Incorporating ARIMA forecasting and service-level based replenishment in RFID-enabled supply chain

This paper focuses on the global supply chain of a company in Taiwan (referred to as Company A), which manufactures thin film transistor liquid crystal display (TFT-LCD) products. A simulated experiment is made 640 times, and a comparison of output analysis is also made. In the experiment, the key performance indicators are the total inventory cost, the inventory turnover rate, and the bullwhip effect. Four supply chain replenishment policies, four customer demand forecasting methods, radio frequency identification (RFID) and non-RFID system are the experimental factors and their levels, which generate 16 combinations of the Taguchi experiment. From the result, we find that the RFID-enabled R-SCI_ARIMA supply chain model which integrates the (s, Q) replenishment policy based on the ARIMA forecasting method and service level is the best: the total inventory cost has a 35.43% reduction, and the inventory turnover rate has a 61.36% increase, compared with that of the non-RFID SCI_ARIMA model.

Keywords: global supply chain; replenishment policy; ARIMA forecasting; RFID

In the traditional supply chain, the upstream manufacturers deal with customer demands on the basis of orders actually received downstream. Because much longer time is needed to reflect market changes, it does not satisfy the changing demand styles and causes the inventory to expire. In addition, barriers are generated while information is being processed in every tier. Information of demands is distorted during the movement among different supply chain tiers, and this leads to the increase of variation in supply chain orders, which produces the bullwhip effect. To solve the problem of bullwhip effect, the sharing and exchanging of information is necessary. One of the ways to achieve this goal is to use the RFID system. With the real-time product visibility and traceability of RFID, the communication of information among tiers can be accelerated, and the lead time of product delivery can be shortened. Moreover, the impact which results from errors in demand forecasting can be reduced, and the effectiveness of inventory management is thus increased. The RFID technology can provide item-level data along with stock keeping unit (SKU) and gain access to relative data with the unique electronic product code (EPC).

In January 2005, Wal-Mart and the US Department of Defense asked their suppliers to implement RFID technology, in order to shorten the lead time, save total costs and make decisions through data provided by RFID tags. For supply chain members, the implementation of the RFID technology does generate lots of effectiveness.

The purpose of this research is to establish a simulated model of a global supply chain, which incorporates the autoregressive integrated moving average (ARIMA) demand forecasting method and its service level as the basic inventory replenishment policy, for the TFT-LCD Company A, to verify that the implementation of an RFID system can best improve its effectiveness.

The remainder of this paper is organised as follows. Section 2 provides a literature review on the supply chain demand forecasting, RFID applications, supply chain simulation and agent model applications, and the impact of the bullwhip effect in a supply chain. The TFT-LCD global supply chain and modelling of Company A is presented in Section 3. The Taguchi methods/design of experiments and verification by simulation are depicted in Section 4. In Section 5, an analysis and comparison of the experimental results of KPI simulation output is made. We conclude the research in Section 6.

This research uses Company A's historical data of customer sales to make good predictions of the future situations. As a result, the time series method is adopted as the customer demand forecasting method. The most common time series methods include naïve, moving average (MA), exponential smoothing (ES), and autoregressive integrated moving average (ARIMA). Dhahri and Chabchoub ([7]) researched issues of the increase of the inventory level and the decrease of the service level, which result from the adoption of methods to weaken the bullwhip effect. Aburto and Weber ([1]) researched a mixed system combining ARIMA and a neutral network, which showed improvement on the accuracy of forecasting. They also designed a replenishment system, which leads to less sales loss and lower inventory level. Sun and Ren ([22]) studied the impact of demand forecasting on the bullwhip effect in supply chain management (SCM). The result showed that an increase in the lead time will increase its variation, and the degree of impact is determined by the forecasting methods adopted. The study of Chandra and Grabis ([4]) indicated that the existence of the bullwhip effect and the increase in the variation of orders result in a lack of efficiency in inventory management. The result showed that parameters considered in the autoregressive models do effectively reduce the bullwhip effect, but the variation of parameters will also affect the degree of the bullwhip effect. Zhang's ([25]) study focused on the impact of forecasting methods on the bullwhip effect. The result showed that an increase in the lead time and underlying parameters of the demand process will strengthen the bullwhip effect, and the degree of effect varies according to different forecasting methods.

