Inventory Write-Down Prediction for ... - Semantic Scholar

Inventory Write-Down Prediction for Semiconductor Manufacturing Considering Inventory Age, Accounting Principle, and Product Structure with Real Settings

Jei-Zheng Wu* Department of Business Administration Soochow University 56, Section 1, Kuei-yang Street, Taipei 10048, Taiwan, R.O.C. *

[email protected] / Tel: +886 (2) 2311-1513 #3403 / Fax: +886 (2) 2382-2326

ACKNOWLEDGEMENTS

This research was partially supported by the National Science Council, Taiwan (NSC99-2410-H-031-002) and Macronix International, Ltd. Special thanks to the domain experts Dr. C.Y. Lu, Paul Hsu, Y.L. Whang, Fred Su, Hero Chen, T.S. Huang, Sophie Li, Ryan Lin, and Professor Chen-Fu Chien.

Wu*, J.-Z. (2013), “Inventory write-down prediction for semiconductor manufacturing considering inventory age, accounting principle, and product structure with real settings,” Computers & Industrial Engineering, 65(1), 128-136. (http://dx.doi.org/10.1016/j.cie.2011.11.020)

1

ABSTRACT

The International Financial Reporting Standards (IFRS) No. 2 has been the worldwide accounting principle for the reduction of inventory to market allowance since January 1, 2005. Using make-to-stock manufacturing strategies and inventory accounting for only approximately 14 % of the total costs, integrated device manufacturers have found maintaining robust records for financial statements increasingly difficult. For example, one company in the case study conducted in this study must write-down losses of 2 to 100 % of the total inventory costs for products with inventory ages of 18 months to 3 years. However, the average cycle time for producing flash memory is approximately three months. In other words, when the system variation and safety stock policy are considered, the company must write-down the reduction of inventory to market allowance for most of work-in-process inventory. However, little research has been done to addressing the practical management of operations according to inventory aging processes. This study develops a polynomial-time-based model to obtain significant features, including inventory ages, accounting principles, and product structures (bill of material), for the accurate prediction of inventory write-downs to reduce the impact of the carrying value fluctuation of inventory. An empirical study was conducted on a Taiwanese semiconductor manufacturer. The results show that predicting three-month inventory write-downs of a complete flash memory production line comprising approximately 8,500 product types can be conducted in less than 10 s, with the mean absolute percentage error (MAPE) less than 3.5 %. Discussions regarding the sensitivity analysis and cost tornado diagrams suggest the priority of affecting factors. The results show the viability of implementing the proposed model to predict inventory write-downs in the semiconductor manufacturing industry.

Keywords: semiconductor manufacturing, supply chain management, inventory age, inventory write-down, accounting principle, bill of material, obsolescence, International Financial Reporting Standards

2

1. Introduction

The semiconductor industry is continuously growing with extensive applications in medical electronics, green energy, car electronics, computers, communication, and consumer electronics (MG+4C). The Semiconductor Industry Association (SIA) reports a double-digit growth of all major semiconductor product categories in 2010 compared to 2009. The global semiconductor chip sales reached a record US$298.3 billion with a nearly 32 % annual increase (SIA, 2011). However, numerous semiconductor memory companies have been exposed to the risk of excess inventory levels and frequent inventory write-downs (Chen et al., 2010). The reason can be ascribed to industry characteristics, including high capital intensity, rapid technology development, and severe supply chain competition (Aizcorbe, 2002; Chien, 2007; Leachman et al., 2007; Chien et al., 2008; Chen et al., 2010). Specifically, high capital expenditure prompts semiconductor manufacturers to fully use capacity, which leads to high levels of inventory accumulation during low-demand periods. Following Moore’s law, the new generation product will dominate prior generations regarding the cost-per-function. This technology migration will accelerate the price decline and replacement of prior generation products, rendering the existing inventory obsolete. Additionally, the increasingly fierce competition has commodified chip sales. The continuous and significant price decline leads to a market value lower than the manufacturing costs, which is another cause of inventory write-downs. On January 1, 2005, the International Financial Reporting Standards (IFRS) Foundation Trustees declared the IFRS No. 2 (2010) the accounting principle for the reduction of inventory to market allowance. However, since its implementation, integrated device manufacturers have experienced even greater difficulty maintaining robust records for financial statements. Inventory write-downs are recorded as part of cost of goods sold that led to disavantagous gross profit. Semiconductor manufacturing is complex and lengthy. For example, the average cycle time of flash memory is approximately three months, including the 50 days for wafer fabrication, 5 days for circuit probing, 7 days for chip assembly, and 22 days for final tests. The average cycle time of semiconductor manufacturing varies across products because of a number of manufacturing strategic decision settings, including demand planning (Chien et al., 2010a), new product ramping schedule and allocation (Chien et al., 2011), capacity planning (Chien and Zheng, 2011), wafer start plan, work-in-process (WIP) level, tool availability (Kuo et al., 2011), outsourcing strategy and order allocations (Wu and Chien, 2008a; Chien et al., 2010b), and scheduling (Wu and Chien, 2008b). In addition, three months to one year of safety stocks are typically established depending on the product market and customer satisfaction level. In these cases, most of the WIP and end products become the amortization item to be written down. Additionally, variations in financial performance may further cause overreactions in the stock market when a significant amount of reduction of inventory to market is reported at once. This overreaction is typically nonreversible. That is, no compensation will be provided even after written-down inventories are sold thereafter. The timing and magnitude of inventory write-downs are crucial to earning management (Chen et al., 2010). However, knowing that prior research addressed inventory write-down estimation primarily based on economic factors and temporal history data, this study develops a multi-period inventory write-down prediction model that captures specific company features, including accounting principles, inventory ages, and product structures. An empirical study was conducted on a semiconductor manufacturer located in the Hsinchu Science Park in Taiwan to demonstrate the viablity of the proposed model. The rest of this paper is organized as follows. Section 2 addresses the literature review of inventory models related to inventory write-downs; Section 3 elaborates features of inventory write-downs in semiconductor manufacturing and

