecasters well trained. Otherwise, effects of adjustment may become undesirable. However, it is also mentioned that judgmental
adjustments are in practice the most common way for incorporate key drivers into disaggregated forecast.
2.4.5 Training forecasters and implementing forecasting packages
One of the most overrated solutions to overcome forecasting problems in training of planners (Makridakis, 1999). As author explains, training should not be only focused on application sophisticated methods but in how to improve performance of models used to forecast the demand. Such as selection of time horizons, how to add judgment, deal with outliers or selecting the level of aggregation to create the forecasts. A different solution that also implies training of the personal is the implementation of forecasting package. Here organizations must to check important points before proceeding with this step. For example, methods available in the package, platform for data, how easy is to learn, possibility to implement new methods, data set capacity, local support and associated costs.
2.4.6 Applying more sophisticated methods
Here authors differ in how effective results the application of more sophisticated methods for sales demand forecasting. One of the first authors to compare accuracy between different forecast methods was Armstrong (1978), in his survey he concluded that there is no evidence to conclude that econometric (explanatory) methods are more accurate than time series methods; furthermore, complex econometric methods did not do better than simple ones. In Makridakis (1999), the authors summarize the study of accuracy comparing between several methods from famous M-Competition study. Concluding that there is no evidence to conclude that post sample accuracy of complex methods is better that accuracy for simple methods. Weatherford, L.R. and Kimes, S.E. (2003) also compared different methods and found that simple methods perform better than complex ones under very specific conditions but performance is not the best across different scenarios.
2.4.7 Collaborative forecasting
information exchange between retailers, DC’s and manufactures. One enabler for information exchange is the utilization of same information systems and forecasting tools.
2.4.8 Product categories
Another effective approach to improve the quality of forecasts explained in Jain, L. (1995) is to divide the products in categories and determine the best model for each category. In some cases, one forecast is generated at top category level; however, the assumption here is to forecast products which demand behaves similarly, this is, if demand of one product goes up the rest of the products in the category need to go up as well. Unfortunately, this is not always the case, for this reason some companies prefer to make their forecast separately. Product categorization also include approach variation such as level of promotion of products (highly and least promoted) or type of customer involved (large and small).
CHAPTER 3 Problem description 3.1 Causes for forecast inaccuracy
Before stating the problem and setting up the objectives for this research the causes of forecast inaccuracy mentioned in literature will be briefly described. According CPIM (APICS, 2006) forecast accuracy can be reduced by:
• Inappropriate forecast method: Many factors should be considered before selection of forecasting method. In time series forecasting, selecting a model that fits data pattern is crucial.
• Lack of participation: All functional areas involved in the forecasting system should participate during the different stages in the process in order to avoid misalignment. • To difficult to understand: The more the complex the method is, the more isolated the
users are from the process. If techniques are difficult to understand it is more probable that method is not fully exploited and improved.
• Lack of compatibility: Between forecast system and organization capabilities. If technique is not trusted decision rely more heavily in other sources of information or feeling.
• Data may be inaccurate: Collection and classification errors. Time and measurement errors do not only reduce reliability of data but also implies time consumption within the forecasting process.
• Some data are inappropriate: Some items should not be forecasted, especially dependant demand for components.
• Lack of monitoring: Occur when companies do not check forecast performance and therefore improvement targets are missing.
During problem exploration, all possible causes for inaccuracy mentioned above were analyzed for Cordis situation. The fishbone diagram (Ishikawa) is a much extended tool used to determine root causes for any problem. Potential causes of forecasting inaccuracy were listed in according standard categories of the diagram.