In conclusion, these studies all show that the ARIMA demand forecasting method can reduce the bullwhip effect in a supply chain. However, they do not take the integration of RFID and the factor of real-time information sharing into consideration. This research combines a supply chain with the real-time and fast responding character of RFID in expectation of enhancing the effectiveness of demand forecasting, in order to make further studies on issues of the service level in replenishment management. Also, the applicability of ARIMA will be simulated and compared with other forecasting methods.

Delen et al. ([6]) analysed RFID data collected from retailers and suppliers in a supply chain to know how to estimate the time needed from the logistics centre to retailers through RFID. Saygin et al. ([20]) designed methods for establishing an RFID supply chain system and emphasised the communications infrastructure necessary to provide seamless data and information flow in order to achieve RFID data-based decision-making at all levels of the supply chain. Mills-Harris et al. ([18]) produced a simulated study on the inventory management of time-sensitive materials, based on data collected by RFID. The forecast integrated inventory model was developed based on a trend adjusted exponential smoothing algorithm. The result showed that a proper adjustment of the two smoothing parameters (α and β) can achieve the system performance demanded. Moreover, in January, 2005, a successful trial of the RFID/EPC system on tagged pallets and cases was done by Wal-Mart and its top 100 suppliers. The University of Arkansas analysed Wal-Mart's success and found that after adopting the RFID/EPC system, there was a 16% decrease in the out-of-stock rate (Wal-Mart Stores Inc. [23]). Hardgrave et al. ([9]) researched 24 retailers of Wal-Mart, divided into two groups, each group consisting of 12 retailers. The result showed that the group which implements RFID has a 26% decrease in the out-of-stock rate and has improved 63% compared to the group without RFID. Lee et al. ([12]) of IBM proved the potential effectiveness of RFID in decreasing the inventory and enhancing the service level with simulation methods based on real data, and the subject is a three-tier supply chain.

The studies above only involve parts of the supply chain tiers without taking the complexity of the whole supply chain into account. Therefore, this research takes Company A's global supply chain of TFT-LCD as an example, designs a platform for the RFID network database of products to simulate the operation model of the RFID-enabled global supply chain, and uses experimental design methods to prove the potential effectiveness of RFID for the improvement on the supply chain inventory management.

Kleijnen ([11]) surveyed four types of supply chain simulation: spreadsheet, system dynamics (SD), discrete-event dynamic system (DEDS), and business games. The survey concluded that the DEDS simulation is an important method in SCM. It can represent individual events and incorporates uncertainties. Borshchev and Filippov ([2]) suggested that the system being modelled contain active objects (people, products, stocks, business units, etc.) with timing and event ordering and that it is suitable to add an agent based model to the DEDS simulation background. Özbayrak et al. ([19]) established a four-tier supply chain. The performance evaluated was mainly based on inventory, the WIP level, backlogged orders, and customer satisfaction. Li and Wei ([14]) offer technology which integrates RFID and a multi-agent to monitor the location of transferring goods and the information of the environment. Liang and Huang ([15]) established a multi-agent system in a supply chain, in which the inventory system is operated through different agents. The result of the agent-based system showed the reduction of total cost and the smoothing of orders variation curve. Emerson and Piramuthu ([8]) established an agent-based dynamic supply chain infrastructure. They also proved through real cases that the performance of a dynamic supply chain infrastructure is better than that of a static supply chain infrastructure.

The traditional management system has difficulty in practising supply chain management due to its complexity, so observation and testing of the system simulation is needed. Therefore, the dynamic system simulation method is used for establishing an RFID-enabled supply chain model, which consists of agents of planning management, stock control and executive operation. Moreover, the RFID agent and the demand forecasting agent are added in order to figure out the most appropriate replenishment policy and demand forecasting method for the RFID-enabled supply chain.

Lee et al. ([13]) concluded that the bullwhip effect results from the distortion of information during the process of communication in the supply chain. They defined that the factors which lead to the bullwhip effect include demand forecasting, order batching, price fluctuation, rationing and shortage gaming. Four methods should be adopted to effectively weaken the bullwhip effect, and they are: avoiding multiple demand forecast updates, breaking order batches, stabilising prices, and eliminating gaming in shortage situations. Luong ([16]) concluded that the bullwhip effect will appear when information of demand moves upstream. In their supply chain, the impact of autoregressive coefficient and lead time on the bullwhip effect is studied through a first-order autoregressive model by the retailers, based on inventory policy. Xu et al. ([24]) used a demand model and time series to prove the bullwhip effect really exists in the three tiers of supply chain, and that the bullwhip effect can be effectively weakened through information exchange and consistent forecasting. Kelle and Milne ([10]) thought the bullwhip effect is a phenomenon that results from demand forecasting and batch orders. Small frequent orders can reduce the effect of high variability and the resulting uncertainty, therefore effectively weakening the bullwhip effect.

