px * px *

*** Default width and height in pixels. Change it to your required dimensions.**

**Contents:**

���>�� ������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������y� ��r]bjbjE�E� 4�'�'�rU�����������������������i$�! 0���������������$Q!��#N�������������^^^������^��^^^��������������^^��0! ^A$$A$^^FA$�����@r^�\������:$���! ������������������������������������������������������������������������A$���������� �: COWAN LRA MODEL FORECAST RESULTS FOR GLOBAL SEMICONDUCTOR SALES � MAY 2007 => DOWNWARD TREND IN YR-O-YR SALES GROWTH CONTINUES

Date: 7-12-07

The Cowan LRA Model, which forecasts (each month) global semiconductor sales, is a mathematically-based model exploiting linear regression analysis of the past 23 years of historical, monthly global semiconductor sales as gathered by the WSTS and published by the Semiconductor Industry Association, SIA. It is a dynamic, mathematically-pure view of near-term forecasted worldwide semiconductor sales looking out over the next five quarters. Linear regression techniques are utilized on the "appropriately transformed" monthly sales numbers that "render" the sales data highly linear and, therefore, very amenable to linear regression analysis. The numerical transformation (of 23 years of monthly, actual sales results -- from 1984 to 2006) that is employed is not a complicated�mathematical expression but is quite straight forward and "makes sense physically" thereby yielding extremely high linear regression correlation coefficients approaching 0.98 and greater.

In exercising the Cowan LRA Model each month, a total of six distinct sets of linear regression parameters (of the format y = mx + b) are employed in order to determine the global semiconductor sales forecast predictions for each of the next five quarters as well as a sales forecast estimate for the next month.

�

The latest run of the Cowan LRA Model has yielded global sales forecast estimates (displayed below) covering the next five quarters -- namely from 2Q07�to 2Q08.� These latest forecast numbers are�based upon May 2007�s�actual global S/C sales of $18.719 billion as released�on July 03, 2007 by the SIA. These newly calculated revenue forecast estimates -- for 2Q07, 3Q07, 4Q07 (and 2007), 1Q08, and 2Q08�-- are detailed in the table below. Also provided is a�sales forecast estimate for next month, namely June 2007 (both an actual and a 3MMA). In addition, each of the forecasted sales numbers are compared to the model�s previous forecast results from last month with the percentage changes in the back-to-back forecast numbers shown in the right-hand column.

�

Additional graphs�are provided below that provide more insight into the latest sales forecast numbers as well as�displaying an historical perspective�to this month's sales forecast estimates.

�

The momentum indicator (see the first graph below)�again decreased from�the previous�month's momentum of�minus 1.66% to minus 3.01 percent, thereby falling further into�negative territory. This drop continues the trend of a further slowing in the sales growth rate over the rest of 2007. (Note - the momentum indicator is defined as the percent difference between the actual sales number for a given month -- in this case May�s actual global sales of $18.719 billion -- and the forecasted sales estimate for May -- that is, $19.300 billion calculated and published last month. This indicator can be either positive or negative and is a measure of the deviation of the actual monthly sales number from the previous month�s modeled prediction based upon the model�s analysis of 23 years of past �historical experience.� When the momentum indicator is negative, it indicates that the industry�s sales growth vector is decreasing.�

Five of the six updated global S/C sales forecast estimates, namely 2Q07, 3Q07, 4Q07, 2007,�and 1Q08 -- as shown in the first table above -- degraded by varying percentages compared to�last�month's calculated forecast estimates as summarized in the last column - Pct Change. Only 2Q08�s updated sales forecast estimate showed a slight positive increase compared to last month�s forecast estimate result.

�

Relative to next month�s global semiconductor sales forecast estimate, the June 2007 actual global sales is forecast�to be�$23.609 billion. This yields a 3MMA (three Month Moving Average) sales forecast estimate for June of $20.236 billion which is normally reported by the SIA in its monthly press release.

�

The latest, updated forecast of year-over-year sales growth for 2007�compared to 2006's final�global semiconductor sales�of $247.716 billion is�now predicted to�be 3.0 percent�with monthly re-iterations of the model still expected over the remainder of�this year. The third graph (see below) shows the by-month evolution of the previous sales forecast estimates�for 2007 over the past ten months starting from the model's first prediction for 2007, namely�with the August 2006 sales number which was published back in the beginning of October 2006.

It should be pointed out that a given month�s global sales number published by the SIA is a "lagging indicator" since it is normally published a full�month after the fact. However, the Cowan LRA Model "turns" this lagging sales number into a "leading indicator"�by virtue of its forecast horizon covering the next five quarters. This is the "beauty" of the model and therefore makes it dynamic since it is renewed each month utilizing the most recent month�s actual global S/C sales number as published by the SIA and thus rigorously �tracks� the near-term future status of the global semiconductor industry on a real-time basis.

The next monthly update of the Cowan LRA Model will be available on Friday, August 3, 2007 following the posting of the June 2007 actual worldwide semiconductor sales number by the SIA.

The following table summarizes the year-over-year (2006 => 07) sales growth forecast estimates from various market researchers and compares them to the Cowan LRA Model results including the most recent, previous, and prior forecasted sales growth rates. As shown in the table the latest Cowan LRA Model forecasted sales growth of 3.0 percent is in the �middle of the pack� relative to the rest of the market researchers, industry associations -- with five, namely Semico, IC Insights, Gartner DQ, WSTS and SIA -- in a relatively narrow range around 2 percent.

Finally, a graph and an additional table are provided in order to display the �predictability� track record of the Cowan LRA Model over the past five years. The fourth graph, shown below, displays May�s sales forecast estimate history for each of the last five years (2002 through 2006) as well as the corresponding percent delta to the actual year end sales number. As shown the variability in the percent delta for May�s historical forecast estimates vary by year from -8.1 percent (in 2003) to -0.09 percent for last year, namely 2006. It should be highlighted that as the year progresses, typically each subsequent monthly forecast �zeros in� on the actual final sales revenue. This is illustrated by the following table which tracks each month�s forecast and percent delta to the year end actual for each of the past five years covering January through November numbers, respectively. Typically, July through September�s forecast results are a pretty good proxy for the entire year�s sales number with typically each subsequent month�s forecast estimate improving more on the predictability as the table shows.

Mike Cowan, the developer of the Cowan LRA Model, is a 40-year semiconductor industry veteran, having retired (Jan 2002) from IBM Microelectronics where he was engaged in both technical and management assignments including strategy development and competitor/competitive analysis. For more details, please contact Mike at: mikedco@attglobal.net.

The following three articles were posted by the editors of The Semiconductor Reporter in order to introduce the Cowan LRA Model to their readers as part of the kickoff of the monthly global S/C sales forecasting feature to their website (semireporter.com) back in August 2002. These articles provide additional insight into their view of the value of the Cowan LRA Model�s forecasting methodology, approach.

Introducing the Cowan LRA Model semiconductor revenue forecast

The Semiconductor Reporter, August 5, 2002, 8 a.m. EDT

Conditions change rapidly and unexpectedly in the semiconductor industry, and market forecasts are hard put to keep up with the changes. How can industry management be sure that a forecast issued two, three, or more months ago is still in line with what's happening in the market?

To address the issue of monitoring forecast accuracy given the latest data, The Semiconductor Reporter presents a new type of semiconductor market forecast, the Cowan LRA Model Forecast, developed by IBM Microelectronics veteran Mike Cowan. Based on linear regression analysis (LRA) of historical worldwide semiconductor revenue data published by the SIA and WSTS, the model is used to generate an updated forecast each month immediately following the monthly release of the SIA/WSTS monthly revenue number.

About Mike Cowan (side bar)

Mike Cowan has a 36-year history at IBM's Microelectronics Division in East Fishkill, N.Y., where he was involved in many facets of semiconductor development and manufacturing engineering, including both technical and management responsibilities.

Over the past ten years, as a senior technical staff member, he has been involved in strategy development and competitive analysis focused on the semiconductor industry, and has developed a number of top-down and bottom-up models to predict the dynamics of the semiconductor industry.

Cowan earned BS and MS degrees in physics at Wayne State University in Michigan, and an MS in electrical engineering at Syracuse University in New York.

The proprietary Cowan LRA Model generates forecast estimates for worldwide semiconductor revenues -- an estimate of what the following month's number will be, as well as estimates for the next four quarters and the current calendar year. The estimates are derived using a dynamic model based on historical actual monthly semiconductor revenues over the past 18 years. "The model incorporates both the cyclic and the seasonal nature of the actual historical data and relies on statistical analysis of the historical data employing linear regression analysis of certain combinations of the actual data," Cowan said.

All market researchers revise their forecasts from time to time, either on a schedule or on an as-needed basis when they see actuals drifting too far off the predicted course. Tellingly, they seldom refer to their previous forecast numbers, which are now "wrong," when they announce their latest forecast.

Cowan, on the other hand, will publish his new forecast every month in a running tally matrix that shows the evolution of revenue forecasts for each quarter over time as each month's new actual-revenue data is fed into the model, refining its "view" of the future. These dynamic estimates are continuously changing, typically by a small amount each month. The method reflects the reality that as a future quarter gets closer, revenues for that quarter can be more accurately predicted -- based on current and historical data.

Based on worldwide semiconductor revenue data released Friday for the month of June 2002 ($13.276 billion), the Cowan LRA Model is currently forecasting estimated 2002 total worldwide semiconductor revenues of $137.05 billion, a 1.4% decrease from the 2001 total. The revenue forecast figures are all up slightly from the forecast generated by the model based on May numbers.

Turning points in the market can be pinpointed in Cowan's running tally, as the "arrow" shifts from up to down -- when the monthly adjustment to the revenue forecast stops pushing the number up and starts pushing it down (and vice versa). Looking at the model's predictions as monthly actuals are fed into it sequentially from 2001 and going into 2002, for example, the turning point was observed in the February 2002 estimates for upcoming quarters: they started getting bigger when February's actuals were factored in.

Even though the model is a new one, its "predictions" for previous quarters and years can be seen by stepping it through the historical data and comparing its forecasts to actual outcomes. Mr. Cowan is willing to share historical "track record" tables with interested readers, and is interested in hearing feedback and fielding questions. In addition, those who wish to dig deeper into the picture developed by the Cowan LRA Model may be interested in studying Mr. Cowan's Running Tally Matrix spreadsheet. Please email him at mikedco@attglobal.net.

What's the point of a dynamic forecasting model? Who needs a forecast that won't sit still?

By Jeremy Young; The Semiconductor Reporter, August 5, 2002, 8 a.m. EDT

There are numerous market analysts applying a range of methodologies to the task of predicting what the semiconductor market is going to do in the near- and long-term future. Why do we need another one? In particular, why do we need a dynamic forecast that changes every month?

After all, most corporations go through one major forecasting exercise each year, and then make adjustments once or twice during the year if conditions change. Who could possibly take on the task of rejiggering their company's plans -- every month? Ouch!

We all hope that will not become necessary. However, in reality the market itself is in a constant state of flux, and forecasts that are made only once every six months, or even every quarter, cannot hope to stay in close step with the market's moves.

Here's the point: even if we don't tear up the budget and the production plan every month and do it over, it pays to check in frequently to make sure that the market reality has not made all our plans obsolete. We may have a forecast that we hang our hat on, and it may be the forecast of record for many months in a row. But if things keep changing, we need to keep checking to see whether market pressures are pushing the numbers outside the boundaries we set for it. We need to know at the earliest possible point when it does become time to tear up the plan and make a new one.

That's why we got excited when we started looking at Mike Cowan's LRA Model forecast. Here is an impressive forecasting methodology based on linear regression analysis that swallows another data point each month when the SIA releases the monthly revenue figures, and instantly puts out a revised forecast. Here is a forecast approach that accepts the reality that things constantly change: the forecast is dynamic, and as we watch it over time, we expect not only to glean important information from its latest forecast, but from the changes that are occurring in the numbers it puts out each month.

How can anything but a dynamic forecast remain true to a dynamic marketplace?

Keep calibrated

Readers who are deeply involved in analyzing market trends will want to examine the details of Mr. Cowan's LRA Model forecast estimates, and analyze its dynamic behavior. They may also want to have a look at his monthly Running Tally Matrix, which lays out in detail how the new data, including monthly revenue figures and revisions to previous months' revenue numbers, produce changes in the forecast numbers. (Please email your request for the Running Tally Matrix to us at editor@semireporter.com.)

Readers who don't have the time to get into it that deeply may simply want to check the forecast once a month to see if the model has spotted a significant market shift or a turning point in the market's progress -- to see if it's just "steady as she goes," or whether something new is showing up in the numbers that bear closer attention.

Market forecasters may want to keep an eye on the monthly forecast changes to get an early handle on whether they need to look at their numbers again. In a sense, the Cowan LRA Model Forecast is a way to keep other forecasts calibrated.

VLSI Research, for one, has recently said that it will update its forecasts every month -- in recognition, we believe, of the same market reality that makes the Cowan LRA Model useful. Market researchers are not actually clairvoyant. And as new data comes out, they need to continuously check and adjust.

We hope you will find The Cowan LRA Model Forecast a useful early warning system for your forecasting and planning process. With luck, you won't have to change course more often -- but it could make you quicker on your feet when the need does arise.

'Back-to-the-future' forecasts aim to clear up crystal ball

By J. Robert Lineback, The Semiconductor Reporter; Sept. 4, 2002, 10 a.m. EDT

After seeing mixed results from chip industry forecasts in the past 10 years, Mike Cowan decided to take a crack at finding new statistical approaches to predicting semiconductor sales based on widely available data and as few assumptions as possible.

The challenge was transforming the most reliable data available--historic sales numbers compiled by the World Semiconductor Trade Statistics (WSTS) organization--into some kind of reliable and useful leading indicator for market trends.

"I went back to the basics, and just looked at historical data--the month's actual sales data as it is reported by WSTS and the SIA," said Cowan, who retired from IBM Corp.'s Microelectronics Division in January after 36 year at the company. "After looking at what I call the fine structure of that data, I discovered that with the right combinations of the data, it could be applied to linear regression analysis and used as the basis of forecasting."

The result is the Cowan LRA Model, which is being used each month on The Semiconductor Reporter to update a dynamic forecast for chip revenues based on historical sales data released by the Semiconductor Industry Association. The new forecast from Cowan's model is being issued today (see the results), based on the July sales data released on Tuesday by the SIA.

Cowan works with the single-month sales data, and his proprietary model incorporates both the cyclic and the seasonal nature of the actual historical data collected over the past 18 years. Linear regression analysis (LRA) is applied to certain combinations of the actual sales data from the 18-year period (see Aug. 5 story).

"Obviously one would like to have the sales data sooner," he said, referring to the time lag involved in collecting it from around the world and then the SIA releasing the information in the U.S. market. "It is certainly a lagging indicator, and it has been a source of frustration for a good many years."

But using all of the sales data from the past 18 years, Cowan said he has worked out the right combinations of data and calculations to turn the historical numbers into a look forward.

According to Cowan, the beauty of this approach is that it is dynamic and provides "momentum vectors" month-to-month, quarterly, and yearly, to "give you a flavor of the direction that the industry is going."

"It is a forecast estimate. No forecast is going to be accurate 100% of the time, all of the time," added Cowan, but he also believes the LRA model has proven that it is not "happenstance statistics" because it works well when stepped through the historical data from 2001--which "was certainly a pivotal year in demonstrating that things had changed," he said.

"It was 2001 when everyone [in the industry] got very frustrated. When you looked at the prognostications and laid them out in a significant way across a wide range of market researchers, you just saw that everyone was way off the mark," Cowan said.

A variety of factors have caused fundamental changes in the chip industry in recent years, including a shift in emphasis from stand-alone PCs to communications and an increase in low-cost systems for consumer markets. Sporadic growth cycles and the unprecedented lack of visibility in 2002 have caused some market research firms to revisit and rework their forecasting techniques, with a few moving to monthly updates of annual outlooks (see Sept. 3 story).

Cowan agrees that semiconductor forecasting has fallen into a certain degree of disarray following the worst chip industry downturn in history last year and the weak recovery period of 2002. Part of the problem, he suggested, can be traced back to top-down and bottom-up forecasting models that start with estimates of system shipments or wafer fab capacity and then apply assumptions against those models.

"The supply-demand approach was okay [in the 1990s], and you could get ballpark numbers, but it left a little bit to be desired," Cowan said. "But then came the big frustration in 2001."

That led Cowan back to widely available and most often historical data, which he figured could say something about the future if only tickled with the right amount of statistical analysis.

Like most researchers, Cowan protects the specifics of his linear regression analysis and models, but he said his approach "leverages existing data so that underlying assumptions are not way out in left field"--meaning off the mark like most were in 2001.

Cowan said he also has other models besides the one that works off monthly WSTS/SIA sales data. Another of his forecasting models uses widely available quarterly results from U.S.-based semiconductor companies and a revenue consensus for leading chip makers from Wall Street chip analysts.

The IBM retiree believes consensus forecasting is currently the best approach for industry projections. The WSTS group holds a worldwide consensus forecast summit twice a year for its fall and spring projections, but the biannual frequency is not enough to keep up with the changing market conditions, Cowan noted.

So he has found that building a "consensus outlook" from the projections of a half dozen or more chip analysts at brokerage firms works well as a consensus element in the forecast built off of quarterly results from U.S. companies. "This is a new technique that also features existing data in terms of providing a forecasting methodology," Cowan explained. "It cleans off the old crystal ball."

!@FJ~����������������ɺ��휊�{i^VNVCN;Ch`P�CJaJh�\!h')�CJaJh��CJaJh-t�CJaJh�Vh')�CJaJ#h�\!h�5�CJOJQJ^JaJh�5a5�CJOJQJ^JaJ#h�\!h�5�CJOJQJ^JaJhwO�5�CJOJQJ^JaJh

5�CJOJQJ^JaJhwO�5�CJOJQJ^JaJ#h��h�L�5�CJOJQJ^JaJ#h��hvR?5�CJOJQJ^JaJ#h��h')�5�CJOJQJ^JaJ��Z[�

�

�����WY��������lm)*_`����������������������������gd��gd')�����D E U c f i l r u � � � � � � � � I

J

7J`d���"Z[mx���������

t

�������������������������������ڙǙ�������h�\!h')�5�CJaJh�\!h')�0JCJaJh�\!h�H�CJaJhU9h')�CJaJh�\!CJaJh�5aCJaJh��CJaJh�\!h&_�CJaJh-t�CJaJh�\!h��CJaJh�\!h')�CJaJh��CJaJh�Y�CJaJ2t

�

�

�

�

�

�

�

�

%1389NVWfijkmnpr��������������Ǽ��������x����m�me���xZh�\!h��CJaJh�Y�CJaJh�\!hl)�CJaJh��CJaJh�\!h��CJaJh��CJaJh�\!h�CJaJh��CJaJh�\!h0r�CJaJh�\!h�OACJaJh�\!hyT�CJaJh�\!hvR?CJaJh�5aCJaJh�I�CJaJhU9h')�CJaJh�\!h')�CJaJhU9CJaJ#����������&*-9:GLYe}������ )456JK��������ǼҼ�ǩ����ҡ��҆~�s���h]h�]h�\!h�%CJaJh�\!hl)�CJaJh�\!h�H�CJaJh`P�CJaJh�I�CJaJh�CJaJh�\!h0r�CJaJh�Y�CJaJh��CJaJh�\!h]7�CJaJh�\!h�CJaJh�\!h�OACJaJh�\!h')�CJaJh�5aCJaJh�\!h��CJaJh��CJaJh��CJaJ$KOo���������������'=WYbuy������·�����v�k`�`U`J`UBh�f�CJaJhOI�h')�CJaJh�\!h�p�CJaJh�\!h')�CJaJh�\!h�CJaJh�\!h�5aCJaJh�5aCJaJh�\!h�%CJaJh�\!hyT�CJaJ/hyT�hyT�5�B*CJ

OJQJ\�^JaJ

phh�)h')�CJaJjh�5ah�#qUaJh�0�h')�CJaJh�#qh�#qCJaJh')�CJaJhcrDCJaJh�\!h0r�CJaJy������������������ 7@EHIP]bfwxyz�������������������w݂�o�o�o�o���dh�\!h��CJaJhO5ICJaJh�\!h�y�CJaJh!�CJaJhU9CJaJh!�5�CJaJh�\!hU95�CJaJh�I�5�CJaJh�I�h�I�5�CJaJh�\!hU9CJaJh�\!h')�CJaJh�\!hyT�CJaJh-

�CJaJh�I�CJaJhcrDCJaJh�\!h�p�CJaJ$z���/0IRZ]defijmv���5Jcotu�������������������������ǿ�ҷ�����ү��������������Ƿ���ڷ��vhOI�h��CJaJh�\!h')�CJaJh�\!hvR?CJaJhw;GCJaJh>P�CJaJh�f�CJaJh�{�CJaJh!�CJaJh�MCJaJh�\!h�p�CJaJh%DPCJaJhO5ICJaJh�I�CJaJh!�h�$ CJaJh�\!h�$ CJaJ.�

#@AKWXY]bghu{�����������������������������������ͺ�����������ݜ���ݑ��y��ݤh�R"CJaJh�CJaJhw;GCJaJhOI�h')�CJaJhOI�CJaJh�\!h�p�CJaJh�\!h

T�CJaJh�\!hvR?CJaJh%DPCJaJh�

CJaJh�MCJaJh�\!h')�CJaJhp*[CJaJh�IOCJaJh�{�CJaJ/����

&5Rjrsz}����������>RSTs�������=>MQ�����������������������ڴ������҉�ҙ�������v��h�\!h�E�CJaJh5GCJaJh�p�CJaJh�R"CJaJh�\!h�)sCJaJh>P�CJaJh�MCJaJh�\!h

T�CJaJhOI�h')�CJaJhp*[CJaJh�{�CJaJh�

CJaJh�CJaJh�\!h�p�CJaJh�\!h')�CJaJ.Q�������

!py�����"%>?�������

*0Lkl����������������������������xpxpxpx�x�h@�CJaJh�1LCJaJh�\!h�h�CJaJh�\!h�L2CJaJh�\!h�OACJaJh�D�CJaJh�\!h�p�CJaJh�CJaJh�MCJaJh�]�CJaJhOI�h�o�CJaJh�\!h�o�CJaJh5GCJaJh�\!h�)sCJaJh�\!h')�CJaJ(lm�������

%()*+e���������'()cxy�������������꼱꩞��������~��v~��׆�~nfvfvfh�XCJaJh@�CJaJh�#qCJaJh��CJaJh�CJaJh�1LCJaJh�C�CJaJh�U5h�1LCJaJh')�CJaJh�\!hp*[CJaJhp*[CJaJh�\!h�)sCJaJh��CJaJh5GCJaJh�\!h�0�CJaJh�\!h�L2CJaJhOI�h�L2CJaJ(������"*\_`abcglqsx{���� ( + = F � � !!!-!Z!!�!�!�!�!"%"E"y"��������������ƻ������������𛃛����������heu�CJaJh�MCJaJh�]�CJaJh�_"CJaJh�cCJaJh�C�CJaJhH!�CJaJh�C�h�C�CJaJh�C�CJ

aJ

h3;$j��h�"h�#qUh�U5h�1LCJaJh�1LCJaJh�XCJaJh�#qCJaJ1`bc�#�#�#�#�#�#�#�#�#�#�#�#,%-%�&'='W(T*p*���������������������� $�da$gd3;$�xgd�V 7$8$H$gd�]�gd�V$a$gd�`gd�C�$a$gd�1Ly"�"�"�"�"�"�"�"�"�"##&#+#T#c#h#�#�#�#�#�#�#�#�#�#�#�#�#�#�#�#�#�#�����������������������������zvk^j4h�5ah�5aUaJh�)shN(xCJaJh�y�jHqh�5ah�I�UhcrDh�CJaJh3;$h�C�CJ

aJ

j��RsRGB���׀IDATx^��Ϯ�Ƒ����{��[/ ����p�6j&yb��Ln���3�Ϻt�eԒ_*����s���h���������%

$�Lv|��s�B䓎�7��/��5S�ｗ�~�S���~t

��(�O�7��v��K~:�'������=��B�{�t��Ώ�5�-�(6x���Gϩ_��{Z�!=T���'���S�Ԛ��_\��I�S��,KN�@q�J\�������'?z)Rq���#:��_����YGV�O�k85�R��Kq����@��C�ɔ�Gs�p�y�5�4ϰk��:��3;/O�Oo���z����O���01��返ὄ��r�L�\��?�e�

E�����Vݶ|���Fn��w�~y��>��eS��:���ȹ������骖KT���RN9���e�f,

�����eE���pŽ��s��y�rj��������k����K�����������N�j��~��5��

��u�"�}��x���z��~K�e�Z�\���@v�E������O��jH� ������U�����\�G��v���_nk��2s�{�N��9�j�t9�ru����{���v�ˠ�{���u�����F���}5��bė����/����~2�#�F�uS�2JƩ��SF�j�Y{f��: ���{b�ߔ�����vJ���ɷ�����F������|�;�P��y+�5H�u����B��s.5pq�>�u~x9Wz����5�P0���(ż_���H��}�/��}���'���bM=�Kc����q3�^�Zp}a�=g��i��������߹ew��n�'~�w����'H���d��Z���:|!���������ʎ&��'�?�Z�}كTq�%����1��x?=����2K�G'�D�a!�{����2iJf*

X��D��7���f�-�JO�듦/!�=_5�Q��С�(�A��3��է�4�j"��+��[�LR�Um/=�o����^��� �쉿�z�f�-D_(����h�����{Q�-��b��]2��½��×��2�\�˩�~����GwIs�iYM�.ҍKc��>=�Ώݝ�֜M��/ķXOTȼ��ߛf��P�S�y���C�;��;\���=��t����S]Ĝ@Y��/W�]M�ec/�0K����19��y�B���#��n�_f�p=_�GB�]�7������Y���/

��'��M�G���s����s���M}�$���ay�E�JE8^l �ɾ,�A��+�3�\nkj_ǈ�%��D��V������h��%|!̃A�7���#O���[e,�u�77c-�͙����)q?QU|a��&;t�ԡ�gNJ��Y-��[�ÃF�a�۳:/�վ�/�&#.���|+��_ɻk�����~u�'��Q�Xf��?[]���2�@6��V��b��h^M���p{fT< ��:�ۣ���s�,�WU-ͮ��

�#������{6�:��gD�[T~9ͬ��ܟ>T�����G�M�Ge���D�����3Vq%'#q�����$�8��K�����)f�w9:ĵ��

x}r�x����w���r�:\TZaG�*�y8I�j�bR��c|XŻ�ǿ�I

u3KG�nD1�NIB�s���

��R��u���K>V�.EL+M2�#'�f��i~�V��vl�{u8��z��H�

�*���:�(W�☕

~��J��T�e\O*�tHG��HY��}KN��P�*ݾ˦���TѼ�9/#��A7�qZ��$*c?���qU��n��w�N��%��O��i�4=3ڗP��

����1�P�m\\9���������Mؓ�2a�D�]�;Yt�\����[x������]�}Wr��|�]��g-���

eW�

�)6-r��CS�j��

i�d �DЇA�ΜIqbJ#x�꺃6k���#��A�Sh��&ʌt(Q�%��p%m��&]�caSl=�X�������\P�1�Mh�9�M��V�dDA��aV�B��[݈fJ�íP|8�քA�V^��f

�H��n���-��"�d>�z���n��Ǌ� �ة�>�b���&����2��v����Kyϼ���D:����,AGm��\nz��i�Ù��.uχYC�6�OMf��3o�r��$��5�����NH�T[XF64�T,ќ���M0�E)`#�5�XY�`�פ;��%�1�U�٥m;���R>QD����D�cp�U�'��&LE�/p���m���%]�����8fi��r�S4�d7y\�`�J�n���ί�I�R���3U�~7+���#��m�q�BiD����������i*�L6�9��m�Y&�����i���HE��=(K&�N!V��.K�e�LD�ĕ�{D ����vEꦚde��NƟ��e�(�MN9ߜR�6����&3(��a����/D��U�z�

Date: 7-12-07

The Cowan LRA Model, which forecasts (each month) global semiconductor sales, is a mathematically-based model exploiting linear regression analysis of the past 23 years of historical, monthly global semiconductor sales as gathered by the WSTS and published by the Semiconductor Industry Association, SIA. It is a dynamic, mathematically-pure view of near-term forecasted worldwide semiconductor sales looking out over the next five quarters. Linear regression techniques are utilized on the "appropriately transformed" monthly sales numbers that "render" the sales data highly linear and, therefore, very amenable to linear regression analysis. The numerical transformation (of 23 years of monthly, actual sales results -- from 1984 to 2006) that is employed is not a complicated�mathematical expression but is quite straight forward and "makes sense physically" thereby yielding extremely high linear regression correlation coefficients approaching 0.98 and greater.

In exercising the Cowan LRA Model each month, a total of six distinct sets of linear regression parameters (of the format y = mx + b) are employed in order to determine the global semiconductor sales forecast predictions for each of the next five quarters as well as a sales forecast estimate for the next month.

�

The latest run of the Cowan LRA Model has yielded global sales forecast estimates (displayed below) covering the next five quarters -- namely from 2Q07�to 2Q08.� These latest forecast numbers are�based upon May 2007�s�actual global S/C sales of $18.719 billion as released�on July 03, 2007 by the SIA. These newly calculated revenue forecast estimates -- for 2Q07, 3Q07, 4Q07 (and 2007), 1Q08, and 2Q08�-- are detailed in the table below. Also provided is a�sales forecast estimate for next month, namely June 2007 (both an actual and a 3MMA). In addition, each of the forecasted sales numbers are compared to the model�s previous forecast results from last month with the percentage changes in the back-to-back forecast numbers shown in the right-hand column.

�

Additional graphs�are provided below that provide more insight into the latest sales forecast numbers as well as�displaying an historical perspective�to this month's sales forecast estimates.

�

The momentum indicator (see the first graph below)�again decreased from�the previous�month's momentum of�minus 1.66% to minus 3.01 percent, thereby falling further into�negative territory. This drop continues the trend of a further slowing in the sales growth rate over the rest of 2007. (Note - the momentum indicator is defined as the percent difference between the actual sales number for a given month -- in this case May�s actual global sales of $18.719 billion -- and the forecasted sales estimate for May -- that is, $19.300 billion calculated and published last month. This indicator can be either positive or negative and is a measure of the deviation of the actual monthly sales number from the previous month�s modeled prediction based upon the model�s analysis of 23 years of past �historical experience.� When the momentum indicator is negative, it indicates that the industry�s sales growth vector is decreasing.�

Five of the six updated global S/C sales forecast estimates, namely 2Q07, 3Q07, 4Q07, 2007,�and 1Q08 -- as shown in the first table above -- degraded by varying percentages compared to�last�month's calculated forecast estimates as summarized in the last column - Pct Change. Only 2Q08�s updated sales forecast estimate showed a slight positive increase compared to last month�s forecast estimate result.

�

Relative to next month�s global semiconductor sales forecast estimate, the June 2007 actual global sales is forecast�to be�$23.609 billion. This yields a 3MMA (three Month Moving Average) sales forecast estimate for June of $20.236 billion which is normally reported by the SIA in its monthly press release.

�

The latest, updated forecast of year-over-year sales growth for 2007�compared to 2006's final�global semiconductor sales�of $247.716 billion is�now predicted to�be 3.0 percent�with monthly re-iterations of the model still expected over the remainder of�this year. The third graph (see below) shows the by-month evolution of the previous sales forecast estimates�for 2007 over the past ten months starting from the model's first prediction for 2007, namely�with the August 2006 sales number which was published back in the beginning of October 2006.

It should be pointed out that a given month�s global sales number published by the SIA is a "lagging indicator" since it is normally published a full�month after the fact. However, the Cowan LRA Model "turns" this lagging sales number into a "leading indicator"�by virtue of its forecast horizon covering the next five quarters. This is the "beauty" of the model and therefore makes it dynamic since it is renewed each month utilizing the most recent month�s actual global S/C sales number as published by the SIA and thus rigorously �tracks� the near-term future status of the global semiconductor industry on a real-time basis.

The next monthly update of the Cowan LRA Model will be available on Friday, August 3, 2007 following the posting of the June 2007 actual worldwide semiconductor sales number by the SIA.

The following table summarizes the year-over-year (2006 => 07) sales growth forecast estimates from various market researchers and compares them to the Cowan LRA Model results including the most recent, previous, and prior forecasted sales growth rates. As shown in the table the latest Cowan LRA Model forecasted sales growth of 3.0 percent is in the �middle of the pack� relative to the rest of the market researchers, industry associations -- with five, namely Semico, IC Insights, Gartner DQ, WSTS and SIA -- in a relatively narrow range around 2 percent.

Finally, a graph and an additional table are provided in order to display the �predictability� track record of the Cowan LRA Model over the past five years. The fourth graph, shown below, displays May�s sales forecast estimate history for each of the last five years (2002 through 2006) as well as the corresponding percent delta to the actual year end sales number. As shown the variability in the percent delta for May�s historical forecast estimates vary by year from -8.1 percent (in 2003) to -0.09 percent for last year, namely 2006. It should be highlighted that as the year progresses, typically each subsequent monthly forecast �zeros in� on the actual final sales revenue. This is illustrated by the following table which tracks each month�s forecast and percent delta to the year end actual for each of the past five years covering January through November numbers, respectively. Typically, July through September�s forecast results are a pretty good proxy for the entire year�s sales number with typically each subsequent month�s forecast estimate improving more on the predictability as the table shows.

Mike Cowan, the developer of the Cowan LRA Model, is a 40-year semiconductor industry veteran, having retired (Jan 2002) from IBM Microelectronics where he was engaged in both technical and management assignments including strategy development and competitor/competitive analysis. For more details, please contact Mike at: mikedco@attglobal.net.

The following three articles were posted by the editors of The Semiconductor Reporter in order to introduce the Cowan LRA Model to their readers as part of the kickoff of the monthly global S/C sales forecasting feature to their website (semireporter.com) back in August 2002. These articles provide additional insight into their view of the value of the Cowan LRA Model�s forecasting methodology, approach.

Introducing the Cowan LRA Model semiconductor revenue forecast

The Semiconductor Reporter, August 5, 2002, 8 a.m. EDT

Conditions change rapidly and unexpectedly in the semiconductor industry, and market forecasts are hard put to keep up with the changes. How can industry management be sure that a forecast issued two, three, or more months ago is still in line with what's happening in the market?

To address the issue of monitoring forecast accuracy given the latest data, The Semiconductor Reporter presents a new type of semiconductor market forecast, the Cowan LRA Model Forecast, developed by IBM Microelectronics veteran Mike Cowan. Based on linear regression analysis (LRA) of historical worldwide semiconductor revenue data published by the SIA and WSTS, the model is used to generate an updated forecast each month immediately following the monthly release of the SIA/WSTS monthly revenue number.

About Mike Cowan (side bar)

Mike Cowan has a 36-year history at IBM's Microelectronics Division in East Fishkill, N.Y., where he was involved in many facets of semiconductor development and manufacturing engineering, including both technical and management responsibilities.

Over the past ten years, as a senior technical staff member, he has been involved in strategy development and competitive analysis focused on the semiconductor industry, and has developed a number of top-down and bottom-up models to predict the dynamics of the semiconductor industry.

Cowan earned BS and MS degrees in physics at Wayne State University in Michigan, and an MS in electrical engineering at Syracuse University in New York.

The proprietary Cowan LRA Model generates forecast estimates for worldwide semiconductor revenues -- an estimate of what the following month's number will be, as well as estimates for the next four quarters and the current calendar year. The estimates are derived using a dynamic model based on historical actual monthly semiconductor revenues over the past 18 years. "The model incorporates both the cyclic and the seasonal nature of the actual historical data and relies on statistical analysis of the historical data employing linear regression analysis of certain combinations of the actual data," Cowan said.

All market researchers revise their forecasts from time to time, either on a schedule or on an as-needed basis when they see actuals drifting too far off the predicted course. Tellingly, they seldom refer to their previous forecast numbers, which are now "wrong," when they announce their latest forecast.

Cowan, on the other hand, will publish his new forecast every month in a running tally matrix that shows the evolution of revenue forecasts for each quarter over time as each month's new actual-revenue data is fed into the model, refining its "view" of the future. These dynamic estimates are continuously changing, typically by a small amount each month. The method reflects the reality that as a future quarter gets closer, revenues for that quarter can be more accurately predicted -- based on current and historical data.

Based on worldwide semiconductor revenue data released Friday for the month of June 2002 ($13.276 billion), the Cowan LRA Model is currently forecasting estimated 2002 total worldwide semiconductor revenues of $137.05 billion, a 1.4% decrease from the 2001 total. The revenue forecast figures are all up slightly from the forecast generated by the model based on May numbers.

Turning points in the market can be pinpointed in Cowan's running tally, as the "arrow" shifts from up to down -- when the monthly adjustment to the revenue forecast stops pushing the number up and starts pushing it down (and vice versa). Looking at the model's predictions as monthly actuals are fed into it sequentially from 2001 and going into 2002, for example, the turning point was observed in the February 2002 estimates for upcoming quarters: they started getting bigger when February's actuals were factored in.

Even though the model is a new one, its "predictions" for previous quarters and years can be seen by stepping it through the historical data and comparing its forecasts to actual outcomes. Mr. Cowan is willing to share historical "track record" tables with interested readers, and is interested in hearing feedback and fielding questions. In addition, those who wish to dig deeper into the picture developed by the Cowan LRA Model may be interested in studying Mr. Cowan's Running Tally Matrix spreadsheet. Please email him at mikedco@attglobal.net.

What's the point of a dynamic forecasting model? Who needs a forecast that won't sit still?

By Jeremy Young; The Semiconductor Reporter, August 5, 2002, 8 a.m. EDT

There are numerous market analysts applying a range of methodologies to the task of predicting what the semiconductor market is going to do in the near- and long-term future. Why do we need another one? In particular, why do we need a dynamic forecast that changes every month?

After all, most corporations go through one major forecasting exercise each year, and then make adjustments once or twice during the year if conditions change. Who could possibly take on the task of rejiggering their company's plans -- every month? Ouch!

We all hope that will not become necessary. However, in reality the market itself is in a constant state of flux, and forecasts that are made only once every six months, or even every quarter, cannot hope to stay in close step with the market's moves.

Here's the point: even if we don't tear up the budget and the production plan every month and do it over, it pays to check in frequently to make sure that the market reality has not made all our plans obsolete. We may have a forecast that we hang our hat on, and it may be the forecast of record for many months in a row. But if things keep changing, we need to keep checking to see whether market pressures are pushing the numbers outside the boundaries we set for it. We need to know at the earliest possible point when it does become time to tear up the plan and make a new one.

That's why we got excited when we started looking at Mike Cowan's LRA Model forecast. Here is an impressive forecasting methodology based on linear regression analysis that swallows another data point each month when the SIA releases the monthly revenue figures, and instantly puts out a revised forecast. Here is a forecast approach that accepts the reality that things constantly change: the forecast is dynamic, and as we watch it over time, we expect not only to glean important information from its latest forecast, but from the changes that are occurring in the numbers it puts out each month.

How can anything but a dynamic forecast remain true to a dynamic marketplace?

Keep calibrated

Readers who are deeply involved in analyzing market trends will want to examine the details of Mr. Cowan's LRA Model forecast estimates, and analyze its dynamic behavior. They may also want to have a look at his monthly Running Tally Matrix, which lays out in detail how the new data, including monthly revenue figures and revisions to previous months' revenue numbers, produce changes in the forecast numbers. (Please email your request for the Running Tally Matrix to us at editor@semireporter.com.)

Readers who don't have the time to get into it that deeply may simply want to check the forecast once a month to see if the model has spotted a significant market shift or a turning point in the market's progress -- to see if it's just "steady as she goes," or whether something new is showing up in the numbers that bear closer attention.

Market forecasters may want to keep an eye on the monthly forecast changes to get an early handle on whether they need to look at their numbers again. In a sense, the Cowan LRA Model Forecast is a way to keep other forecasts calibrated.

VLSI Research, for one, has recently said that it will update its forecasts every month -- in recognition, we believe, of the same market reality that makes the Cowan LRA Model useful. Market researchers are not actually clairvoyant. And as new data comes out, they need to continuously check and adjust.

We hope you will find The Cowan LRA Model Forecast a useful early warning system for your forecasting and planning process. With luck, you won't have to change course more often -- but it could make you quicker on your feet when the need does arise.

'Back-to-the-future' forecasts aim to clear up crystal ball

By J. Robert Lineback, The Semiconductor Reporter; Sept. 4, 2002, 10 a.m. EDT

After seeing mixed results from chip industry forecasts in the past 10 years, Mike Cowan decided to take a crack at finding new statistical approaches to predicting semiconductor sales based on widely available data and as few assumptions as possible.

The challenge was transforming the most reliable data available--historic sales numbers compiled by the World Semiconductor Trade Statistics (WSTS) organization--into some kind of reliable and useful leading indicator for market trends.

"I went back to the basics, and just looked at historical data--the month's actual sales data as it is reported by WSTS and the SIA," said Cowan, who retired from IBM Corp.'s Microelectronics Division in January after 36 year at the company. "After looking at what I call the fine structure of that data, I discovered that with the right combinations of the data, it could be applied to linear regression analysis and used as the basis of forecasting."

The result is the Cowan LRA Model, which is being used each month on The Semiconductor Reporter to update a dynamic forecast for chip revenues based on historical sales data released by the Semiconductor Industry Association. The new forecast from Cowan's model is being issued today (see the results), based on the July sales data released on Tuesday by the SIA.

Cowan works with the single-month sales data, and his proprietary model incorporates both the cyclic and the seasonal nature of the actual historical data collected over the past 18 years. Linear regression analysis (LRA) is applied to certain combinations of the actual sales data from the 18-year period (see Aug. 5 story).

"Obviously one would like to have the sales data sooner," he said, referring to the time lag involved in collecting it from around the world and then the SIA releasing the information in the U.S. market. "It is certainly a lagging indicator, and it has been a source of frustration for a good many years."

But using all of the sales data from the past 18 years, Cowan said he has worked out the right combinations of data and calculations to turn the historical numbers into a look forward.

According to Cowan, the beauty of this approach is that it is dynamic and provides "momentum vectors" month-to-month, quarterly, and yearly, to "give you a flavor of the direction that the industry is going."

"It is a forecast estimate. No forecast is going to be accurate 100% of the time, all of the time," added Cowan, but he also believes the LRA model has proven that it is not "happenstance statistics" because it works well when stepped through the historical data from 2001--which "was certainly a pivotal year in demonstrating that things had changed," he said.

"It was 2001 when everyone [in the industry] got very frustrated. When you looked at the prognostications and laid them out in a significant way across a wide range of market researchers, you just saw that everyone was way off the mark," Cowan said.

A variety of factors have caused fundamental changes in the chip industry in recent years, including a shift in emphasis from stand-alone PCs to communications and an increase in low-cost systems for consumer markets. Sporadic growth cycles and the unprecedented lack of visibility in 2002 have caused some market research firms to revisit and rework their forecasting techniques, with a few moving to monthly updates of annual outlooks (see Sept. 3 story).

Cowan agrees that semiconductor forecasting has fallen into a certain degree of disarray following the worst chip industry downturn in history last year and the weak recovery period of 2002. Part of the problem, he suggested, can be traced back to top-down and bottom-up forecasting models that start with estimates of system shipments or wafer fab capacity and then apply assumptions against those models.

"The supply-demand approach was okay [in the 1990s], and you could get ballpark numbers, but it left a little bit to be desired," Cowan said. "But then came the big frustration in 2001."

That led Cowan back to widely available and most often historical data, which he figured could say something about the future if only tickled with the right amount of statistical analysis.

Like most researchers, Cowan protects the specifics of his linear regression analysis and models, but he said his approach "leverages existing data so that underlying assumptions are not way out in left field"--meaning off the mark like most were in 2001.

Cowan said he also has other models besides the one that works off monthly WSTS/SIA sales data. Another of his forecasting models uses widely available quarterly results from U.S.-based semiconductor companies and a revenue consensus for leading chip makers from Wall Street chip analysts.

The IBM retiree believes consensus forecasting is currently the best approach for industry projections. The WSTS group holds a worldwide consensus forecast summit twice a year for its fall and spring projections, but the biannual frequency is not enough to keep up with the changing market conditions, Cowan noted.

So he has found that building a "consensus outlook" from the projections of a half dozen or more chip analysts at brokerage firms works well as a consensus element in the forecast built off of quarterly results from U.S. companies. "This is a new technique that also features existing data in terms of providing a forecasting methodology," Cowan explained. "It cleans off the old crystal ball."

!@FJ~����������������ɺ��휊�{i^VNVCN;Ch`P�CJaJh�\!h')�CJaJh��CJaJh-t�CJaJh�Vh')�CJaJ#h�\!h�5�CJOJQJ^JaJh�5a5�CJOJQJ^JaJ#h�\!h�5�CJOJQJ^JaJhwO�5�CJOJQJ^JaJh

5�CJOJQJ^JaJhwO�5�CJOJQJ^JaJ#h��h�L�5�CJOJQJ^JaJ#h��hvR?5�CJOJQJ^JaJ#h��h')�5�CJOJQJ^JaJ��Z[�

�

�����WY��������lm)*_`����������������������������gd��gd')�����D E U c f i l r u � � � � � � � � I

J

7J`d���"Z[mx���������

t

�������������������������������ڙǙ�������h�\!h')�5�CJaJh�\!h')�0JCJaJh�\!h�H�CJaJhU9h')�CJaJh�\!CJaJh�5aCJaJh��CJaJh�\!h&_�CJaJh-t�CJaJh�\!h��CJaJh�\!h')�CJaJh��CJaJh�Y�CJaJ2t

�

�

�

�

�

�

�

�

%1389NVWfijkmnpr��������������Ǽ��������x����m�me���xZh�\!h��CJaJh�Y�CJaJh�\!hl)�CJaJh��CJaJh�\!h��CJaJh��CJaJh�\!h�CJaJh��CJaJh�\!h0r�CJaJh�\!h�OACJaJh�\!hyT�CJaJh�\!hvR?CJaJh�5aCJaJh�I�CJaJhU9h')�CJaJh�\!h')�CJaJhU9CJaJ#����������&*-9:GLYe}������ )456JK��������ǼҼ�ǩ����ҡ��҆~�s���h]h�]h�\!h�%CJaJh�\!hl)�CJaJh�\!h�H�CJaJh`P�CJaJh�I�CJaJh�CJaJh�\!h0r�CJaJh�Y�CJaJh��CJaJh�\!h]7�CJaJh�\!h�CJaJh�\!h�OACJaJh�\!h')�CJaJh�5aCJaJh�\!h��CJaJh��CJaJh��CJaJ$KOo���������������'=WYbuy������·�����v�k`�`U`J`UBh�f�CJaJhOI�h')�CJaJh�\!h�p�CJaJh�\!h')�CJaJh�\!h�CJaJh�\!h�5aCJaJh�5aCJaJh�\!h�%CJaJh�\!hyT�CJaJ/hyT�hyT�5�B*CJ

OJQJ\�^JaJ

phh�)h')�CJaJjh�5ah�#qUaJh�0�h')�CJaJh�#qh�#qCJaJh')�CJaJhcrDCJaJh�\!h0r�CJaJy������������������ 7@EHIP]bfwxyz�������������������w݂�o�o�o�o���dh�\!h��CJaJhO5ICJaJh�\!h�y�CJaJh!�CJaJhU9CJaJh!�5�CJaJh�\!hU95�CJaJh�I�5�CJaJh�I�h�I�5�CJaJh�\!hU9CJaJh�\!h')�CJaJh�\!hyT�CJaJh-

�CJaJh�I�CJaJhcrDCJaJh�\!h�p�CJaJ$z���/0IRZ]defijmv���5Jcotu�������������������������ǿ�ҷ�����ү��������������Ƿ���ڷ��vhOI�h��CJaJh�\!h')�CJaJh�\!hvR?CJaJhw;GCJaJh>P�CJaJh�f�CJaJh�{�CJaJh!�CJaJh�MCJaJh�\!h�p�CJaJh%DPCJaJhO5ICJaJh�I�CJaJh!�h�$ CJaJh�\!h�$ CJaJ.�

#@AKWXY]bghu{�����������������������������������ͺ�����������ݜ���ݑ��y��ݤh�R"CJaJh�CJaJhw;GCJaJhOI�h')�CJaJhOI�CJaJh�\!h�p�CJaJh�\!h

T�CJaJh�\!hvR?CJaJh%DPCJaJh�

CJaJh�MCJaJh�\!h')�CJaJhp*[CJaJh�IOCJaJh�{�CJaJ/����

&5Rjrsz}����������>RSTs�������=>MQ�����������������������ڴ������҉�ҙ�������v��h�\!h�E�CJaJh5GCJaJh�p�CJaJh�R"CJaJh�\!h�)sCJaJh>P�CJaJh�MCJaJh�\!h

T�CJaJhOI�h')�CJaJhp*[CJaJh�{�CJaJh�

CJaJh�CJaJh�\!h�p�CJaJh�\!h')�CJaJ.Q�������

!py�����"%>?�������

*0Lkl����������������������������xpxpxpx�x�h@�CJaJh�1LCJaJh�\!h�h�CJaJh�\!h�L2CJaJh�\!h�OACJaJh�D�CJaJh�\!h�p�CJaJh�CJaJh�MCJaJh�]�CJaJhOI�h�o�CJaJh�\!h�o�CJaJh5GCJaJh�\!h�)sCJaJh�\!h')�CJaJ(lm�������

%()*+e���������'()cxy�������������꼱꩞��������~��v~��׆�~nfvfvfh�XCJaJh@�CJaJh�#qCJaJh��CJaJh�CJaJh�1LCJaJh�C�CJaJh�U5h�1LCJaJh')�CJaJh�\!hp*[CJaJhp*[CJaJh�\!h�)sCJaJh��CJaJh5GCJaJh�\!h�0�CJaJh�\!h�L2CJaJhOI�h�L2CJaJ(������"*\_`abcglqsx{���� ( + = F � � !!!-!Z!!�!�!�!�!"%"E"y"��������������ƻ������������𛃛����������heu�CJaJh�MCJaJh�]�CJaJh�_"CJaJh�cCJaJh�C�CJaJhH!�CJaJh�C�h�C�CJaJh�C�CJ

aJ

h3;$j��h�"h�#qUh�U5h�1LCJaJh�1LCJaJh�XCJaJh�#qCJaJ1`bc�#�#�#�#�#�#�#�#�#�#�#�#,%-%�&'='W(T*p*���������������������� $�da$gd3;$�xgd�V 7$8$H$gd�]�gd�V$a$gd�`gd�C�$a$gd�1Ly"�"�"�"�"�"�"�"�"�"##&#+#T#c#h#�#�#�#�#�#�#�#�#�#�#�#�#�#�#�#�#�#�����������������������������zvk^j4h�5ah�5aUaJh�)shN(xCJaJh�y�jHqh�5ah�I�UhcrDh�CJaJh3;$h�C�CJ

aJ

j��RsRGB���׀IDATx^��Ϯ�Ƒ����{��[/ ����p�6j&yb��Ln���3�Ϻt�eԒ_*����s���h���������%

$�Lv|��s�B䓎�7��/��5S�ｗ�~�S���~t

��(�O�7��v��K~:�'������=��B�{�t��Ώ�5�-�(6x���Gϩ_��{Z�!=T���'���S�Ԛ��_\��I�S��,KN�@q�J\�������'?z)Rq���#:��_����YGV�O�k85�R��Kq����@��C�ɔ�Gs�p�y�5�4ϰk��:��3;/O�Oo���z����O���01��返ὄ��r�L�\��?�e�

E�����Vݶ|���Fn��w�~y��>��eS��:���ȹ������骖KT���RN9���e�f,

�����eE���pŽ��s��y�rj��������k����K�����������N�j��~��5��

��u�"�}��x���z��~K�e�Z�\���@v�E������O��jH� ������U�����\�G��v���_nk��2s�{�N��9�j�t9�ru����{���v�ˠ�{���u�����F���}5��bė����/����~2�#�F�uS�2JƩ��SF�j�Y{f��: ���{b�ߔ�����vJ���ɷ�����F������|�;�P��y+�5H�u����B��s.5pq�>�u~x9Wz����5�P0���(ż_���H��}�/��}���'���bM=�Kc����q3�^�Zp}a�=g��i��������߹ew��n�'~�w����'H���d��Z���:|!���������ʎ&��'�?�Z�}كTq�%����1��x?=����2K�G'�D�a!�{����2iJf*

X��D��7���f�-�JO�듦/!�=_5�Q��С�(�A��3��է�4�j"��+��[�LR�Um/=�o����^��� �쉿�z�f�-D_(����h�����{Q�-��b��]2��½��×��2�\�˩�~����GwIs�iYM�.ҍKc��>=�Ώݝ�֜M��/ķXOTȼ��ߛf��P�S�y���C�;��;\���=��t����S]Ĝ@Y��/W�]M�ec/�0K����19��y�B���#��n�_f�p=_�GB�]�7������Y���/

��'��M�G���s����s���M}�$���ay�E�JE8^l �ɾ,�A��+�3�\nkj_ǈ�%��D��V������h��%|!̃A�7���#O���[e,�u�77c-�͙����)q?QU|a��&;t�ԡ�gNJ��Y-��[�ÃF�a�۳:/�վ�/�&#.���|+��_ɻk�����~u�'��Q�Xf��?[]���2�@6��V��b��h^M���p{fT< ��:�ۣ���s�,�WU-ͮ��

�#������{6�:��gD�[T~9ͬ��ܟ>T�����G�M�Ge���D�����3Vq%'#q�����$�8��K�����)f�w9:ĵ��

x}r�x����w���r�:\TZaG�*�y8I�j�bR��c|XŻ�ǿ�I

u3KG�nD1�NIB�s���

��R��u���K>V�.EL+M2�#'�f��i~�V��vl�{u8��z��H�

�*���:�(W�☕

~��J��T�e\O*�tHG��HY��}KN��P�*ݾ˦���TѼ�9/#��A7�qZ��$*c?���qU��n��w�N��%��O��i�4=3ڗP��

����1�P�m\\9���������Mؓ�2a�D�]�;Yt�\����[x������]�}Wr��|�]��g-���

eW�

�)6-r��CS�j��

i�d �DЇA�ΜIqbJ#x�꺃6k���#��A�Sh��&ʌt(Q�%��p%m��&]�caSl=�X�������\P�1�Mh�9�M��V�dDA��aV�B��[݈fJ�íP|8�քA�V^��f

�H��n���-��"�d>�z���n��Ǌ� �ة�>�b���&����2��v����Kyϼ���D:����,AGm��\nz��i�Ù��.uχYC�6�OMf��3o�r��$��5�����NH�T[XF64�T,ќ���M0�E)`#�5�XY�`�פ;��%�1�U�٥m;���R>QD����D�cp�U�'��&LE�/p���m���%]�����8fi��r�S4�d7y\�`�J�n���ί�I�R���3U�~7+���#��m�q�BiD����������i*�L6�9��m�Y&�����i���HE��=(K&�N!V��.K�e�LD�ĕ�{D ����vEꦚde��NƟ��e�(�MN9ߜR�6����&3(��a����/D��U�z