Environment Measurements Methods Material Machines Personnel Lack_of_participation Inaccurate_data Inappropriate_data Lack_of_compatibility d Difficult_to_understan ethod Innappropriate_fcst_m Lack_of_monitoring Cause-effect diagram inaccuracy Forecast
Figure 11. Cause effect diagram of forecasting inaccuracy
Currently, forecast performance is tracked and reviewed monthly basis, this report enables to pinpoint areas of attention. Furthermore, stakeholders are aware about products families with reduced accuracy, therefore, this point is discarded as root cause. Inappropriate data is referred to the fact that not all components need to be forecasted (such as dependent demand), in this case, SC-OUS is forecasting demand of finished products, and these numbers are later used as to create MRP requirements, therefore, this cause is discarded as well. In regards accuracy of data, we can mention that worldwide demand is gathered and stored in a shared IS platform which is updated continuously on the basis of demand realized at DC’s and regional affiliates, this system enables availability of timely and reliable data. Furthermore, managers and
forecasters participates in meetings on a weekly basis with the purpose of analyze and review forecasting-supply issues. These meetings serve to discuss and solve the impact of issues generated by large forecast errors.
reliability and flexibility in the model. In the products where error is historically large, the statistical forecast is considered just a reference because planners do not trust on it.
Under the light of current situation and based on previous research existing, this research will tackle the forecast accuracy issue combining different approaches mentioned in chapter 2. The first approach will be the bottom up analysis of demand in order to understand sources of variation and the second approach will be the categorization of demand in order to select a forecast method that better fits with demand characteristics.
Research will be structured in this order, problem definition, data gathering, bottom up analysis, demand categorization, forecasting scenarios and model testing, forecast error comparison, selection of model per categories, roadmap for implementation, conclusions and limitations and room for future improvements.
3.2 Problem statement
Currently, same forecast model is applied across all DFU’s existing in Cordis, however, demand characteristics between DFU’s are different. It has been observed that forecast inaccuracy has been historically larger for some products. When large enough the forecast error causes backorders impacting the service level or considerably increase holding cost from excess of inventories. Furthermore, there is lack of deeper understanding about causes of inaccuracy and which approaches can be applied in current situation in order to improve forecast accuracy.
3.3 Research objectives
Understand the causes of forecast error and the sources of demand variation (at disaggregated level). Classify the demand in a proper way that reflects patterns variation. Propose and test different a forecast models (methods and parameters) including Lewandowski method focusing on those DFU’s with historical large errors. Create a framework that allows to the company to understand interaction between demand classifications and forecast method capabilities aiming to an effective reduction of current forecast error. The final step will be to outline an algorithm form forecast model selection and make recommendations for future implementation.
3.4 Preliminary research questions
• Which factors are interacting with forecasting accuracy?
• How to measure demand variability for product DFU’s? – data mining strategy • How to classify demand variability?
• Which is the relation between forecast error and demand characteristics? • Which DFU’s will be part of scope?
• Which demand patterns exists according demand classification? • Which forecast methods fit to these patterns (based on literature)?
• Does current forecast model can be improved by parameters modification?
3.5 Research question
The research question should reflect the main objectives of the project. Therefore, research question is stated as follows:
3.6 Conceptual model
The conceptual model includes the variables interacting in the forecast process. It shows in a simple way process inputs and outputs. The model also shows the information levels identified in the process. The first level corresponds to the Control Level which is the level where planner creates and adjusts the forecast of demand. Here we find 4 inputs used to adjust the forecast. The planner review and consider the inputs and his or her expertise in order to adjust the model parameters. Although these adjustments are based on quantitative information, adjustment by itself is judgmental, because it depends on experience and perception (weighting) that planner has from events. In the DFU-level the disaggregated forecasts are calculated. These forecasts are produced in the reconciliation process (explained in chapter 4). Finally, the last level
corresponds to the demand level. Here demand realized (also known as real) is recorded at the end of each week and used to update demand records and accuracy reports. We used SIPOC methodology in order to identify inputs and outputs of the process (see Appendix 1 for further details). Next figure shows conceptual model for this research:
Demand cathegorization
Figure 12. Conceptual model of forecasting process
The conceptual model also shows the scope for this project (highlighted processes) . Currently, planners are adjusting parameters and trying to improve accuracy at control level without looking at DFU level, however, the weakness in the current approach is that differences
in chapter 4 with more detail). The previous figure, shows a blue dotted line that connects the historical data with DFU’s forecast model, this link (which is currently missing) will be the basis for analysis and categorization of demand for selection of forecasting model.