3

introduces the proposed model; Section 4 explains the proposed model; Section 5 presents the data collection and analytical results based on an empirical study; Section 6 addresses discussions on sensitivity analysis and applications of cost tornado diagrams; and lastly, Section 7 offers a conclusion.

2. Fundamentals

Semiconductor industry is very capital intensive, in which the chip makers strive to increase productivity and enhance capacity utilization for maintaining their capital effectiveness and competitive advantages (Chien et al., 2007). Indeed, manufacturing strategic decisions of semiconductor companies involve the interrelated elements including pricing strategies (P), demand forecast and demand fulfillment planning (D), capacity planning and capacity portfolio (C), capital expenditure (C), and cost structure (C) that will affect their overall financial return (R), as illustrated in the PDCCCR conceptual framework of Fig. 1 (Chien et al., 2010a). On one hand, semiconductor companies have to forecast future demands to provide the basis for capacity strategic decisions including new fab construction, technology migration, capacity transformation and expansion, tool procurement, and outsourcing. On the other hand, given demand uncertainty and forecast errors, companies often carry a safety stock in terms of the days of in the semiconductor supply chain. As shown in the Bullwhip Effect (Lee et al., 1997), the variations are amplified as moving upstream in the supply chain. Thus, it is critical for semiconductor companies to develop robust demand fulfillment strategies and manage the inventory to mitigate the negative impacts of the Bullwhip Effect. However, the demand fluctuation due to shortening product life cycle and increasing product diversification in the consumer electronics era make the demand forecast problem increasingly difficult (Chien and Chen, 2011), which complicates the present problem for inventory write-down prediction considering inventory age, product characteristics, and cost structure, according to the accounting principle. Figure 1. Conceptual Framework of PDCCCR (Chien et al. 2010a) The characteristic of a product value decreasing over time has been studied in inventory management. Recent research has considered perishability, deterioration, and obsolescence, and developed optimal inventory policies accordingly. Agricultural goods are among the classic products facing this problem (Blackburn and Scudder, 2009; Lodree Jr. and Uzochukwu, 2008), and their perishability might affect the demand itself. Concerning supply chain management, obsolescence strongly affects reorder policies, which have been analyzed and compared (Deniz et al., 2010; Emsermann and Simon, 2007; Song and Zipkin, 1996). Ferguson and Koenigsberg (2007) suggested a two-period inventory model to determine the optimal production and price. The primary features for classifying these models include single/multiple items, static/varying demand, single/multiple periods, purchase/production models, with/without backordering, single/multiple buyers, and constant/changing deterioration rates (Raafat, 1991). However, inventory ages are occasionally not recorded to reflect the inventory value for a write-down. Existing studies have explored echelon (stage) inventory policy aimed at optimizing the stock levels and order quantity of each echelon supply chain to reduce stock and inventory holding costs (Clark and Scarf, 1960; Cohen et al., 1989; Graves et al., 1993). Regarding specific application areas, the inventory costs of electronic products can be divided into component devaluation costs, price protection costs, product return costs, obsolescence costs, and inventory holding costs. Different inventory costs have different proportions in different product categories. Therefore, different supply chain management policies should be used for inventory-driven cost reduction (Callioni et al., 2005).

4

Since a decline in product demand leads to a loss of orders, and thus, high stock levels, examining the product life cycle of electronic components and adjusting stock levels accordingly to reduce inventory overdue costs is necessary (Solomon et al., 2000). To further improve the evaluation of inventory management, effective information systems were suggested, including implementation of the Just-in-Time (JIT), material requirement planning (MRP), mass customization, and manufacturing flexibility and modularity strategies (Rabinovich et al., 2003). To obtain the structure of a significant bill of material (BOM), a flow-network approach was developed (Yenisey, 2006). However, the previous studies focused on ordering and production decisions to minimize total costs, which is not applicable to predicting inventory write-downs. Chen et al. (2010) investigated the causes and the effects of inventory write-downs in the semiconductor industry, which faces a relatively high likelihood of write-downs compared to other industries. A regression of inventory levels was proposed to estimate the differences between the actual inventory and expected inventory, and was found to be highly correlated to the actual write-downs. Their research suggested that inventory write-downs were related to both production-demand impairment and incentives to increase the performance of future financial statements; however, a number of questions still remain. First, the model could not assist or provide suggestions of how to bridge the gap between production and demand. Second, though abnormally high write-downs can improve the financial performance in the next period, a stable growth instead of fluctuating profit is more attractive to investors. Third, the semiconductor industry is well-known for its rapidly changing business environment; thus, long-term regression may not fully explain the current status of various product life cycles and their complicated production processes. In summary, little research has been done regarding the effect of inventory write-downs in a specific supply chain environment, hindering the correct prediction of inventory write-downs under the IFRS No. 2.