In conclusion, due to changing customer demand, the supply chain becomes instable and hard to control. Thus, it is difficult to forecast the change of orders. So, this research uses some demand forecasting methods, like ARIMA, to forecast the future demand in the hope of effectively weakening the bullwhip effect.

According to the Materialsnet ([17]), the Industry Economic Knowledge Center of Industrial Technology Research Institute in Taiwan produced statistics showing that the total production value of the flat display panel in Taiwan was 40.3 billion USD in the year of 2007, which overtakes the 34.5 billion USD in South Korea and the 22.5 billion USD in Japan. This makes Taiwan the real top flat display panel manufacturer. At the same time, the total production value of the flat display industry of 2007 in Taiwan was 1.78 trillion NTD, which is a 39.8% increase compared to that of 2006.

The structures of the upstream, midstream, and downstream of the TFT-LCD industry are quite enormous. The components manufactured upstream include crystal, glass substrate, colour filter, driver IC, polariser, and back light. They are assembled in the midstream, and then can be applied in electronic appliance, consumer products, communication, transportation, computer, and business products.

In this case study, the global operations structure of Company A includes eight branch warehouses, three regional distribution centres, five LCD monitor manufactories, and four LCD panel manufactories. Its global inventory is computed every day, and the data is transmitted to the headquarters in Taiwan to be organised. In order to solve the problem in inventory management, this research establishes a demand forecasting model based on an inventory replenishment policy with a certain service level. The purpose is to forecast the future demands of customers and to solve the problem in inventory management by implementing RFID to transmit instant information.

This research adopts AnyLogic, a system simulation tool, to establish a mechanism of supply chain automatic replenishment simulation with a demand forecasting model. AnyLogic is professional simulation software that can be applied for discrete, continuous and hybrid system modelling. The interface of the simulation model R-SCIARIMA established in this research is shown in Figure 1.

Graph: Figure 1. The simulation main screen of R-SCIARIMA model.

In this research, the mechanism of inventory replenishment simulation is operated through functions of agents to monitor the entire supply chain system and collect instant information with RFID. The agents can be sorted into three categories: planning management, stock control, and executive operation. The detailed R-SCIARIMA multi-agents simulation procedures are described as follows:

• 1. When a number of units for customer demand generated by the system with a Weibull distribution model, the order check agent will notify the demand forecasting agent after the demand is confirmed.
• 2. The demand forecasting agent will process the calculation of forecasted demand based on the ARIMA model and then offer the number of forecasted units for customer demand to the order management agent.
• 3. The order management agent then asks the finished goods agent to release finished goods with required forecasted units. If the finished goods on-hand units are in short supply, existing finished goods will be issued at once. Besides, the delayed time of the issuing operation will be generated.
• 4. A real-time monitoring of releasing finished goods is performed through the RFID agent simultaneously. The stock units monitor agent will obtain real-time transactions information of the finished goods on-hand units. The elapsed time of retrieving and transferring RFID tagged data is assumed to be 0.025 week. On the contrary, it is enlarged to be 0.15 week for the bar-code system.
• 5. The out-of-stock units will be recorded by the order management agent and the production management agent as notified by the stock units monitor agent. Thus, the out-of-stock units must be released first when newly finished goods are received.
• 6. In the simulation run, the supply chain management agent will publish the inventory replenishment policy (s, Q) to the stock units monitor agents of each tier in the supply chain. The amount of maximum raw materials on-hand S for any tier's member is also assigned by the supply chain management agent.
• 7. The stock units monitor agent monitors the transactions of on-hand units constantly. If the on-hand units are found to be fewer than the reordering point s, the requirement of replenishment Q is sent to the order management agent.
• 8. The order management agent sends replenishment demands (=Q purchasing units) to the upper tier's supplier in the supply chain. (The procedure of demand orders handling in each tier of the supply chain follows that of steps 2 to 8.)
• 9. When the replenishment of raw materials arrives, the stock units monitor agent of each tier will receive the information of the gain of on-hand units sent by the RFID agent. Then, the total of the present on-hand units and the scheduled receipt units will be checked to see if it exceeds the amount of the maximum raw materials on-hand units. The purchased units agent is only responsible for the release and carrying operation of the raw materials in stock.
• 10. If the stock units monitor agent finds that the total of the raw materials on-hand units and the scheduled receipt units exceeds the assigned maximum raw materials on-hand units S, it will notify the order management agent to cancel the purchasing orders which have not been released in the preceding tier.
• 11. When the production management agent of the manufacturing plant tier receives the information of an out-of-stock in the finished goods stock, it will ask the production agent to begin production.
• 12. When the production operation is being performed by the production management agent in the manufacturing plant, the work-in-process units increase and the raw materials units decrease. There will be delayed time in the production operation. The units of finished goods will be increased at the end of the production process.
• 13. The stock units monitor agents will be simultaneously informed of the work-in-process units by the production management agent.