3.7 DMAI²C (Define, Measure, Analyze, Innovate and Improve, Control)
DMAI²C is a 6 sigma approach used in JNJ as guideline for process improvement. It consists in 5 steps that form a cycle for continuous improvement. This methodology was chosen as
research methodology because of the next reasons:
1. The approach and tools used are known by company stakeholders
2. DMAI²C tools are based on statistical and quality concepts, therefore they have validity. 3. The DMAI²C steps allow accomplishing the research objectives.
In JNJ methodology is known as DMAI²C process because it also includes Innovation as part of improvement (fourth step), therefore the superscript “2” is located after letter I. Next picture shows the process.
Figure 13. DMAIC process7
3.8 Research model
According to Schönsleben (2007) the forecasting technique should be selected based on actual consumption, previous demand figures and comparison of actual planning resources and possibilities in the company. Furthermore, forecast selection is the main objective in forecast management procedure. Therefore, we included part of the forecasting management model found in Schönsleben (2007) in the fourth step of our research design. The next figure depicts the resulting model:
Not part of the scope
Figure 14. Research model
The grey arrow (Control) is beyond the scope of this project, this means that implementation of the forecasting model will not be part of the scope. In this chapter, we have covered the first step of DMAI²C process. In chapter 4, variables will be measured as part of analysis of current situation. In chapter 5 and 6 analysis of demand patterns and forecasting methods for time series will be done. The Innovate & Improve stage will be related to the main goal of this project that is the forecast method selection according to demand pattern characteristics and
3.9 Scope
The product families selected for this research will be the X and Y products8. The main goal of
this research will be to provide and effective framework reduce forecast error in those DFU’s with historical higher forecast error (slow movers), which are considered as more difficult to forecast. As for life cycle perspective, this research will only include mature products, excluding new or end cycle life products, this because time series for these products is not suitable for statistical approaches analyzed in this research. The demand categorization will be done for all DFU’s, however, due to resources and time limitations, forecasts scenario testing will be done base on sampling and for a limited number or models.
Furthermore, since forecast accuracy of new scenarios won’t be evaluated for upcoming demand, accuracy will be measured against historical forecast performance.
The final step of this project is to deliver comprehensive forecast algorithm for model selection according demand categorization that will work as guideline for future implementation and upgrade of forecasting system, however, implementation will not be part of the scope.
8 The reason to exclude cardiology products is because Star products (which represent 80% of the turnover for
CHAPTER 4
ANALYSIS OF CURRENT SITUATION 4. 1 Planning hierarchies in SIOP process
Before explaining forecasting process in detail, it is useful to define the meaning and function of the planning hierarchies within the demand SIOP process. In the planning system, the
combination form by one item code and country name is known as DFU. The DFU is the demand unit used in the planning process and corresponds to the most detailed level of information of demand available in the APS system.
As mentioned before, within the SIOP process a major component is to correctly forecasting the sales in the different countries and regions where Cordis does Business and translating this back to MPS’s triggering production in the plants. SC-OUS plans the demand for more than 31,000 DFU’s. For this reason, the forecasting process is a task that requires an especial organization. The solution for this was the creation of control level (CL). The CL’s make the forecasting process a manageable activity. As explained in Flides, R. (2009), it results practically impossible for planners to give individual attention to each SKU, therefore, aggregation of information is required is required.