3. Features of Inventory Write-down Prediction in the Semiconductor Industry

Predicting inventory write-downs in the semiconductor industry is difficult because of complex interrelated considerations. First, the accounting principles restrict procedures for writing down inventory costs. Specifically, inventories must be written down to the net realizable value by individual items following IFRS No. 2 (2010). Inventories are not allowed to be written down on the basis of an inventory classification, such as all finished goods or all inventories in a particular operating segment. Consequently, complex product structures are required for estimation. The semiconductor manufacturing supply chain involves not only complex BOM and product substitution, but also various raw wafer release schedules (wafer start), turnaround times (lead times), production plans, safety stock strategies, and end product-demand forecasts (Denton et al., 2006). The quantity of wafer starts for different products at various times are based on sophisticated calculations in the master production schedule that respond to market dynamics and customer needs. Because lead times are typically lengthy, WIP always remains in the production system. Therefore, the current WIP levels of all products of different ages are critical initial input data to assess inventory quantity and value at specific times. By contrast, demand for the end product (finished good) is the driving force for reducing WIP levels and pushing the material flow to forward operations. In addition, master production schedules are also rolling-based to enable revision according to demand at various periods. For example, at the beginning of the first month, managers plan a six-month master production schedule that comprises the first month’s plan, frozen for execution, and plans for the later months as a reference for material preparation and other financial purposes. Nevertheless, a multi-period demand forecast can assist in assessing costs, thus setting profit targets.

5

The direction and path of the material flow is controlled according to BOM, on which each product can be transferred from a parent product to its descendants and gradually developed into end products. BOM enables the estimation of future WIP levels and distribution regarding individual production stages (echelons) and product nodes. With the technological developments of the semiconductor industry, new and old manufacturing processes can coexist and be used to produce the same products simultaneously; however, this requires recording the specific substitution relations among products. Postponement strategy is a well-known approach to delay the differentiation of numerous end products (Brown et al., 2001). Subsequently, more than one stocking stage exists in the supply chain. For example, a die-bank is the common push/pull boundary. In other words, upward stages build the WIP inventory on a die-bank while waiting for specified demand from the end-product stage. Products in different product nodes have varying operating costs because of the investment of different resources and value-added levels. Consequently, the unit cost information of the product structure is required to holistically assess the inventory costs (Hansen, 2007). However, a standard cost assessment is not straightforward for complex production systems such as semiconductor manufacturing. The standard cost equals the estimated total expenses divided by the total production quantity, which is derived from the relationship between capacity and estimated use. This use is estimated based on experience or the average of past records. Therefore, numerous factors, such as demand fluctuations and learning curves, affect use, thus influencing the standard cost assessment. As previously mentioned, accounting principles restricts the methods of writing down losses of inventory value while the inventory is aging. Following the implementation of IFRS No. 2, the last-in first-out (LIFO) rule is not allowed when assessing the inventory valuation loss allowance. Therefore, the first-in-first-out (FIFO) or weighted inventory rule should be used to predict the distribution of the WIP inventory. Finally, inventories should be measured by the lower cost or net realizable value representing market value (LCM). Accounting practices suggest the use of accrued provision rates to reflect the potential write-downs before net realizable values are reached. The accrued provision rate represents the write-down percentage of inventory costs regarding the inventory’s age. These features demonstrate the difficulties of predicting inventory write-downs.

4. Inventory Write-down Prediction Model Functions

 x  Floor of x is the largest integer not greater than x. [ x]  max( x, 0) Primitive Parameters and Sets A S T

I Is

The largest inventory age in the system Number of stages Number of planning periods Set of all products Set of products in stage s = 1, …, S. This set can be generated based on the BOM

Parameters

i Ci

Lead time flowing down to product i  I . Cost of product i

6

Dit Fi

Demand of end product

i  I S , t  1,..., T

 i , the discrete lead time can be 0  Fi  1   i   i  1 .

Downward flow rate of product i  I . Given a continuous lead time, defined as  i   i  where 0   i   i  1 . Consequently,

Pa

On-hand inventory of product i  I at the beginning of t  1,..., T ; the inventory of the substituted product is adjusted into the substitution product before this procedure. On-hand inventory of product i  I at age a  1,..., A at the beginning of t  1,..., T ; the inventory of the substituted product is adjusted into the substitution product before this procedure. Accrued provision rate for inventory at age a

Rit

Release of raw wafer i  I1 , t  1,..., T

Oit Oiat

t  1,..., T ; Sit  0, t  0 and Sit  0, i  I S

Safety stock of end product i  I S by the end of

t i

S Yi

Yield rate of product i  I

Variables

biat

Supply of product i  I, t  1,..., T at age a  1,..., A

dit

Downward flow calculation of product i  I, t  1,..., T ; initially, di

fi t

Downward material flow of product

fiat

Downward material flow of product i  I, t

git

Downward lag material flow of product i  I, t

t

 0, t  0 represents shortage

i  I, t  1,..., T ; initially, fi t  0, t  0 represents shortage  1,..., T at age a  1,..., A ; fiat  0, t  0  1,..., T ; initially, git  0, t  0 represents

shortage

g ha

Downward lag material flow of product i  I, t

 1,..., T at age a  1,..., A ; giat , t  0 Temporary material flow of a given product at age a  1,..., A ;

rit

Shortage of release of raw wafer i  I1 , t

siat

Safety stock of end product i  I S at age

wit

WIP inventory of product i  I by the end of t

t ia

 1,..., T

a  1,..., A by the end of t  1,..., T ; siat  0, t  0

wiat

 1,..., T ; wit  0, t  0 WIP inventory of product i  I at age a  1,..., A by the end of t  1,..., T

qiat

Aggregate WIP inventory of product i  I S at age a  1,..., A by the end of t

 1,..., T ;

q  0, a  0 t ia

OBS The obsolescence inventory write-down by the end of t  1,..., T t

This study develops an inventory write-down prediction model integrating five sequential polynomial-time heuristic procedures. The first four procedures collectively predict inventory distribution during the planning horizon. Then, the final procedure multiplies the WIP inventory with the unit cost and accrued provision rate according to the inventory age to estimate the temporal inventory write-down. Specifically, the first two procedures predict the flows of a no-age inventory (aggregate inventory without considering inventory age). The output comprises the downward flow calculation, material flow, WIP inventory, and safety stock inventory. The procedures recognize flow interrelation regarding product structure and planning times. The product structure depicts product relations in multiple stages, including the raw wafer stage, push/pull boundary stage, and end product stage (Fig. 2). Each stage is further divided into two substages to symbolize the beginning and end; however, the end-product stage contains an additional node to express the relationships among the downward flow,

7

demand, and safety stock. The “beginning stage” receives downward flows and distributes them into corresponding times according to the product lead time. A lagged material flow reflects situations when a portion of WIPs delay its movement and enters the “end stage” one period late. The “end stage” distributes the received material flow based on downward demand. Each downward flow must reflect the discount quantities because of the different yield rates of products (Chien and Hsu, 2011). Only the push/pull boundary stage and the end product stage maintain a static inventory, which flow downward only when demand is recognized. The end-product stage builds additional safety stock that increases inventory costs across planning times. Other stages force material distribution downard according to the relative downward demand. Additionally, the downward flow in the first stage represents the raw wafer release, which pushes the material flow downward. In summary, Procedure 1 first estimates the downward flow derived from the end product demand based on the pull strategy, whereas Procedure 2 pushes the raw wafers from the first stage to revise the material flow and inventories (Fig. 3). After all the no-age material flows and inventories are estimated, Procedure 3 specifies the age inventory distribution using the FIFO rule. In other words, the older inventories have priority to flow downward to meet the end-product demand, whereas the younger inventories remain in the upward stages as required. The following Procedure 4 aggregates inventory with flows for ease of computation. These two procedures collectively predict the WIP distribution with age information (namely age of WIP) during the planning periods. With the age WIP distribution forecast, the final Procedure 5 can be used to summarize inventory write-downs.

Procedure 1: No-age Inventory Demand Requirement (Pull Strategy) input

i , Dit , Fi , Oit , Sit , and Yi t

t

t

t

output di , f i , gi , and wi

begin for stage s = bottom stage (S) to upper stage (1) do for time t = 1 to end period (T) do for product i in I s do if s = bottom stage (S) then 

wit   Sit 1  Oit  wit 1  git i 1  Dit  Sit  (update WIP inventory) 

fi t i   Dit  Sit  Oit  wit 1  git i 1  (update the downward material flow) else 

  w  Oit  wit 1  git i 1   d tj / Y j  (update the WIP inventory) jSi   t i

fi

t i

     d tj / Y j  Oit  wit 1  git i 1   jSi 



(update the downward material flow)

end if s = S

dit i  fi t i / Fi (update the downward flow calculation) git 1  dit   f i t  (update the downward lag material flow) end for i end for t end for s

8

end begin Figure 2. No-age inventory distribution prediction Figure 3. A prediction logistics comparison between pull and push strategy procedures

Procedure 2: No-age Inventory Demand Fulfillment (Push Strategy) input  i , Di , Fi , Oi , Si , Yi , and Ri t

t

t

t

t

t

t

t

output di , f i , gi , wi , and ri

t

begin for stage s = 1 to S do for time t = 1 to T do for product i in I s do if s = 1 then

rit   dit  Rit 



dit  max(dit , Rit ) end if

fi t  Fi dit git  (1  Fi )dit if s is Push/Pull boundary or end product then if s = S (in case that there is no Push/Pull boundary) then

wit   Sit 1  Oit  wit 1  f i t i  git i 1  Dit  Sit  else

  w  Oit  wit 1  f i t i  git i 1   d tj / Y j  jSi   t i

end if s = S else (no WIP inventory in this case)

d  t j

Y j d tj

 d tj / Y j

dit , j  Si

jSi

end if s is Push/Pull boundary end for i end for t end for s end begin Procedure 3: Age Inventory Distribution input  i , Di , Fi , Yi , Si , di , f i , gi , wi , and Oia t

t

t

t

t

t

t

t

t

t

t

t

output bia , f ia , gia , ha , sia , and wia begin for stage s = 1 to S do for time t = 1 to T do

9





for product i in I s do if s = S then

bit1  dit end if

ha  Oiat , a  1,..., A for a = 1 to A then if gi  bia then t

t

git  git  biat giat  giat  biat else

giat  giat  git break for a = 1 to A end if end for a for a = A to 1 then if f i  bia then t

t

fi t  fi t  biat fiat  fiat  biat else

fiat  fiat  fi t break for a = A to 1 end if end for a

ha i  ha i  f iat i , a  1,..., A ha i 1  ha i 1  g iat i 1 , a  1,..., A if s = S then

ha  ha  siat 1 , a  1,..., A end if if s is Push/Pull boundary or end product then

ha  ha  wiat 1 , a  1,..., A end if for a = 1 to A and s = S do if Si  ha then t

Sit  Sit  ha

10

siat 1  siat 1  ha ha  0 else

siat 1  siat 1  Sit ha  ha  sit break for a = 1 to A end if end for a for a = 1 to A and (s is Push/Pull boundary or end product) do if wi  ha then t

wit  wit  ha wiat 1  wiat 1  ha ha  0 else

wiat 1  wiat 1  wit ha  ha  wit break for a = 1 to A end if end for a

biat 

Y j d tj

d jSi

t j

/ Yj

ha , j  Si

end for i end for t end for s end begin

11

Procedure 4: Age Inventory Aggregation input  i , Oia , bia , f ia , g ia , sia , and wia t

t

t

t

t

t

t

output qia begin for stage s = 1 to S do for product i in I s do for age a = 1 to A do for time t = 1 to T do

qiat  siat  wiat for time p = 1 to t-1 do

qip,a t  p 1  qip, a t  p 1  Oiat end for p for lead time  = 1 to  i do 1 t 1 t t qit, a   qi , a   f ia  g ia

end for p t i t i qit, a i 1  qi , a i 1  g ia

end for t end for a end for i end for s end begin Procedure 5: Inventory Write-down Estimation t

input qia , Ci , and Pa t

output OBS begin for time t = 1 to T do

OBSt 



iI , a 1,..., A

Pa Ci qiat qiat

end for t end begin 5. Empirical Study and Results

This study was motivated by the need for a robust inventory write-down estimation model to reduce the impact of inventory-carrying value fluctuations of a semiconductor manufacturing company in Taiwan. This integrated device manufacturer is among the world’s leading providers of Mask ROM and flash memory for various applications in consumer electronics. This case study company applies the “reserve approach” to write-down obsolete inventory. To implement and validate the proposed solution, an empirical study of this company was conducted based on real data of flash memory production. However, the company’s internal costs, production planning, and inventory data are sensitive and confidential. To oblige the nondisclosure agreement, all presented data have been transformed into relative values,

12

which do not affect the generality for further explanation. Specifically, each inventory write-down value (OBS) is presented as an OBS index, denoting the OBS estimation divided by the actual OBS of January 2010. Data collection involved actual OBSs and beginning on-hand inventories of the 12 months of 2010. The flash memory supply chain included four major production stages, among which the die-bank is the push/pull boundary in the third stage. Each stage comprised approximately 360, 580, 1,100, and 5,600 product types. The range in turnaround times (lead time) of each stage was 0 to 20 days, 7 to 12 days, 6 to 20 days, and 44 to 65 days. The average lead time to complete the entire process was approximately 3.5 months. Approximately 6,600 BOM and 1,200 substitution records existed. Each month contained approximately 2,700 types of on-hand inventories that were categorized according to product and age. Effective inventory ages ranged from 1 to over 24 months. The safety stock was 0.6 times the demand. The value of the age of inventory is partially written down once the age is greater than three months, and is fully written down when the age is greater than 18 months (see the first and second columns in Table 1). This accrued provision rate setting best represented the need to estimate the obsolescence inventory write-down (OBS) for the case company in 2010. For example, the other three arbitrarty settings resulted in OBSs with a higher deviation from the actual write-down considering lower of cost or market (LCM) values (Fig. 4). Nevertheless, applications of the provided accrued provision rates enabled OBS forecasts by predicting the inventory distribution.

Table 1. The accrued provision rates for inventory at different ages Figure 4. The relations among LCM and OBSs using different accrued provision rate scenarios for the actual inventory The proposed model using the collected data was tested on a personal computer with an Intel© Core™2 Quad CPU Q8200 and 2.33 G and 3.21 GB RAM. From January 2010, nine 4-month OBSs were predicted. Only the first three months were collected and compared with actual OBSs; that is, the sum of the actual inventory times corresponding to the accrued provision rates. The three-month forecast can be used for the quarterly financial forecast. The elapse time of each computation was less than 10 s, which was suitably efficient to generate an OBS forecast, and thus, helpful for further what-if analysis and problem solving. The computation results show that the three-month forecast can capture the trend and turning point of the actual OBS (Fig. 5). The mean absolute percent errors (MAPEs) of the first-, second-, and third-month forecasts are all less than 3.5 % (Table 2). The efficiency and effectiveness outcome supports the viability of the proposed heuristic-based model.

Figure 5. The relationships among the actual OBS and the three-month forecasts Table 2. OBS prediction accuracy for 2010 6. Discussions Based on the Sensitivity Analysis and Cost Tornado Diagram

The accuracy of OBS prediction relies on an accurate demand forecast and estimation of operating costs and yield rate. In addition, though FIFO rules are assumed for inventory estimation, execution might be different because of limited controls, which also increase the realized average inventory ages. Conversely, the different safety stock strategies also varies OBS forecasts. Therefore, the following sensitivity analysis examines and compares different settings for the presence of these factors during September 2010.

13

As shown in Figs. 6 to 10, the changing factors had a greater impact on the OBS of the third month compared to that of the first month. The impact gradually increases in conjunction with the forecast horizon. This phenomenon warns management of the delayed effect of uncertainties. A number of semiconductor memory companies discovered a significant amount of inventory write-downs too late. However, most factors linearly influence OBS, except the demand factor (Fig. 8). The OBS with a 40 % demand in the third month’s forecast declined significantly compared to the forecasts for the second month and the third month with a 60 % demand. This highlights the need to be particularly aware of demand changes.