This research is based on the global supply chain operation process of Company A, with an RFID mechanism in addition. The receiving and shipping point of each location is equipped with RFID, including one reader and two antennas. The tagged EPC data of each product, affiliated with RFID tagged records like the receiving time and shipping time, is transmitted to the reader. The data is then transmitted through local networks and saved in a MySQL database. The tagged EPC code is disassembled into the RFID EPC records, and the EPC records saved in each location are then transmitted to the central server via the internet for tracking, tracing and stock units update of the supply chain. This platform uses Labview as a software development tool. The whole RFID-enabled application infrastructure is shown in Figure 2. In this research, the RFID mechanism in the global supply chain simulation model of Company A established by AnyLogic will be gradually simulated with reference to the network platform operation flow above.

Graph: Figure 2. Conceptual structure of RFID-enabled supply chain applications.

The case of Company A is a multi-tier global supply chain model. The simulation period T is set to be 52 weeks. Based on the actual demand in 52 weeks of the 17-inch TFT-LCD of Company A, the statistics distribution model analysis tool, Stat:Fit, is used to make an analysis. The result shows the customer demand statistics distribution model of Company A approximates the Weibull (min = 6980, α = 4.13, β = 128).

In this research, the inventory replenishment policies include the continuous review (s, Q), (s, S) and the periodic review (R, s, S), (R, S), in reference to the formulae designed by Chopra and Meindl ([5]), and Simchi-Levi et al. ([21]), with the prerequisite of the desired cycle service level (CSL) to be 95%. The reorder point (s), order-up-to level (S) and order quantity (Q) are calculated individually. The result is shown in Table 1. The continuous review policies of (s, Q), (s, S) are described as follows:

Graph

Graph

Graph

Graph

Graph

Graph

Where:

• D average weekly demand faced by each tier member;
• σ standard deviation of weekly demand;
• L lead time for replenishment;
• D L average weekly demand during lead time;
• σ L standard deviation of weekly demand during lead time;
• ss safety stock;
• s reorder point;
• S order-up-to-level;
• PC purchasing cost;
• CC carrying cost.

Table 1. Input parameters of (s, Q) and (s, S) replenishment policy with 95% CSL.

This research focuses on the forecasting of the end customer's order demand from the retailers. It is based on the actual customer demand offered by Company A, and the forecasting model consists of ARIMA, trend-corrected exponential smoothing, simple exponential smoothing and the naïve forecasting methods. The model is used for calculating the customer demand of the next week.

This research uses SPSS to establish the ARIMA model, based on the four steps by Box and Jenkins ([3]): identification, estimation, diagnosis and forecasting. The complete establishment flow of the model as shown in Figure 3 is as follows:

• 1. _B_Identification

Graph: Figure 3. The flowchart of ARIMA forecasting model.