The CL is defined as a combination of one product hierarchy (P) and one location hierarchy (L). Product hierarchies are designated with names P0, P1, P2, P3, P4 and P5, while location hierarchies with names L0, L1, L2, L3 and L4. The P0 and L0 hierarchies correspond to the most detailed level of information or level of disaggregation. Higher numbers contain information aggregated from smaller hierarchies. For example, a P1 hierarchy contains several P0’s or item codes. A P2 hierarchy may contain several P1’s and therefore several P0’s. The same occurs for location hierarchies. For example, L2 hierarchy specifies region of demand (APAC, EMEA, etc.), while L0 shows the country of demand. One hierarchy L1 contains different L0’s. Planning hierarchies are depicted in next table:
Table 1. Planning hierarchies
Product level Location level
P5 (Family) - P4 L4 (Worldwide) P3 L3 (US/OUS) P2 L2 (Region) P1 L1 (Area) P0 (Item) L0 (Country)
Confidential
Figure 15. Hierarchies for control levels in planning process9
The CL is the aggregation of information used to create the forecasts. The planner selects the CL based on a table. The CL’s are also related to the amount of visibility (or resolution) of sources of demand. For example, a forecast generated with a control level of P3-L2 has more aggregated information therefore; visibility of disaggregated unit of demand is less when compared with a forecast created at P1-L1 level. There is a trade for the aggregation level of one forecast. Table 2 shows pros and cons of level of aggregation. As mentioned in chapter 2, in pooled time series it is difficult to understand the sources of variation. For this reason, organizations cannot look for further improvements in their forecasting process. In the other side, it results impossible to produce separately the forecast for each item code; therefore disaggregation/aggregation level is set depending on available resources. In the case of Cordis, CL’s are set at P3_L2 level for X products and P1_L2 level for Y products. The CL’s were defined based on cycle life considerations and resources available in the SC department. Next table shows guidelines used in Cordis for CL’s selection:
Table 2. Criteria for CL’s selection
Aggregation level When Pros Cons
Less aggregation desired
• Families of items recently introduced • Families of items with
changeable
conditions in demand (i.e. country variation, erratic replenishment)
• More resolution of information (planner can track better important changes in demand for specific DFU’s)
• Forecast as DFU level is in some way more reactive to changes done at control level (might become also a con).
More aggregation desired
• Families with mature items
• Families with items with stable demand and non-changeable conditions
• Less workload for planner
• Forecasting process less complex as result of standardized CL’s
• Results more difficult for planner to track important changes in demand for specific DFU’s
• Forecast at DFU level is less reactive to changes made at CL (might become also a pro).
Control levels are usually not modified in the system; however, SC management defines them based in considerations presented in previous table. This project seeks to increase accuracy at disaggregated level by setting appropriate forecast models, in order to avoid the need of forecasting at different control levels. If good forecast are produced at forecast level, less need to work at disaggregated levels, this point results very important from complexity perspective, because having less controlling levels, forecasting activity results less complex and
manageable.
4.2 The forecasting process
The forecasting process is divided in 6 main steps. These steps are described in the next paragraphs.
4.2.1 Retrieval of historical information and parameters adjustment
This is the first stage in the forecasting process and it is related to the retrieval and revision of all information required to create the forecast. There are 5 inputs for this 1st step of the
forecasting process:
• Historical data: includes the historical demand per item, the historical demand is also known as real. The data is aggregated depending on the DFU_MAP, this relation is also known as parent-child model. All the demand is registered and stored in the APS and IT is later used for the calculation of statistical forecast. By default 2 years of historical data are required for model calculation. This is the minimum length that APS system needs to run the model without exceptions. When possible, system also allows changing the amount of history for calculation. If historical data is shorter than 2 years system generates one exception called ShortHist (short history) which pinpoint the DFU’s enabling the planner to perform additional adjustments if required. An important point to remark here is that history is not cleaned before using it for calculation of statistical forecast, therefore outliers are taken in consideration for creation of statistical forecast. • Forecast accuracy report: This information is available on monthly buckets form
it is measured indeed is the APE. The forecast accuracy is measured and reported every month, the report measures the accuracy in lag 2, which corresponds to forecast lag. The calculations to measure Net deviation (PE) and APE are shown as follows:
100
*
Re
Re
1∑
=−
=
n t t t tal
al
FCST
PE
(1)100
*
Re
Re
1∑
=−
=
n t t t tal
al
FCST
APE
(2)Where t corresponds to week of demand (t=1, t=2, … tn), Real is the demand observed in the week t and FCST is the demand forecasted for that week. Forecast are produced at lag 2, therefore, the demand is forecasted two months in advance period t. Analyzing both formulas we can also observe that (1) is measuring difference between demand and forecast without absolute transformation, this means that number will be negative when demand results greater than forecast. Therefore this formula is good to measure under forecasting and over forecasting. Another purpose of the PE is to measure the BIAS in the forecast; this is the presence of a constant over or under forecasting in consecutive periods. Generally speaking forecast is considered as BIAS when present a continuous under forecast or over forecast for the last 3 months.