Figure 6. Sensitivity analyis of the third-month OBS predictions of changing ages Figure 7. Sensitivity analyis of the third-month OBS predictions of changing cost rates Figure 8. Sensitivity analyis of the third-month OBS predictions of changing demand Figure 9. Sensitivity analyis of the third-month OBS predictions of changing yield rates Figure 10. Sensitivity analyis of the third-month OBS predictions of changing safety stock as a percentage of demand

The cost tornado diagram is an effective tool for summarizing sensitivity results and prioritizing important factors (Chien, 2005). The three-month OBS tornado diagrams consistently show that age and cost are the two primary factors affecting OBS (Figs. 11 to 13). In addition, lowering the age of the two-month diagram provided a relatively greater benefit than reducing 20 % of manufacturing costs did. Because reducing 20 % of cost is difficult in a matured manufacturing environment, management should establish a systematic approach to ensure the execution of FIFO rules. The results also show that demand factor is not one of the two primary influencing factors. After investigating the root causes, we discovered a significant portion of non-active inventory in the production system. The OBS value of non-active inventories are not altered concerning demand change. This is also the reason the safety stock ratio does not significantly affect OBS prediction. Finally, in the third month prediction, the demand factor becomes the third highest factor rather than the fourth, as in the previous two predictions. This result also responds to the increasing effect of demand changes.

Figure 11. Cost tornado diagram of the first month OBS prediction for September 2010 Figure 12. Cost tornado diagram of the second month OBS prediction for September 2010 Figure 13. Cost tornado diagram of the third month OBS prediction for September 2010

7. Conclusion

Because predicting inventory write-downs in the semincoductor manufacturing industry is vital but complex, a systematic model that captures critical features, including the accounting principles, inventory ages, and product structures must be developed. To provide an efficient and accurate solution, this study developed an integrated polynomial-time heuristic-based model. The empirical study shows the advantages of using the proposed model, including rapid computation less than 10 s and a MAPE less than 3.5 % for a complete production line three-month prediction. Using this efficient and effective model, management can perform comprehensive what-if analysis regarding various improvement alternatives and supply chain strategies, such as the decisions on push/pull boundaries.

14

Discussions regarding the sensitivity analysis and cost tornado diagrams suggest the priority of affecting factors. To prevent unexpected inventory write-downs, management should focus on implementing the FIFO rules, reducing manufacturing costs, and monitoring market demand changes. Additionally, management should be aware that the delayed effects of the current inadquate strategy will significantly increase write-downs in the future. Though the proposed model assumes linearity and deterministic parameters and variables, more general models can be developed by introducing simulation and evolutionary algorithms. This extension will enable the evaluation of accounting rule changes and supply chain decision effects.

15

REFERENCES

Aizcorbe, A. (2002). Why are semiconductor prices failing so fast? Industry estimates and implications for productivity measurement. Federal Reserve Board, FEDS Discussion Paper No. 2002-20. Blackburn, J. & Scudder, G. (2009). Supply chain strategies for perishable products: the case of fresh produce. Production & Operations Management, 18(2), 129-137. Brown, A.O., Ettl, M., Lin, G.Y., Pctrakian R. & Yao, D.D. (2001). “Inventory allocation at a semiconductor company,” in Song, J.-S. and Yao, D.D. eds, Supply Chain Structures: Coordination, Information and Optimization, 283-309. Callioni, G., de Montgros, X., Slagmulder, R., Van Wassenhove, L.N. & Wright, L. (2005). Inventory-Driven Costs. Harvard Business Review, 83(3), 135-141. Chen, H., Ramnath, S., Rangan, S. & Rock, S. (2010). Inventory write-downs in the semiconductor industry. The Fourth Accounting Research Conference, December 19-20, Gachibowli, Hyderabad, India. (http://www.isb.edu/AccountingResearchConference/File/InventorySrinivasaRangan.pdf) Chien, C.-F. (2005). Decision Analysis and Management, Yeh-Yeh Book Gallery, Taipei, Taiwan. (in Chinese) Chien, C.-F. (2007). Made in Taiwan: Shifting paradigms in high-tech industries. Industrial Engineer: IE, 39(2), 47-49. Chien, C.-F. & Chen, Y.-J. (2011). Manufacturing intelligence and UNISON analysis for modeling non-volatile memory products demand forecast problem in semiconductor manufacturing. Proceedings of International Conference on IML2011, 1-7. Chien, C.-F., Chen, Y.-J., & Peng, J.-T. (2010a). Manufacturing intelligence for semiconductor demand forecast based on technology diffusion and product life cycle. International Journal of Production Economics, 128(2), 496-509. Chien, C.-F., Chen, H.-K., Wu, J.-Z., & Hu, C.-H. (2007). Constructing the OGE for promoting tool group productivity in semiconductor manufacturing. International Journal of Production Research, 45(3), 509-524. Chien, C.-F., Dauzère-Pérès, S., Ehm, H., Fowler, J.W., Jiang, Z., Krishnaswamy, Lee, T.-E., Mönch, L., & Uzsoy, R. (2011). Modeling and analysis of semiconductor manufacturing in a shrinking world: challenges and successes. European Journal of Industrial Engineering, 5(3), 254-271. Chien, C.-F. & Hsu, C. (2011). UNISON analysis to model and reduce step-and-scan overlay errors for semiconductor manufacturing. Journal of Intelligent Manufacturing, 22(3), 399-412. Chien, C.-F., Wu, J.-Z., & Weng, Y.-D. (2010b). Modeling order assignment for semiconductor assembly hierarchical outsourcing and its decision support system. Flexible Services and Manufacturing Journal, 22(1-2), 109-139. Chien, C.-F., Wu, J.-Z., & Wu, C.-C. (2011). A two-stage stochastic programming approach for new tape-out allocation decisions for demand fulfillment planning in semiconductor manufacturing. Flexible Services and Manufacturing Journal. (DOI: 10.1007/s10696-011-9109-0) Chien, C.-F. & Zheng, J. (2011). Mini-max regret strategy for robust capacity expansion decisions in semiconductor manufacturing. Journal of Intelligent Manufacturing. (DOI: 10.1007/s10845-011-0561-1) Clark, A.J. & Scarf, H. (1960). Optimal Policies for a Multi-Echelon Inventory Problem. Management Science, 6(4), 475-490. Cohen, M.A., Kleindorfer, P.R. & Lee, H.L. (1989). Near-Optimal Service Constrained Stocking Policies for Spare Parts. Operations Research, 37(1), 104-117. Deniz, B., Karaesmen, I. & Scheller-Wolf, A. (2010). Managing Perishables with Substitution: Inventory Issuance and Replenishment Heuristics. Manufacturing & Service Operations Management, 12(2), 319-329. Denton, B.T., Forrest, J. & Milne, R.J. (2006). IBM Solves a Mixed-Integer Program to Optimize Its Semiconductor Supply Chain. Interfaces, 36(5), 386-399. Emsermann, M. & Simon, B. (2007). Optimal Control of an Inventory with Simultaneous Obsolescence. Interfaces, 37(5), 445. Ferguson, M.E. & Koenigsberg, O. (2007). How should a firm manage deteriorating inventory? Production and Operations Management, 16(3), 306-321. Graves, S.C., Rinnooy Kan, A.H.G. & Zipkin, P.H. (1993). Logistics of Production and Inventory. In: A.H.G.R.K. S.C Graves & P.H. Zipkin, Logistics of Production and Inventory (Vol. Volume 4, pp. v-vii). Amsterdam: Elsevier. Hansen, D.R. (2007). Cost Management: Accounting & Control, 6e, South-Western Cengage Learning, Mason, Ohio, USA. Kuo, C.-J., Chien, C.-F., & Chen, J.-D. (2011). Manufacturing intelligence to exploit the value of production and tool data to reduce cycle time, IEEE Transactions on Automation Science and Engineering, 8(1), 103-111. 16