First, the time series chart will be observed through its autocorrelation coefficient to check whether the chart is stationary or semi-stationary. The autocorrelation function (ACF) and partial autocorrelation function (PACF) will be used to check the orders of p and q (p is an autoregressive order, and q is a moving average order). In the check of the time series, the result shows that the actual customer demand data of Company A is of trend-affiliated time series. So, the first-difference is calculated to wipe out the trend to make it a stationary time series. The difference d is 1, as observed in the ACF and PACF graph, and candidate models (p, d, q) are generated such as (1, 1, 0), (1, 1, 1), (1, 1, 2), (2, 1, 0), (2, 1, 1) and (2, 1, 2). The Akaihere's information criteria (AIC) value of each candidate model is quite close. Therefore, residual ACF and PACF graphs are generated to check if the residual is white noise. The result shows that the ACF and PACF fall in the confidential interval when (p, d, q) = (2, 1, 2) and (p, d, q) = (1, 1, 1). The AIC value of (1, 1, 1) is 540.882, which is larger than the 536.614 of (2, 1, 2). Therefore, the ARIMA = (2, 1, 2) is selected as the customer demand forecasting model.

• 1. _B_Estimation

The parameters estimation will be carried out to find out the degree of impact of each lag variable on the forecasting series. The model of the ARIMA (2, 1, 2) can be induced by using first-difference model Δx(t) = x(t) − x(t − 1) and ARIMA (2, 2) model:

Graph

Therefore, the forecasting model is as follows:

Graph

Suppose the forecasting value is the current value, x(t) is the current value, x(t − 1) is the last period, x(t − 2) is the period before last, x(t − 3) is the two periods before the last, and u(t) is the residual, then u(t) ∼ N(0, σ2).

• 1. _B_Diagnosis

A test of model fitness will be carried out to make a check with its residual distribution. This research tests the acceptability of ARIMA = (2, 1, 2) with the Portmanteau Q-test model suggested by Ljung-Box:

Graph

In this formula, n is the number of actual residual (n = N − d = 60 − 1 = 59, N is the number of time series data, d is the order of difference), k is the number of residual ACF, is the residual (k = 1, 2, ..., K). Finally:

Graph

shows that there is no significant relation between residuals. Therefore, the forecasting model should be ARIMA (2, 1, 2).

• 1. _B_Forecasting

We can design the forecasting model as follows:

Graph

The forecasting value of ARIMA (2, 1, 2) is calculated as follows:

Graph

According to the forecasting value of period 53 for Company A, the actual demand of period 52 was , the actual demand of period 51 was , the forecasting value of period 52 was , the forecasting value of period 51 was , therefore, the forecasting value of period 53 is as follows:

Graph

According to the trend-corrected exponential smoothing method designed by Chopra and Meindl ([5]), from the demand Dt and the time period of linear regression, the level and the initial value of the trend appear as follows: Dt = at + b. With the constant b, the demand of period t = 0 can be estimated, which is also an estimate of the initial level L0. The constant a represents the change rate of each period's demand and the initial estimate of trend T0. With the regression function of Excel, the value of initial level L0 can be estimated from the INTERCEPT worksheet function, and the value of trend T0 can be estimated from the parameter of variable t. In the case of Company A, the result is L0 = 7105.038 and T0 = −1.34332. α and β are the smoothing coefficients of the level and trend, and there are nine combinations with the values of 0.2, 0.5 and 0.8. From the calculation of mean squared error (MSE), the MSE is the smallest when α = 0.2 and β = 0.2. In the case of Company A, the result of the forecasting estimate of the 53rd week based on the actual demand of t = 52 is as follows:

Graph

Graph

Graph

According to the simple exponential smoothing method designed by Chopra and Meindl ([5]), the initial estimation of level L0 can be calculated with the average of all historic data. Based on the demand data with period n = 52, the initial estimation can be presented as follows:

Graph

The selection of the smoothing coefficient α is determined by the smallest value of MSE, which falls at the value when α = 0.9. Therefore, the estimation of the level of the 52nd week calculated with equation is as follows:

Graph

As a result, the forecasting value of period 53 is .

The naïve forecasting method uses the single value of the preceding period in the time series as the basis of forecasting. It can be applied in stationary time series, seasonal variation or trends. In other words, the forecasting value of any period equals to the actual value of the preceding period: . In the case of Company A, the forecasting value of period 53 is .