APE is measuring % of absolute deviation between demand and forecast for the period t. Here the absolute value of the difference is taken into consideration for each group of products. In the forecast accuracy report, APE reported as MAPE is available at different hierarchy levels and can be aggregated or decomposed up to P1 level in order to
analyze the contribution per families at lower hierarchy levels, allowing the planner to see the pinpoints for accuracy improvement. A detail in this formula, is that it accounts for either negative and positive values in the calculation, therefore considers addition of errors from all DFU’s in ane specific product family. Therefore, it is useful to measure the accuracy in terms of mix. In contrast, the PE offsets negative and positive values when they are added arithmetically (Makridakis, 1998). For this reason PE is used to measure the forecast error in terms of volume. As from now and in order to avoid confusions, term of APE and PE will be used as reference for both error metrics.
• Forecast adjustments: This database tracks forecast adjustments made in the short term. In this worksheet, the planner can observe the trend that the forecast followed in the previous 4 months and the forecast values for the upcoming 2 months. This analysis is called waterfall analysis and it is used to calculate the % of change of forecast with one period of lag. Positive percentages indicate that demand forecasted with actual model has increased in comparison with the forecast created one period before. Negative values have the opposite meaning.
• Top line overview: This report includes a snapshot with the last forecast draft stored in the system, including the demand planned for the coming 18 months. The report shows forecast trend in the long term giving visibility for strategic decisions and make
This report also allows to the planners to detect anomalies or overreaction in the statistical forecast which might result in a strong trend or seasonal pattern. The tool is primarily used for strategic decisions such as capacity management and decision where capital is required.
• Demand outliers: This dynamic report extracts the historical data at different
aggregation levels, tracking the historical demand. When forecast present one pattern that does not match with demand figures in previous period, planner verifies if there are outliers influencing the forecast result. If an outliers exist, planner adjust that forecast parameters in order to reduce or correct overreaction. Usually, the outliers are not deleted from history.
4.2.2 Forecast calculation
Forecast calculation at CL is done by system, each period forecast is calculated for a horizon of 18 months ahead. The system used to create the demand forecast is Manugistics. Manugistics is an APS with friendly interface where navigation easy supporting different user views for the main processes at the supply chain department. Manugistics counts with one specific user view where forecast are created. This user view is called DSM_FORECASTING_OUS and it counts with different screens of information related to the forecast. After selecting the desired product family and location of demand to be forecasted, Manugistics shows the screen in shown in Figure 16. This screen presents a graph that includes the historical demand (green area) and the demand forecasted (blue line) at the CL.
The quantitative forecast methods available in Manugistics are: Fourier, Multiple Linear
Regression and Lewandowski. Although three models are available, all products are forecasted using the Lewandowski method. The reason for this is that APS manual states that Fourier method works better when demand is constant and it’s changes occur at constant rate (level, trend and seasonality change with constant rates). In contrast, MLR methods work better when causal variables of demand are known (i.e. price, weather, demographics and economic population). Finally, the APS manual states that Lewandowski method is designed to forecast demand of products with changes in ales patterns or sporadic demand. Due to the fact that the first statement cannot be affirmed for all DFU’s and that the causal variables of demand are unknown, choosing Lewandowski method was a logical business decision for current situation. For the creation of forecast, planner selects the history and afterwards he or she runs the
default forecast model (lewandowski model10) in order to update the forecast for the next period.
A new version of forecast is created and system notifies with the message “Model fit complete”. This version is known as FCST_DRAFT and is the version that planner reviews before creating the definitive forecast.
Figure 16. DSM_Forecasting screen
4.2.3 Review and adjust statistical forecast at Control Level
After system updates the forecast and the screen shows the updated FCST_DRAFT, planner reviews the demand forecasted using eyeballing technique. During this step he or she decides if the forecast model requires further adjustments. If not further requirements are needed, forecast model does not suffer any further modification and it is saved in the system. When modifications in model are required, planner opens the forecast toolkit window and adjusts the Lewandowski parameters available in Manugistics. The parameters listed in forecast toolkit window for Lewandowski method are:
Parameters for dynamic mean
• Mean impact: Specifies how reactive the mean is to recent history. A low value means less reaction and it goes from 0.01 to 0.9
• Mean value max: Specify how reactive the mean to all historical demand is. Value goes from 0.01 to 0.9.