Leachman, R., Ding, S. & Chien, C.-F. (2007). Economic efficiency analysis of wafer fabrication IEEE Transactions on Automation Science and Engineering, 4(4),501-512. Lee, H.L., Padmanabhan, V., & Whang, S. (1997). The bullwhip effect in supply chains, Sloan Management Review, 38(3), 93-102. Lodree Jr, E.J. & Uzochukwu, B.M. (2008). Production planning for a deteriorating item with stochastic demand and consumer choice. International Journal of Production Economics, 116(2), 219-232. International Financial Reporting Standards (IFRS) No.2. (2010). Inventories. IFRS Foundation Trustees, IFRS Interpretations Committee. Semiconductor Industry Association (SIA) (2011). Press Releases. www.sia-online.org. San Jose, CA. Raafat, F. (1991). Survey of literature on continuously deteriorating inventory models. The Journal of the Operational Research Society, 42(1), 27-37. Rabinovich, E., Dresner, M.E. & Evers, P.T. (2003). Assessing the effects of operational processes and information systems on inventory performance. Journal of Operations Management, 21(1), 63-80. Solomon, R., Sandborn, P.A. & Pecht, M.G. (2000). Electronic part life cycle concepts and obsolescence forecasting. IEEE Transactions on Components and Packaging Technologies, 23(4), 707-717. Song, J.-S. & Zipkin, P.H. (1996). Managing inventory with the prospect of obsolescence. Operations Research, 44(1), 215-222. Wu, J.-Z. & Chien, C.-F. (2008a). Modeling strategic semiconductor assembly outsourcing decisions based on empirical settings. OR Spectrum, 30(3), 401-430. Wu, J.-Z. & Chien, C.-F. (2008b). Modeling semiconductor testing job scheduling and dynamic testing machine configuration. Expert Systems with Applications, 35(1/2), 485-496. Wu, J.-Z., Chien, C.-F., Huang, Y.-S. & Huang, H.-Y. (2010). A multi-period inventory model to incorporate with inventory age, accounting principle, and product structure: A case study in a make-to-stock semiconductor integrated device manufacturer. Proceedings of the 40th International Conference on Computers and Industrial Engineering, July 25-28, Awaji Island, Japan. Yenisey, M.M. (2006). A flow-network approach for equilibrium of material requirements planning. International Journal of Production Economics, 102(2), 317-332.