This research assigns the total inventory cost, the inventory turnover rate and the bullwhip effect as key performance indicators (KPIs). The subject for simulation is the five tiers of the supply chain of Company A, and the simulation time lasts for 52 weeks. The KPI equations are as follows:

• • Total inventory cost = production cost + inventory replenishment cost + backorder cost + delivery cost:

Graph

• • Inventory turnover rate = sales amount ÷ inventory cost:

Graph

• • Bullwhip effect:

Graph

where Var(QN) represents the deviation of demand orders issued by the Nth tier in the supply chain, and Var(D) represents the deviation of end customer demand.

The following notation is used in the global supply chain simulation model.

• N the set of members in each tier of the supply chain;
• PAQ nt the throughput units of the production management agent of manufacturer n in period t ;
• IAQ nt the issued units by the finished goods agent of manufacturer n in period t ;
• SC n the unit sale price of finished goods produced by manufacturer n ;
• MAQ nt the raw materials on-hand units monitored by the purchased units agent of manufacturer n in period t ;
• MC n the raw materials unit cost of manufacturer n ;
• MSC n the raw materials carrying cost per unit of manufacturer n ;
• PAPQ nt the work-in-process units of the production management agent of manufacturer n in period t ;
• PC n the work-in-process cost per unit of manufacturer n ;
• AC n the issuing activity unit cost of manufacturer n ;
• OAQ nt the raw materials purchased units issued by the order management agent of manufacturer n in period t ;
• MOC n the raw materials purchasing unit cost of manufacturer n ;
• IASQ nt the backorder units recorded by the finished goods agent of manufacturer n in period t ;
• SOC n the backorder unit cost of the finished goods of manufacturer n ;
• QAQ nt the scheduled shipping units monitored by the stock units control agent of manufacturer n in period t ;
• ShC n the shipping cost per unit of manufacturer n ;
• IAIQ nt the amount of finished goods on-hand units monitored by the finished goods agent of manufacturer n in period t ;
• GC n the finished goods unit cost of manufacturer n ;
• D nt the demand units received by the order management agent of manufacturer n in period t.

This research adopts the Taguchi method to carry out experiments effectively with fewer combinations of experiments. There are three phases in the Taguchi method: planning, execution, and the analysis and verification of the output.

• 1. _B_Experimental planning phase

In this experiment, the quality characteristic of the global supply chain model of Company A is the total inventory cost. Its characteristic is the smaller-the-better, meaning that the cost should be smaller. The definition of the smaller-the-better signal-to-noise (SN) ratio is as follows:

Graph

In the equation, MSD is the mean square deviation and yi is the output value. The control factors and their levels are listed in Table 2. In this experiment, there are two level-4 factors and one level-2 factor in Table 3, which is an L16(4)3 combination known from Minitab.

• 1. _B_Experimental implementation phase

Table 2. Control factors and levels in Taguchi experiments.

Table 3. L16(4)3 experiment combinations.

The 16 combinations determined by the last step are being experimented and replicated 40 times, and the total simulation experiment times are 16×40 = 640 times. The summarised table of the experimental output data is shown in Table 4.

• 1. _B_Experimental outputs analysis and confirmation

Table 4. Partial experimental output data.

The experimental result is shown in the last column of Table 4, and the total average of the 16 SN ratios is:

Graph

The average SN ratios of the three factor levels can be calculated by using the SN ratios in Table 5. The average SN ratio of level-1 factor A is ; the average SN ratio of level-1 factor B is ; the average SN ratio of level-1 factor C is , and the others likewise. According to the definition of the SN ratio, the larger the SN ratio is, the better the quality is. So, a level of larger SN ratio is selected. The optimal level combination of this research is A1B1C1. The result of the analysis of variance (ANOVA) shows the F value of factor B is 0.3886, and the F value of factor C is 0.4594. Both are smaller than the significance level of 0.05. Thus, factors B and C are significance factors, and factor A is not a significant factor and is to be merged as errors. The SN ratio under the optimal conditions is

Graph

Table 5. The simulation results of four models.

In order to effectively estimate the observed values, the confidence interval (CI) must be calculated. The confirmation of the expected average value of experiments is:

Graph

That is, it can be concluded with 95% confidence in this experiment that the boundary of the expected SN ratio is . The SN ratio average of the five experiment confirmations is −175.446, which falls in the confidence interval above. This means that the selected factors B and C and their levels are adequate. When the success in the experiment is confirmed, the combination of the optimal level of control factors is included in Company A's global supply chain system. The specification is to implement the (s, Q) inventory replenishment policy, ARIMA demand forecasting method and the RFID-enabled system. It is called the R-SCIARIMA (RFID-enabled supply chain inventory demand forecasting: ARIMA).

This research plans to verify if the R-SCIARIMA is the optimal model through four global supply chain inventory management models.

• • SCI (supply chain inventory)

This model represents the current environment and simulates the current inventory operation of Company A. The current operation implements the (s, Q) replenishment policy and checks the weekly inventory level of each supply chain member to increase the purchasing orders to Q units.

• • R-SCI (RFID-enabled supply chain inventory)

This model supposes each inventory item is recorded on an RFID tag, and its visibility is 100%. The fixed baseline is similar to that of the SCI model, and RFID is only for monitoring rather than dynamically modifying the inventory level.

• • SCIARIMA (supply chain inventory demand forecasting: ARIMA)

This model implements the (s, Q) replenishment policy and the ARIMA demand forecasting method. It uses forecasting methods to estimate the future demand in order to reduce the inventory cost by decreasing variation of orders.

• • R-SCIARIMA (RFID-enabled supply chain inventory demand forecasting: ARIMA)

In addition to using ARIMA demand forecasting methods, this model improves demand management with the immediacy of RFID. Forecasts can be done and information of demand can be received at any time on the basis of instant demand information. However, this research uses a simulation method, in which the time of the simulation process is short, and the data RFID can read is quite large. Therefore, the virtual RFID system method is implemented here.

In this research, RFID equipment, including two antennas and a reader, are set in the receiving/shipment location of each supply chain tier. The price of each antenna is about $290, and the price of each reader is about $995. The proposed life of RFID is five years, the salvage value is 10%, the simulation time is one year, the maintenance expense of RFID is 12%, and the operation cost is 15%.

The RFID equipment cost of each supply chain tier is 2 × 290 + 1 × 995 = $1575. Based on the accelerated depreciation method, the salvage value of one RFID system at the end of the fifth year will be 1575 × 10% = $157.5. The maintenance cost is 1575 × 12% = $189, the operation cost is 1575 × 15% = $236.25, and the depreciation cost equals (1575 − 157.5) × [(5 − 1 + 1) / (1 + 2 + 3 + 4 + 5)] = $472.5. Thus, the total cost of one-year simulation for a set of RFID equipment = the maintenance cost + the operation cost + the depreciation cost = $189 + $236.25 + $472.5 = $897.75. The cost of one tag is about $1, and the average throughput in one year of LCD panel manufactories is 2,701,171 pieces. The attached tags are for entire usage of the supply chain, so the cost of RFID tags is about $2,701,171. The individual RFID cost of each tier is: LCD panel manufactories = 897.75 × 4 × 1 + 2,701,171 = $2,704,762, LCD monitor manufactories = 897.75 × 5 × 2 = $8977.5, regional distribution centre = 897.75 × 3 × 2 = $5,386.5, branch warehouses = 897.75 × 8 × 2 = $14,364, and retailers = 897.75 × 8 × 2 = $14,364.

The simulation output analysis and comparison of the R-SCIARIMA model is carried out with the Bernoulli experiment, the value of each KPI and the sensitivity analysis, to verify if the R-SCIARIMA model is the optimal.

From the result of the analysis of the normal distribution graph with Minitab, the total inventory costs and inventory turnover rates of R-SCIARIMA and SCIARIMA fall in the 95% confidence interval, which are in accordance with a normal distribution. The Bernoulli experiment is carried out in order to test the difference between the average total inventory costs of R-SCIARIMA and SCIARIMA, and the test statistics are as follows:

Graph

Graph

Under two-tailors test (α/2 = 0.05) in the 95% confidence level, Z = −111.6789, which is less than Z(α/2 = 0.025) = −1.96 and falls in the reject area. Therefore, we accept the alternative hypothesis H1: there are differences between the total inventory costs of the two groups.

The Bernoulli experiment is carried out in order to test the difference between the average inventory turnover rates of R-SCIARIMA and SCIARIMA, and the test statistics are as follows:

Graph

Graph

Z = 146.96, which is larger than Z (α/2 = 0.025) = 1.96 and falls in the reject area, does not accept the null hypothesis H0: no difference exists in the average inventory turnover rate of the two models. In conclusion, at a 95% confidence level, there are differences between the inventory turnover rates of the two models.