Parameters for trend
• Trend parameter: Sets the type of calculation for the mean of future demand • Constant trend: Mean is calculated based on a dynamic value without calculating a
• Linear trend: Mean is calculated using q dynamic value with system assigning trend to the model. Value in the system is 2.
• Quadratic trend: Mean is calculated using a dynamic value with quadratic trend calculated exponentially. Value in the system is 3.
• Automatic trend: System selects the best model suited to fit the DFU. Value in the system is 4.
• Dynamic trend 1: Uses constant model for the first cycle and then linear for the rest of the periods.
• Dynamic trend 2: Start with constant trend for first cycle, changes to linear for second cycle and finishes with quadratic trend for the rest of the periods.
Parameters for history
• Forecast trend combination: States how much weight is place to recent history. High values mean more weight to recent history. Value goes from 0.01 to 0.9
• Hybrid factor: Indicates how much weight is put into remaining history (medium-long term) Value goes from 0.01 to 0.9
Parameters for seasonality
• Seasonality: When seasonality option is activated. Specify how much seasonal influence will be considered in the forecast model and the starting point, value goes from 0.01 to 9.99.
Parameters for algorithm control
• Optimization steps: Determines the number of times that systems attempt to improve the forecast. Value should be integer between 1 and 6.
4.2.4 Incorporate regional information
Once forecast draft has been created, planner considers the constraints and requirements from demand based on information coming from OUS regions (such as recent increments or
reductions on demand) in this case, planner considers relevant information that may impact on the plan and make further adjustments in the forecast if required. Once planner has modified the forecast in order to include this information, the forecast is saved and stored in the system until its publication.
4.2.5 Reconciliation process (Statistical forecast at DFU level)
The reconciliation process translates the CL forecast into DFU level forecast (allocation forecast). In order to produce the forecasts by allocation, parent-child relations are defined in DFU_MAP’s. The reconciliation process is the step that ensures that forecasts for lower levels in a DFU_MAP (for instance, items) equals to the sum of the forecast from the aggregated level. Next picture shows a screenshot from MANU which graphically shows the DFU forecast and how they contribute to the aggregated forecast pattern.
Figure 18. Screenshot of a forecast of DFU_MAP
The forecast for each DFU is generated by allocation of the CL forecast. The fcts at CL is split (allocated) across the DFU’s depending on the % of volume from historical demand. The
reconciliation is an automatic process that runs every week considering demand occurred at the end of each week.
4.2.6 Publish final version of forecast
After the last 3rd week of SIOP calendar, the final forecast version that was already stored in the
4.3 Forecast error
Forecast error is the difference between actual demand and the forecasted demand (APICS CPIM, 2000). As mentioned before, Cordis measures the forecast error through APE and PE. According main stakeholders in Cordis, this error has been historically higher for specific
product families, especially for those items that are considered slow movers. X and Y families of products are characterized for being composed for a high diversity of products. The speed of demand for the products varies greatly. In Cordis, the speed of demand is related to volume of demand. In other words, fast movers have higher volumes of demand than slow movers. Apparently, the forecast accuracy is related to volume/speed characteristics. The next scatter plot shows the relation between the two variables. In the horizontal axis the error (APE) is plotted and Y axis shows the average volume of demand.
Figure 19. Relation between Volume and forecast error at family level11
The scatter plot shows data at P1 level (or family level), this is currently the most detailed level of information available in Cordis for accuracy metrics, this means that APE and PE are not measured at DFU levels. After analyzing the previous scatter plot it is easily noticed that there is a larger number of families from Y families with high errors. The scatter plot also shows X family has a portfolio with fast and slow movers and the same diversity occurs in terms of forecast error. By looking at the graph we can also infer that there is apparent relation between the volume of sales and error. In other words, fast movers present smaller forecast error than slow movers. This observation is very important because it tell us that current forecast model works better for certain demand characteristics.
4.3.1 PE and APE performance from top-down perspective
As already mentioned, PE measures forecast accuracy in terms of volume and APE in terms of mix with respect of demand. Historically, some Y families have presented higher error than others, especially compared with X families. In the next page, charts comparing historical PE and APE from X and Y families are shown.