17

FIGURES

Figure 1. Conceptual Framework of PDCCCR (Chien et al. 2010a)

Dit Sit

Dit  Sit

d jSi

t j

/ Y j or Dit  Sit

Sit 1

wit

Oit g it i 1

wit 1

fi t i

fi t

git dit

Figure 2. No-age inventory distribution prediction

18

d jSi

t j

d

/ Y j or Dit  Sit

jSi

wit

Oit

Oit git i 1

/ Y j or Dit  Sit

Sit 1

wit

Sit 1

t j

wit 1

f i t i

g

git i 1

wit 1

f i t i

fi t

t i 1 i

git

dit i

dit

Figure 3. The prediction logistics comparison between pull and push strategy procedures

1.200 APR S1 APR S3 LCM

1.100

APR S2 APR S4

OBS Index

1.000 0.900 0.800 0.700 0.600 1

2

3

4

5

6

7

8

9

10

11

12

Month (2010) Figure 4. The relationships among LCM and OBSs using different accrued provision rate scenarios for actual inventory

19

1.200

Actual

FCST 01

1.150

FCST 02

FCST 03

1.100

FCST 04

FCST 05

FCST 06

FCST 07

FCST 08

FCST 09

OBS Index

1.050 1.000 0.950 0.900 0.850 0.800 0.750 1

2

3

4

5

6

7

8

9

10

11

12

Month (2010) Figure 5. The relationship among the actual OBS and three-month forecasts

Age - 2 Age - 1 Age + 0 Age + 1

1.4

OBS Index

1.2

Age + 2

1.0 0.8 0.6 1 2 3 Month of Prediction Starting from 2010.09

Figure 6. Sensitivity analysis of the three-month OBS predictions of changing ages

20

Cost * 0.80

1.3

Cost * 0.85 Cost * 0.90 Cost * 0.95 Cost * 1.00 Cost * 1.05 Cost * 1.10 Cost * 1.15 Cost * 1.20

OBS Index

1.2 1.1 1.0 0.9 0.8 0.7 1

2 3 Month of Prediction Starting from 2010.09

Figure 7. Sensitivity analysis of the three-month OBS predictions of changing cost rates

Demand * 0.6 Demand * 0.8 Demand * 1.0 Demand * 1.2 Demand * 1.4 Demand * 1.6 Demand * 1.8 Demand * 2.0 1

2 3 Month of Prediction Starting from 2010.09

Figure 8. Sensitivity analysis of the three-month OBS predictions of changing demand

Yield * 0.80 1.2

Yield * 0.85 Yield * 0.90

1.1 OBS Index

OBS Index

Demand * 0.4 1.30 1.25 1.20 1.15 1.10 1.05 1.00 0.95 0.90 0.85

Yield * 0.95 Yield * 1.00 Yield * 1.05 Yield * 1.10 Yield * 1.15 Yield * 1.20

1.0 0.9 0.8 1 2 3 Month of Prediction Starting from 2010.09

Figure 9. Sensitivity analysis of the three-month OBS predictions of changing yield rates 21

SafetyRatio: 0.2

1.10

SafetyRatio: 0.4 SafetyRatio: 0.6

OBS Index

1.05

SafetyRatio: 0.8 SafetyRatio: 1.0

1.00

0.95

0.90 1

2 3 Month of Prediction Starting from 2010.09

Figure 10. Sensitivity analysis of the three-month OBS predictions of changing safety stock as a percentage of demand

OBS Index of the 1st Month Prediction 0.7

0.8

0.9

1.0

1.1

1.2

Age - 2

+2

Factor

Cost x 0.8

Yield

1.3

x 1.2

x 0.8

x 1.2

Higher Lower

Demand

Safety

x 2.0

0.2

x 0.4

1.0

Figure 11. Cost tornado diagram of the first month OBS prediction for September 2010

22

OBS Index of the 2nd Month Prediction 0.7

0.9

1.0

1.1

1.2

1.3

-2

Age

+2

x 0.8

Cost

Factor

0.8

x 1.2

x 0.8

Yield

Higher

x 1.2

Lower x 0.4

x 2.0

Demand

0.2

Safety

1.0

Figure 12. Cost tornado diagram of the second month OBS prediction for September 2010

OBS Index of the 3rd Month Prediction 0.8

0.9

1.0

1.1

1.2

1.3

1.4

+2

Age -2

Factor

Cost

Demand

x 0.8

x 1.2

x 2.0

x 0.4

Higher Lower

Yield

Safety

x 0.8

x 1.2

0.2

1.0

Figure 13. Cost tornado diagram of the third month OBS prediction for September 2010

23

TABLES Table 1. Scenarios of accrued provision rates for inventory at different ages age S1 S2 S3 S4 1 0% 0% 0% 0% 2 0% 0% 0% 0% 3 0% 0% 0% 0% 4 2% 2% 0% 0% 5 2% 2% 0% 0% 6 2% 2% 0% 0% 7 10 % 5% 0% 0% 8 10 % 10 % 0% 0% 9 10 % 15 % 0% 0% 10 20 % 20 % 2% 0% 11 20 % 25 % 2% 0% 12 20 % 30 % 2% 2% 13 50 % 35 % 40 % 2% 14 50 % 40 % 50 % 2% 15 50 % 45 % 60 % 90 % 16 50 % 50 % 70 % 100 % 17 50 % 55 % 80 % 100 % 18 50% 60 % 90 % 100 % 18+ 100 % 100 % 100 % 100 %

Table 2. OBS prediction accuracy for 2010 Month of Absolute Percent Error of Prediction OBS prediction 1 2 3 4 5 6 7 8 9 First month 1.75 % 2.72 % 5.18 % 4.91 % 2.35 % 5.06 % 2.51 % 3.31 % 0.29 % Second month 2.65 % 3.39 % 2.39 % 0.79 % 0.81 % 2.98 % 4.10 % 2.42 % 0.06 % Third month 2.55 % 1.18 % 3.03 % 2.94 % 2.39 % 1.94 % 1.93 % 2.51 % 3.40 %

24

MAPE 3.12 % 2.18 % 2.43 %