From the simulation outputs shown in Table 5, based on the experiment output of R-SCIARIMA, it can be known that the R-SCIARIMA is the best of the four models. It has the lowest production cost, replenishment cost, backorder cost and total inventory cost.

Table 6 shows the improvement rate of each tier member in the supply chain model. It shows that the costs in every subject are improved from the comparison of the R-SCIARIMA and SCIARIMA models. Among them, the improvement rates of production cost and backorder (out-of-stock) cost relatively increase a lot. Take retailers for example, in the R-SCIARIMA model, the production cost has a 28.42% decrease, and the backorder cost has a 50.32% decrease. The total inventory cost of the R-SCIARIMA model has a 31.93% decrease. Table 7 shows that the costs in every tier are improved from the comparison of the R-SCIARIMA and R-SCI models.

Table 6. The R-SCIARIMA vs SCIARIMA improved costs of each tier's member.

Table 7. The R-SCIARIMA vs R-SCI improved costs of each tier's member.

Table 8 shows the improved value of the inventory turnover rate of each tier member in the supply chain model. It shows that the inventory turnover rate in each tier is improved from the comparison of the R-SCIARIMA and SCIARIMA or R-SCI models. Take retailers for example: the inventory turnover rate has a 10.65% and a 3.95% increase, respectively. Take branch warehouses for example: the inventory turnover rate has a 90.13% and a 14.40% increase, respectively.

Table 8. The improved inventory turnover rate of each tier by R-SCIARIMA vs SCIARIMA and R-SCI.

According to the definition of the bullwhip effect, the bullwhip effect values of the R-SCIARIMA, SCIARIMA, R-SCI, and SCI models are calculated in Table 9, respectively. Take LCD panel manufactories for example. In the R-SCIARIMA model, the bullwhip effect value = LCD panel manufactories demand variation ÷ customer order deviation = 904,601/26,870 = 33.67. Table 9 shows the output of the calculation of the bullwhip effect of each model according to the definition. In conclusion, the RFID system and ARIMA demand forecasting method both being implemented together in the supply chain can achieve a significant decrease degree of bullwhip effect.

Table 9. The comparison of bullwhip effect by four supply chain inventory models.

This research carries out the analysis of the sensitivity of lead time (LT) and service level (SL) with the optimal model: the R-SCIARIMA model. The lead time is analysed by the method of the original setting LT ± 5%, which affects the reorder point (s). From the comparison of total inventory costs in Table 10, it can be known that if the lead time increases, the total inventory cost increases.

Table 10. The sensitivity analysis of total inventory cost by lead time.

From the comparison of inventory turnover rates in Table 11, it can be known that the result of the variation is the same as that of the total inventory cost.

Table 11. The sensitivity analysis of inventory turnover rate by lead time.

The analysis of sensitivity is carried out with service levels of 99%, 97%, 95%, 93%, and 90%. As service level increases, the risk of backorder will decrease. From Table 12, it can be known that the higher the service level is, the less the out-of-stock cost and risk will be.

Table 12. The comparison of backorder cost by service levels.

This research focused on the simulation analysis of the TFT-LCD global supply chain model of Company A. It was found that the existence of the bullwhip effect will affect the effectiveness of the whole supply chain, especially in the domain of inventory management. From the result of the Taguchi experiment, we learn that there is a combination of optimal levels of the control factors in the total inventory cost of the R-SCIARIMA model. The R-SCIARIMA has a 35.43% decrease in the total inventory cost and a 61.36% increase in the inventory turnover rate, in comparison with the SCIARIMA model. This shows that the implementation of an RFID system has a significant effect of improvement on the supply chain. The R-SCIARIMA has a 13.08% decrease in the total inventory cost and an 11.08% increase in the inventory turnover rate, in comparison with the R-SCI model. This shows the adoption of the ARIMA forecasting method in addition to RFID can effectively enhance the performance of forecasting and gain significant effectiveness of improvement in inventory management. In the evaluation of the bullwhip effect, the establishment of the R-SCIARIMA model in Company A can weaken the bullwhip effect and gain optimal effectiveness of improvement. Finally, we conclude that the real-time transmitting capability and visibility for acquiring RFID-enabled tagged product data across a supply chain can synchronously trigger the ARIMA forecasting process and eventually decrease the supply chain total inventory cost.