Confidential
Figure 20. Error comparison between P4 families (X and Y)12
Analyzing previous graphs we did not find significant BIAS for both families, this tell us that amount of forecasted demand has a positive behavior or forecast deviation is properly corrected when it occurs. This also suggest that forecast parameters are properly adjusted at CL.
Further analyzing these graphs, we see that Y families present more variation in error values across the time. At the same time, Y products show higher forecast error. The dotted line in these graphs represents the internal goal for error. This goal is set as 30% for aggregated demand levels (+/- in the case of PE). We can also observe that APE values for X families average somewhere is between 10% and 20% while for Y is somewhere in between 20% and 30% with more peaks in the lines. These graphs show clearly the forecasting issue that Cordis’s management perceives.
4.3.2 PE and APE performance from bottom-up perspective
From bottom-up perspective, the PE and APE numbers look very different; this is mainly caused by the fact that forecasts are less accurate at disaggregated level. As explained before,
forecasts are generated by allocation, this means that planners review and adjust the forecasts at the control level. However, the system automatically generates the disaggregated forecasts. The forecast error at DFU level especially impacts the accuracy in terms of mix. This means that system is not allocating demand with enough accuracy. The control levels are forecasted within
a reasonable level of accuracy, this means that distribution of demand planned by the system can be improved and be more evenly distributed between DFU’s.
Confidential
Figure 21. APE comparison at DFU level13
We observe that APE classifications between both families are similar. However, error trend to be much higher at disaggregated level. More than the half of the DFU’s for both families present an APE value between 101% and 200%. The 2nd biggest classification corresponds to DFU’s
with bad APE, which represents 16% and 19% respectively. The third group corresponds to those DFU’s with good APE. Finally, we see a very small portion of DFU’s with very bad and very good performance and the smallest portion in this classification is represented by DFU’s with excellent APE.
The most important conclusion from these figures is that currently, CORDIS is producing a “good” forecast at the control level. This is especially true when considering accuracy in terms of volume. In other words, pooled forecasts are matching with less error the demand for regional wide demand per complete families. The large APE values observed at disaggregated levels mean that forecast allocation is not accurate; this does not necessarily means that Cordis is in big troubles because current forecasting process. According to these numbers, quantities forecasted to fulfill regional demand are reasonably ok, what needs to be improved is the accuracy for allocation forecasts. Furthermore, DRP process in Cordis is supporting inventory fare sharing at DC’s, this process ensures that demand for specific countries is fulfilled using available inventories.
4.4 Demand variability
In order to understand sources of variation, aggregated data should be decomposed into
disaggregated data (Gordon, T. Morris, J. and Dangerfield 1997 and Geurts, M. et al 1999). The unit of demand in Cordis is the DFU. Currently there are nearly 30,000 DFU’s for X and Y products. Businger, Read (1999) proposes one method for demand categorization for
thousands of SKU’s. This study takes a similar approach. Furthermore, since time series may
differ greatly between themselves; demand variability is measured using free scale metrics. The next paragraphs show the formulas used for calculations, firstly, lets define , ,…, as the demand occurring in weekly periods for each DFU.
1
y y2
y
n4.4.1 Volume variability (lumpiness)
The metric proposed for this purpose was the Coefficient of variation (CV). The CV measures the variation in terms of volume for the periods with demand. Another name to describe volume the variation is lumpiness. When demand presents high lumpiness is considered as erratic demand; in contrast the term non-erratic is applied when demand is more stable. The formula of CV is described as follows.
y
s
CV
=
(3) Where
y
represents the average demand from week to week and s represents the standard deviation of demand week to week . Note that this calculation takes the standard deviation and divides it by the historical average, by doing this it give us an estimation of the size of the bandwidth variation in each time series.1
y
y
n1
y
y
n4.4.2 Frequency variability (intermittency)
Another natural indicator used in supply chains centers that measures frequency (or
infrequency) of demand is the count of periods with demand equal to 0 (Businger, 1999). This indicator measures the intermittency or continuousness of demand and we can directly contect with XYZ classification presented in Schönsleben (2002). By counting the number of ceros we can determine which items are fast, intermediate and slow moving. The formula used to measure demand continuousness is:
