1、A Primer on AnalysisOverviewConfidential Document,TABLE OF CONTENTS,IntroductionGeneral analytical techniquesGraphsDeflatorsRegression analysisSupply side analysisCost structuresDesign differencesFactor costsScale, experience, complexity and utilizationSupply curvesDemand side analysisCustomer under
2、standingsegmentation and “Discovery”conjoint analysismulti-dimensional scalingPrice-volume curves and elasticityDemand forecastingtechnology/substitution curvesWrap-up,LOGIC AND ANALYSIS CRITICAL TOSTRATEGY DEVELOPMENT,Key to strategy development is laying out “logic” toUnderstand what makes busines
3、s workeconomicsinteractions across competitors, segments, time, . Conceptually organize client goalsDevise ways to achieve clients goalsHelp client “make it happen”A tightly developed piece of this logic is analysisReducing complex reality to a few salient pointsIsolating important economic elements
4、,ANALYSIS IS MORE THAN NUMBER CRUNCHING,Analysis is. Integrating quantitative and qualitative knowledgeSeeing the bigger pictureThinkingcreativelyconceptuallyNot . . .Endless calculationsLetting statistics dictate/rule“Classic” scientific rigor,ANALYTICAL BIAS,“Everything can be quantified”Not reall
5、y, butMost “qualitative” effects are based in economicsexplicit or opportunity costsaccurately quantifiable or notClient hires us to analyze and objectifyQuantitative analysis is the basis,CREATIVITY AND ANALYTICAL PERSEVERANCE AREIMPORTANT TRAITS FOR SUPERIOR ANALYSTS,Strive to address a problem us
6、ing different approaches to test hypotheses and find inconsistenciesTriangulate on answersNever believe a data series blindlyNever stop at first obstacleClients often stop short of good analysis because they quickly surrender in the absence of good, readily available dataWe never surrender to the un
7、availability of dataYour case leader does not want to hear that “there is no data,” but rather what can be developed, in how much time, and at what cost,WHERE THIS PRIMER FITS,No document can teach you to be a great analystAnswers look easy, but process of getting there painfulEach problem somewhat
8、different from examplesA primer canGive flavor of expected analysesShow which analyses have been most productive historicallyExplain basic techniques and warn of common methodological errorsBest training comes fromExperience in project team workDiscussions with John Tang and othersYou are expected t
9、o locate knowledge on your own initiative,DONT LIMIT YOURSELF TO THESE TOOLS,They are a sample of the most commonly used toolsOthers will be of use in specific situationsValue management (CFROI, asset growth, etc.)Additionally, no tool can substitute for a new creative approach,TABLE OF CONTENTS,Int
10、roductionGeneral analytical techniquesGraphsDeflatorsRegression analysisSupply side analysisCost structuresDesign differencesFactor costsScale, experience, complexity and utilizationSupply curvesDemand side analysisCustomer understandingsegmentation and “Discovery”conjoint analysismulti-dimensional
11、scalingPrice-volume curves and elasticityDemand forecastingtechnology/substitution curvesWrap-up,RELATIONSHIPS HAVE MOST IMPACT WHEN DISPLAYED VISUALLY,Graphs and charts should be easily understandable to a “nonquantitative” clientDisplay one main idea per graphMake the point as directly as possible
12、Demonstrate clear relevance to accompanying material and clients businessClearly label title, axes, and sourcesTailor graph to its audience and purposeExplorationPersuasionDocumentation,CHOOSE GRAPH SCALE THOUGHTFULLY,Match chart boundaries to relevant range of the data as closely as possibleSelect
13、scale to facilitate thinking about proposed relationshipsUse same scale across charts if you intend to compare them,LINEAR VS. LOG,On a linear scale, a given difference between two values covers the same distance anywhere on the scaleOn a logarithmic scale, a given ratio of two values covers the sam
14、e distance anywhere on the scale,1,2,4,8,16,One Cycle,Linear,Log,Log,The ratio of anything to zero is infinite. Zero cannot appear on a log scale.,DATA RELATIONSHIP DETERMINES SELECTION OF SCALEThree Scales Most Common,Linear,Log,Log,Linear,Linear (usually time),Log,Linear,Semi-Log,Log-Log,Constant
15、Rate of Change,Constant Growth Rate,Constant “Elasticity”,Given no prior expectation about the form of a relationship, plot it linearly,y = mx + b,log y = mx + b,log y = mlog x + b,WHEN SHOULD A LINEAR GRAPH BE USED?,Linear graphs are best when the change in unit terms is of interest, e.g.,Market sh
16、are over timeProfit margin over timeForty-five degree downward sloping lines on linear graph represent points whose x and y values have constant sumRays through origin represent points with common ratio,Market Share (%),Linear Graph,Hardware,Software,WHEN SHOULD A SEMI-LOG PLOT BE USED?,Semi-log gra
17、phs are generally used to illustrate constant growth rates, e.g.,Volume of sales growth over time,Year,Source: Agricultural Statistics,U.S. Corn Yield (Bushels/ Acre),R=.95,Semi-Log Graph,WHEN SHOULD A LOG-LOG PLOT BE USED?,Log-log graphs are generally used to plot “elasticities,” e.g.,Price elastic
18、ity of demandScale slopeForty-five degree downward sloping lines show points with common product,Salaried and Indirect hourly Employees/ Billion Impressions of Capacity,Printing Capacity (Billions of Impressions),78% Scale SlopeR=.636,1,000,100,10,CIRCLE OR BUBBLE CHARTS OFTEN USED TO SHOW A THIRD D
19、IMENSION,Third dimension should be related to x and y axesCommon examples include:Market sizeAssetsCost flowCircle area (not diameter) is proportional,BUBBLE CHART EXAMPLECategory Growth Versus Gross Margin Versus Size,1980-84Real CAGR (%),Gross Margin (%),= $1B sales,Consumer Electronics,Toys,House
20、wares/Gifts,Jewelry,SportingGoods,SmallAppliances,Camera/Photo,Source: Discount Merchandiser,TABLE OF CONTENTS,IntroductionGeneral analytical techniquesGraphsDeflatorsRegression analysisSupply side analysisCost structuresDesign differencesFactor costsScale, experience, complexity and utilizationSupp
21、ly curvesDemand side analysisCustomer understandingsegmentation and “Discovery”conjoint analysismulti-dimensional scalingPrice-volume curves and elasticityDemand forecastingtechnology/substitution curvesWrap-up,DEFLATORS CORRECT EFFECTS OF INFLATIONConverts Variables from “Nominal” to “Real”,Time se
22、ries data in dollars with high or widely fluctuating inflation rates distort picture of growthDeflating data removes some of the distortionUsing a deflator index list, currency data are multiplied by the ratio of the base year deflator index to the data year deflator index, e.g.,1979 sales (1993 $)
23、= 1979 (1979 $) x,Deflator 1993Deflator 1979,SELECT APPROPRIATE DEFLATOR DEPENDING ONTHE QUESTION YOURE TRYING TO ANSWER,G.N.P. deflator is best for expressing dollars in terms of average real value to the rest of the economyCurrent (variable) weightsMeasured quarterlyC.P.I. is best only for express
24、ing value in relation to consumer spending on a fixed market basket of goods (1973 base)Measured monthlyIndustry or product-specific indices are best for converting dollars into measures of physical outputAvailable from Commerce Dept. for broad industry categoriesCan be constructed from client or in
25、dustry data for narrow categories,BE CAREFUL WHEN MIXING EXCHANGERATES AND INFLATION ACROSS COUNTRIES,First convert each countrys historical data to constant local currencyE.g., Japan1993 yenW. Germany1993 DMU.S.A.1993 dollarsThen convert to single currency (dollars, for example) at fixed exchange r
26、ate,EXAMPLE: AN INTEGRATED CIRCUIT MANUFACTURER,Reported SalesG.N.P. DeflatorAverage I.C.Average I.C.Year($M)(1987 = 1.00)Price ($)Transistor Price (),19877861.0001.001.0519885951.033.92.7219897301.075.99.4919908331.119.98.3419911,0621.161.90.2419921,4231.193.98.1819931,8381.2271.14.16,Reported sale
27、s $15.2%Real sales $11.4%I.C. unit sales8.9%Transistor sales52.4%,Growth Rates (per year),TABLE OF CONTENTS,IntroductionGeneral analytical techniquesGraphsDeflatorsRegression analysisSupply side analysisCost structuresDesign differencesFactor costsScale, experience, complexity and utilizationSupply
28、curvesDemand side analysisCustomer understandingsegmentation and “Discovery”conjoint analysismulti-dimensional scalingPrice-volume curves and elasticityDemand forecastingtechnology/substitution curvesWrap-up,REGRESSION ANALYSIS IS A POWERFUL TOOL FORUNDERSTANDING RELATIONSHIP BETWEEN TWOOR MORE VARI
29、ABLES,Regression analysis:Explains variation in one variable (dependent) using variation in one or more other variables (independent)Quantifies and validates relationshipsIs useful for prediction and causal explanationBut . . .Must not substitute for clear independent thinking about a problemUse as
30、single element in portfolio of analytical techniquesCan be morass“lose forest for trees”,ANY RELATIONSHIP BETWEEN VARIABLES X AND Y?,Used alone, graphical methods provide only qualitative and general inferences about relationships,PercentACV,80%,70%,60%,50%,40%,30%,20%,10%,0%,Annual Number of Purcha
31、ses by Consumer,X:Annual number of purchases by buyerY:Percent ACV,Percent ACV is the volume weighted average percent of grocery stores which carry the category.Sources: ScanTrack; IRI Marketing Factbook; BCG Analysis,REGRESSION ANALYSIS ANSWERS THESE QUESTIONS,What is relationship between X and YHo
32、w big an effect does X have on Y?What is the functional form?Is effect positive or negative?How strong is relationship?How well does X “explain” Y?How well does my model work overall?How well have I explained Y in general?Are there other variables that I should be including?,WHAT IS RELATIONSHIP BET
33、WEEN X AND Y?,PercentACV,Annual Number of Purchases by Customer,Regression fits a straight line to the data pointsPercent ACV = -0.2790 + 0.2606 annual purchasesOne more annual purchase will raise percent ACV by 0.2606 percentage pointsSlope of line (here 0.2606) indicates size of effect; sign of sl
34、ope (here positive) indicates whether effect is positive or negative,R2 = 0.69,Multiple R0.83354R Square (%)69.48Adjusted R Square (%)68.35Standard Error0.10394Observations29,Regression Statistics,Regression10.664000.6640061.4641.98146E-08Residual270.291680.01080Total280.95568,Analysis of Variancedf
35、Sum of SquaresMean SquareFSignificant F,Intercept(0.27901)0.06286(4.439)0.00013(0.40799)(0.15003)X10.260560.033247.8401.5372E-080.192370.32876,CoefficientsStandard Errort StatisticP-valueLower 95%Upper 95%,Sources: Scantrack; IRI Marketing Factbook (1990); BCG Analysis,Microsoft Excel Regression Out
36、put,HOW STRONG IS RELATIONSHIP?,t-statistic measures how well X explains YSimply calculated as slope divided by its standard error Closer slope is to zero, and/or higher standard error (variability), the weaker the relationshipA short-cut: t-statistic greater in magnitude than 2 means relationship i
37、s very strong (i.e., roughly 95% confidence level). Between 1.5 and 2, relationship is relatively strong (i.e., roughly 85-95% confidence level). Under 1.5, relationship is weak.,Multiple R0.83354R Square (%)69.48Adjusted R Square (%)68.35Standard Error0.10394Observations29,Regression10.664000.66400
38、61.4641.98146E-08Residual270.291680.01080Total280.95568,Regression Statistics,dfSum of SquaresMean SquareFSignificance F,Intercept(0.27901)0.06286(4.439)0.00013(0.40799)(0.15003)x10.260560.033247.8401.5372E-080.192370.32876,CoefficientsStandard Errort StatisticP-valueLower 95%Upper 95%,Analysis of V
39、ariance,HOW WELL DOES MY MODEL WORK OVERALL?,R2 measures proportion of variation in Y that is explained by the variables in the model - here just XIndicates overall how well model explains YBased on how dispersed the data points are around the regression lineR2 measured on scale of 0 to 100% 100% in
40、dicates perfect fit of regression line to the data pointsLow R2 indicates current model does not fit the data wellsuggests there are other explanatory factors, besides X, that would help explain Y,Multiple R0.83354R Square (%)69.48Adjusted R Square (%)68.35Standard Error0.10394Observations29,Regress
41、ion10.664000.6640061.4641.98146E-08Residual270.291680.01080Total280.95568,Regression Statistics,dfSum of SquaresMean SquareFSignificance F,Intercept(0.27901)0.06286(4.439)0.00013(0.40799)(0.15003)x10.260560.033247.8401.5372E-080.192370.32876,CoefficientsStandard Errort StatisticP-valueLower 95%Upper
42、 95%,Analysis of Variance,USE MULTIPLE REGRESSION TO SORT OUT EFFECTSOF SEVERAL INFLUENCES,UseWhen several factors have an impact simultaneouslyTo help distinguish cause from correlationDont use as “fishing expedition”,MULTIPLE REGRESSION CAN ENHANCEPREDICTIVE ABILITY,% ACV with Features and/or Disp
43、lays,Brand Size,Percent of Households Buying,Annual Number of Purchases per Year,% ACV with Features and/or Displays,% ACV with Features and/or Displays,Brand Size ($M),Percent of Households Buying,Annual Number of Purchases/Year,R=.67,R=.51,R=.69,R=.87,Predicted % ACV with Features and/or Displays,
44、Actual % ACV with Features and/or Displays,Brand Size, Reach, andPurchase Freqency,Sources: Scantrack; IRI Marketing Factbook 1990; BCG Analysis,OTHER REGRESSION EXAMPLES,Very Low R*,PercentACV,U.S. Corn Yield (Bushels/ Acre),U.S. Corn Yield (Bushels/ Acre),Retailer Margin on Deal,Average Number of
45、Days on Deal,Total Annual Purchases (M),Negative Slope*,Nonlinear Raw Data*,After Log Transformation*,*Sources: IRI Marketing Factbook; Certified Price Book; Nielsen; BCG Analysis*Source: Agricultural Statistics,R=.64,R=.002,R=.95,QUESTIONS TO ASK BEFORE RUNNING A REGRESSION,Which variable is the pr
46、edictive (or dependent) variable?Often straightforward but sometimes requires thoughtConsider direction of causationWhat explanatory variables do I believe are appropriate to include?Avoid spurious correlationsthink independently about what factors are logical to includeAvoid including explanatory v
47、ariables that are highly correlated with each otherShould the regression have an intercept term?How far can the data be reasonably extrapolated?Should the regression line cut through the origin?Does a zero value of explanatory variable imply a zero value for predictive variable?Have I plotted the data?Watch out for outliersLook for form of data (linear, exponential, power, etc.)Do I have enough observations?Rough rule of thumb: 10 observations for each explanatory variable,
