FINANCIAL MODELING FOR EQUITY RESEARCH A Step-by-Step Guide to Earnings Modeling Second Edition JOHN MOSCHELLA, CFA, CPA Copyright © 2017 Gutenberg Research LLC All rights reserved About the Author John has more than a decade of experience analyzing companies in various capacities After earning a BSBA in Finance, MS in Accounting and MBA from Northeastern University, he began his professional career at PricewaterhouseCoopers (PwC) in New York as an Assurance Associate in the Financial Services practice He also participated in a rotational assignment in the Financial Service Research Institute at PwC where he studied bank mergers during the financial crisis After PwC, John spent five years at UBS Investment Bank where he worked first as a Capital Specialist, and then as an Equity Research Associate In his research role, he maintained earnings models for companies in the Semiconductor and Semiconductor Equipment Industries, contributed to research reports, and participated in investor conferences John moved to General Electric Capital Corp in early 2014 as a Lead Risk Analyst where he built regression models to predict asset losses based on various macroeconomic scenarios After the sale of the majority of GE Capital’s assets, John started a consulting firm which provides capital planning support to investment banks, in addition to running Gutenberg Research About Gutenberg Research Gutenberg Research is a web-based, interactive earnings modeling community, which provides professional analysis based on modern portfolio theory and fundamental valuation techniques, with the mission of making earnings models available to all investors The Gutenberg name and philosophy are inspired by the fifteenth century visionary and inventor of the printing press, Johannes Gutenberg Gutenberg's press forever altered the state of communication and flow of information through the mass production of books, changing literacy from a luxury of an elite few, to a right of the masses Now, more than 500 years later there are still many limitations on information For example, investment banks restrict the distribution of their earnings models to top paying clients The Gutenberg Research community is building an inventory of models which all investors can access for free Model Templates Throughout this book images of models are used to demonstrate the topics covered These Microsoft Excel-based files are available for download when you register your book at www.GutenbergResearch.com/book The cell references in the text correspond to the cells within the spreadsheets This will enable you to drill into the cells and trace through the equations, or alter the files to suit your modeling needs File 1—Blank Model Template: Use this template to create your own earnings model None of the cells are locked so you can change the accounts and equations to fit any company you choose File 2—Apple Inc Back of the Envelope Model: The majority of the examples covered in this book are based on Apple Inc This beginner model is perfect for those who have not had prior modeling experience This model is covered in Chapter File 3—Apple Inc Tier Earnings Model: This version of the model is more sophisticated and includes a breakdown of the company’s products The Tier model is covered in Chapter File 4—Apple Inc Tier Earnings Model: The Tier model is geared toward advanced analysts and includes financial statement integration, as well as a discounted cash flow valuation This model is covered in Chapter File 5—Equity Risk Premium (ERP) Model: Using this simple model you can quickly estimate the market ERP based on volatility, changes in interest rates, and market return expectations You can then derive a discount rate using your ERP estimate, and the Capital Asset Pricing Model (CAPM) This model is covered in Chapter File 6—Apple Inc Beta Calculation: This file demonstrates the calculation of beta, using an Excel-based regression, which is covered in Chapter File 7—Boeing Co Regression Model: This file calculates a simple linear regression and related statistical tests using Boeing’s revenue and the U.S Census Bureau’s Industrial Production report The Boeing regression model is covered in the Appendix File 8—Starbucks Inc Regression Model: This file demonstrates a multiple regression using Starbucks’ revenue and the U.S Census Bureau’s Advanced Retail & Food Sales Report, Food Services & Drinking Places sales data The Starbucks regression model is covered in the Appendix Terms of Use This book, and all associated models, files, and content published in text or on GutenbergResearch.com is for demonstration purposes only, and is presented “as is” Neither Gutenberg Research, the author, nor any Gutenberg Research agents or associates are liable for any errors, delays, incompleteness of data presented, or for actions taken based on reliance on any information contained in this book, associated files and templates, or information presented on GutenbergResearch.com The information presented does not represent investment advice Investors should consult a professional investment adviser prior to making investment decisions Preface—Your Financial Modeling Toolbox Equity research is a difficult field, with a steep learning curve for new associates The hours are long, clients are demanding, and modeling errors are unacceptable When I first joined the semiconductor research team at a New York-based investment bank I struggled with the workload Between the 100+ hour workweeks during earnings season, getting up-to-speed with the fundamentals of a highly technical industry, and studying for the FINRA exams, there was no additional time to learn the basics of the day-to-day tasks, most importantly how to model a company’s earnings Most senior research analysts have little time for training, so new associates tend to teach themselves, as I did, by reverse-engineering their team’s models I wrote this book and designed the Excel templates as resources for new associates to assist in the financial modeling learning process This book is based on my experience as a Sell-Side Research Associate, with input from the quantitative methods I used as a Risk Analyst, and a unique consideration of financial reporting from my time as a Public Accountant The methods covered are primarily geared toward sell-side earnings modeling, however, many of the topics are also applicable to careers in investment banking, asset management, or any other field which requires knowledge of forecasting or valuation Remember modeling is part science and part art The chapters to follow describe the modeling approach I use in practice There are many different methods, variations, and techniques you can use to forecast earnings If you have prior modeling experience, feel free to incorporate your own spin on the steps as you work through each section If you are new to modeling, you may find some of the concepts difficult at first, but if you follow each step you will be able to build an earnings model for nearly any company Table of Contents Chapter 1: Introduction to Financial Modeling Section 1: What is an Earnings Model Section 2: The Fundamental Principles of Modeling Section 3: How to Use This Book Section 4: The Earnings Cycle Section 5: Timing of Publication Section 6: Basic Excel Functionality Section 7: Types of Models Section 8: Anatomy of a Model Chapter 2: Building a Basic Financial Statement Model Step 1: Getting Started Step 2: How to Setup Your Model Step 3: Filling in the Historic Data Step 4: Assumptions & Forecast Equations Chapter 3: Converting to a Ti e r E a r n i n g s M o d e l Step 5: Build the Earnings Engine Step 6: Link Product or Segment Details Step 7: Complete the Income Statement Step 8: Adjust for Non-GAAP Items Chapter 4: Converting to a Ti e r E a r n i n g s M o d e l Step 9: Complete the Historic Financials Step 10: Balance Sheet Modeling—Assets Step 11: Balance Sheet Modeling—Liabilities Step 12: Balance Sheet Modeling—Equity Step 13: Cash Flow Statement Modeling Step 14: Primary Financial Statement Links Chapter 5: Model Calibration & Forecasting Step 15: Incorporate Historic Trends Step 16: Adjust for Seasonality Step 17: Adjust for New Product Launches Step 18: Guidance and the Consensus Step 19: Incorporate Your Opinions Step 20: Consider Leading Indicators Chapter 6: The DCF Inputs ( B e t a , E R P, C A P M , & WA C C ) Step 21: Calculate the Equity Risk Premium Step 22: Derive Beta Using the Regression Function in Excel Step 23: Calculate the Required Return on Equity using Beta and the CAPM Step 24: Calculate the WACC Chapter 7: Discounted C a s h F l o w Va l u a t i o n Step 25: Calculate the Stage-One DCF Step 26: Calculate the Stage-Two DCF Step 27: Calculate the DCF Valuation Step 28: Understanding the DCF Valuation Chapter 8: Market Multiple B a s e d Va l u a t i o n Step 29: Choose a Multiple Step 30: Separate Net Cash Step 31: Calculate the Historic Multiple Step 32: Create a Price Band Chapter 9: How to Use Yo u r E a r n i n g s M o d e l Step 33: Run Scenario Analysis Step 34: Perform Sensitivity Analysis Step 35: Analyze Guidance & Consensus Step 36: Prepare for the Release Step 37: Updating The Model Step 38: Regular Model Maintenance Appendix 1: Using Regression Analysis to Predict Earnings Step R1: Select Your Data Set Step R2: Run the Regression Step R3: Test the Model Step R4: Analyze the Output Step R5: Perform Back Testing Step R6: Model Limitations CHAPTER 1: INTRODUCTION TO FINANCIAL MODELING Section 1: What is an Earnings Model Section 2: Fundamental Principles Section 3: How to Use This Book Section 4: The Earnings Cycle Section 5: Timing of Publication Section 6: Basic Excel Functionality Section 7: Types of Models Section 8: Anatomy of a Model Chapter Overview: An earnings model is a representation of a company’s financial position, disaggregated to a level which can be analyzed based on historic results and future expectations The level of disaggregation depends on the subject company For example, if someone asks you whether or not Apple Inc will beat the consensus Earnings Per Share (EPS) estimate next quarter, it would be very difficult to answer without thinking about how many iPhones, Macs, and iPads the company will sell, the average price for each, and the profit margin An earnings model shows all of these components and how they work together to get to a final EPS estimate Section 1: What is an Earnings Model The term model is broadly used to explain many different types of financial representations, which can vary depending on the type of task the model is performing For example, an economist may use a regression model to predict the likelihood of an economic downturn based on several macro- “R1” (Regression-1) This is because the financial model has been completed in the previous chapter The steps in this appendix are only relevant for those who wish to use regression analysis to predict inputs into their earnings model Step R1: Select Your Data Set The first step is selecting the appropriate variables to use in the model It can be difficult to find variables with strong explanatory power Typically, we try to find data which is effective at predicting top-line revenue, since anything below the revenue line can be distorted over time due to fluctuations in profit margins, and other factors Ideally, we should find variables with limited impact from outside forces Given the nature of financial data, this is much easier said then done For example, you will see in the Boeing regression I have used revenue as the dependent variable, but revenue is subject to changes over time due to fluctuations in inflation, foreign exchange rates, and revenue recognition principals A better variable would be inputs into revenue, such as the number of units shipped which is not impacted by these factors Unfortunately, it can be very difficult to find a statistically significant relationship among variables, so sometimes we must make with what we can find Start by generating a list of potential leading indicators for a company’s revenue Then, test each variable using the steps in this chapter to find the best predictors For Boeing and Starbucks, I found the U.S Census Bureau’s reports to be helpful in providing leading indicators For Boeing, I selected the Aircraft & Parts Shipments section, including defense and non-defense shipments, in the Industrial Production report, which is published by the U.S Census Bureau on a monthly basis The final monthly report for each calendar quarter is typically released about a week before Boeing reports quarterly results which makes it a good data point for prediction The two variables (Boeing’s revenue and the shipment data) appear to have a strong correlation based on the time series plot below I will test this correlation in Step R2 In this example, Boeing’s revenue is what we would like to predict, so it is the dependent variable, also know as the “Y” variable, or the variable which “depends” on the results of another variable Aircraft & Parts Shipments from the Industrial Production report is the independent variable, also known as the “X” variable Exhibit 64—Boeing Revenue vs Aircraft & Parts Shipments Source: Company reports, U.S Census Bureau Manufacturers’ Shipments, defense & non-defense For Starbucks, I chose the Food Services & Drinking Places sales data, contained in the Advanced Retail & Food Sales report published by the U.S Census Bureau Similar to the Boeing example, this report is also available prior to the company’s quarterly earnings release and appears to show a strong correlation to Starbucks’ revenue In the chart, below there is a clear de-coupling between the two variables in the fourth calendar quarter each year This is because Starbucks’ fourth calendar quarter sales typically outperform the Food Service & Drinking Places total results If an adjustment to account for this seasonality is not included, then the model will likely have a larger standard error, which is the standard deviation of the residuals (residuals are the differences between each observation and the model predicted value for that observation) The smaller the standard error, the better the model will be at predicting results I will be using a second independent variable to account for the seasonality, which is a “dummy” variable to represent the fourth calendar quarter A dummy variable has a value of or In the Starbucks example, a value of represents the fourth calendar quarter Exhibit 65—Starbucks Revenue vs U.S Food Services Retail Sales Source: Company reports, U.S Census Bureau Advanced Retail and Food Sales Step R2: Run the Regression The next step is to run the regression models for Boeing and Starbucks I will be using Excel for this analysis but feel free to use your favorite statistical software If you have not installed the Excel Data Analysis add-in, refer to the installation instructions in Step 22b in Chapter Starting with the Boeing model, follow these steps: Step R2a: In Excel, on the “Data” tab, click on “Data Analysis”, then “Regression” Select the quarterly percentage change in Boeing’s revenue column for the Y input range (dependent variable), and the percentage change in shipments column for the X input range (independent variable) Step R2b: Check the “Labels” box, which indicates that the Y and X ranges include the data headings Step R2c: For this regression, we want to set the constant in the regression equation equal to zero, so check the “Constant is Zero” box You can try running the regression with this box unchecked to see which result is better The constant is the term before the beta coefficient in the regression equation, which will be explained in greater detail in Step R4 Step R2d: Select the “Output Range” where you would like the regression results to be populated Step R2e: Click the residual details you would like to analyze We will discuss how to analyze residuals in Step R3 Step R2f: Click “OK” to run the regression We follow a similar procedure to run the Starbucks multiple regression, except the X range includes two columns of data, since it is a multiple regression, and we not check the “Constant is Zero” box Step R3: Test the Model Before we can draw any conclusions about the correlation between the variables, we must test the validity of the models To this, we perform a series of statistical reviews, including the following (starting with the Boeing Model): Step R3a—P-Value and t-Stat Review: These two statistics are used to test whether or not the relationship between the variables in the regression is different from zero If the t-stat is greater than +2.79 or less than -2.79, then a statistically significant relationship exists at the 99% confidence interval (also known as the 1% level of significance) The 2.79 figure comes from the Critical t-value of a distribution chart representing 0.005 probability in each tail with 25 degrees of freedom In this case, the model’s t-stat of 13.71 (refer to cell S19 in Exhibit 66) is well above the critical t-value of 2.79; therefore, a statistically significant relationship exists between the variables It is important to note that pure statisticians would not say that a relationship between the variables exists, but that we can reject the notion that no relationship exists, or we reject the null hypothesis The P-value represents the lowest level of significance the model can reach while still showing a statistically significant relationship In this case, the Pvalue is close to zero (cell T19 in Exhibit 66), which means we would reach the same conclusion even at a confidence interval greater than 99% Exhibit 66—Boeing Regression Model Output: PValue & t-Stat Step R3b—Breusche-Pagan Test: To make sure the independent variables not explain variation in the residuals (a situation known as heteroskedasticty), run the Breusche-Pagan test by regressing the independent variable against the squared residuals The resulting R-square is relatively low with a high p-value, and a t-stat between the critical-t points (refer to cells AE19 and AF19 in Exhibit 67 below); therefore, heteroskedasticity is not an issue with this model To put it another way, the correlation between shipments and the squared residuals is not statistically different from zero Exhibit 67—Boeing Model Breusche-Pagan Results In addition to the Breusche-Pagan test, we can also plot the residual with a linear trend line showing the R Square to determine if the independent variables explain variation in the residuals (this was the box we checked in Step R2e) In the Boeing example, the trend line and R Square are close to zero so the residuals and independent variable are not correlated Exhibit 68—Boeing Model Residual Plot Step R3c—Serial Correlation Test: Serial correlation is the relationship of a variable with itself over time If serial correlation is present, past observations will influence future results To ensure the model does not exhibit signs of serial correlation, run the Durbin-Watson (DW) test, by calculating the DW statistic To calculate the Durbin-Watson statistic, first square the residuals from the regression output Then, sum the difference of the actual residuals and the residuals on a one-period lag squared Next, divide the result by the squared residuals Durbin-Watson statistics range in value between and A value near indicates positive serial correlation, a value near indicates negative serial correlation, and a value near indicates no significant serial correlation Note that these ranges represent approximations For exact measures refer to a DW Significance Table The Boeing model DW statistic is near 2, which indicates that serial correlation is not likely an issue for this model Step R3d—The Normal Distribution Assumption: To test that the data does not violate the assumption of a normal distribution, calculate the skew and kurtosis A distribution is considered to be skewed if more observations are found on one side of the mean A perfect normal distribution would have a skew of A skew of +/-1 can be considered approximately normally distributed Kurtosis is the measure of the distribution’s peak A kurtosis of less than can be considered approximately normally distributed Excel has equations for these formulas: =SKEW(residuals) and =KURT(residuals) If you feel you need additional evidence, you can also run a Chi-sq test, and plot the residuals in a histogram to see if they appear to be normally distributed In the Boeing model the skew is 0.02 (which is below +/-1), and the kurtosis is -0.53 (which is below 3) These statistics show that the sample is approximately normally distributed Step R3e—Model Stability: To test the stability of the model, remove two random observations from the sample and re-run the regression The new regression should pass all the prior tests with a similar standard error compared to the first model This will prove that the regression model is fairly stable Step R3f—Other Considerations: Textbook examples for regression analysis tend to be relatively straightforward In the real world, data and external conditions will change over time which makes a regression model less powerful, and in most cases ineffective Even if your model passes all the statistical tests, it is important to think qualitatively about whether or not there is a true economic reason for the variables’ correlation and if the direction of the coefficients makes sense Using the Boeing example, it stands to reason that U.S Aircraft & Parts Shipments could have predictive power for Boeing’s revenue, and the positive regression coefficient makes sense as it implies a direct relationship (when shipments increase Boeing’s revenue increases) In addition, it is important to assess whether or not the population has changed over time If there have been significant changes, for example, if Boeing began selling medical devices or hamburgers, then the model could not be used (violates the homogenous population condition) Another potential issue is that the Boeing regression model uses U.S Shipment data as a proxy for global revenue Using a proxy to represent a larger population of data has problems (refer to limitations in Step R6) In addition, if Boeing’s global share of revenue shifts significantly, foreign exchange rates fluctuate, or revenue recognition principals change, this model may fail to provide a reliable estimate The first five tests are relevant for both of the Boeing and Starbucks regression models Since the Starbucks example is a multiple regression, we have a few additional tests to run, including the following: Step R3g—F-Stat Review: This statistic is similar to the t-stat, except it is used for multiple regressions If the F-stat is greater than or less than the critical F-values, then a statistically significant relationship exists Critical Fvalues are published by degrees of freedom on a critical F chart which is available in most statistical textbooks In the Starbucks model, the F-stat of 29.88 (refer to cell U17 in Exhibit 69 below) is well above the critical Fvalue of 2.86; therefore, a statistically significant relationship exists between the variables (or we reject the null hypothesis that no relationship exists) Exhibit 69—Starbucks Regression Model Output Step R3h—Significance F: The Significance F for a multiple regression has the same interpretation as the P-value for a simple linear regression: It represents the lowest level of significance the model can reach while still showing a statistically significant relationship In this case, the Significance F value is close to zero (refer to cell V17 in Exhibit 69 above) which means we would reach the same conclusion even at a confidence interval greater than 99% Step R3i—Multicollinearity: To test for multicollinearity, regress the two independent variables against each other to ensure they are not correlated In the Starbucks regression model, the resulting regression between the calendar fourth quarter dummy variable and the percentage change in U.S Food Service and Drinking Places Sales is not statistically significant at the 95% confidence interval; therefore, multicollinearity is not a significant issue Step R4: Analyze the Output Now that we have determined that our regression models are valid, we can analyze the results and use them to predict future period revenue, which can be included in the earnings model built in the previous chapters The R Square, also known as the coefficient of determination, represents the amount of variation in the dependent variable, explained by the independent variable In the Boeing example, we can conclude that 84% of the variation in the percentage change in Boeing’s revenue is explained by the percentage change in U.S Defense and Non-Defense Aircraft & Parts Shipments, based on the R Square in the regression output (refer to Exhibit 66 cell Q8) Since we ran the Boeing regression without a constant intercept, the regression equation is: Percentage change in Boeing Revenue = 0.88 × the percentage change in U.S Aircraft & Parts Shipments The 0.88 in the equation represents the coefficient in the regression output (refer to Exhibit 66 cell Q19) We can now use this equation to predict Boeing’s future quarterly revenue by inputting the quarterly percentage change in Aircraft & Parts Shipments, which is published before Boeing releases quarterly results The R Square in our Starbucks Model is 0.76, which means that 76% of the variation in Starbucks revenue is explained by changes in U.S Food Services & Drinking Places Sales (refer to Exhibit 69 cell R11) The Standard error gives the predicted range of +/- 4% (Exhibit 69 cell R12) The regression equation is: Percentage change in Starbucks’ revenue = -0.02 + 0.89 × percentage change in Food Services Sales + 0.15 × 4Q Dummy Variable The coefficients from the equation come from cells R22, R23, and R24 Plugging in the new U.S Census data and the dummy variable for each quarter will result in next quarter’s predicted revenue value, which we could incorporate into our earnings model ahead of the company’s release Step R5: Perform Back Testing Using the regression equation for Boeing, we can back test what the model would have predicted based on the shipment data for the last three quarters The data in Table 13 below shows that the model predicted revenue outperformed the consensus revenue estimate in two of the last three quarters used in the regression model Table 13—Boeing Regression Model Back Testing Quarter March 2015 June 2015 September 2015 Model Predicted Revenue ($M) $22,471 Consensus Estimate ($M) $22,488 Actual Reported Revenue ($M) $22,149 $24,600 $24,216 $24,543 $24,725 $24,738 $25,849 Analysis More accurate then consensus More accurate then consensus Less accurate then consensus The results for the Starbucks model, shown in Table 14 below, were not as good as the Boeing model The regression beat the consensus estimate in only one of the last four quarters used in the model This is due in part to the fact that consensus estimates for Starbucks tend to match management's guidance, and since guidance is generally reliable for Starbucks, earnings surprise is usually relatively low The regression model would be most useful when sales change sharply after management has already issued guidance Table 14—Starbucks Regression Model Back Testing Quarter December 2014 March 2015 June 2015 September 2015 Model Predicted Revenue ($M) Consensus Estimate ($M) Actual Reported Revenue ($M) $4,801 $4,797 $4,803 $4,700 $4,530 $4,564 $4,821 $4,863 $4,881 $4,722 $4,897 $4,915 Analysis More accurate then consensus Less accurate then consensus Less accurate then consensus Less accurate then consensus Step R6: Model Limitations As with all regression analysis, there are limitations to our models which should be considered when deciding whether or not to use regression analysis Some examples of limitations include the following: Sample: Regression analysis draws conclusions based on a limited sample of observations, typically at least 30 observations if possible As a result, there is always a degree of error expected Ahead of an earnings release, we quantify the expected error with a range of values, but actual results can exceed the model’s standard error, particularly if the assumptions in the model are not held constant in the actual results Historic Trends: Regression analysis assumes that historic relationships will hold constant in the future and does not consider changes such as market share or currency rate fluctuations The residuals from the predicted values show instances in the past where the model predicted revenue differs significantly from the actual result Similar instances will likely occur in the future Revisions: Economic data is often revised after it is released, which can result in impaired predictability Company Policies: Revenue recognition policies can change over time and may not align with shipment data (Boeing example) Proxies: The Boeing model uses U.S Aircraft & Parts Shipments as a proxy for global sales of products and services The percentage of Boeing’s revenue generated outside of the U.S fluctuates over time, as does the ratio of service-to-product revenue Both of which will impact the model’s ability to accurately predict revenue The Starbucks model uses Food Service & Drinking Places Sales as a proxy for global sales The percentage of Starbucks’ revenue generated outside of the U.S fluctuates over time, which will impact the model’s ability to accurately predict revenue There are a few ways we could improve the predictive power of these models For example, we could segment the revenue between geographic regions, and products versus services, and re-run the regression To accomplish this, we would need to find similar reliable economic statistics for each region to arrive at the total revenue number We could also include additional observations, although the further back you go the more likely it is that your population attributes have changed Due to the inherent limitations, most equity research analysts not rely on regression analysis in their forecasts Despite the limitations, regression models can provide important insight to help forecast results, and if nothing else provide a few extra data points to consider in our earnings sensitivity analysis ahead of quarterly releases ... Reporting Language (XBRL) tags for financial statement disclosures These tags are used by third party data aggregators to make inputting financial data into an Excel file as easy as refreshing... model 2) Balance Sheet and Cash Flow Statement Analysis: Having all three financial statements available in the forecast allows for much greater detail in future period analysis including cash conversion... the advanced material, you can add the Balance Sheet and Cash Flow Statement and transform your Tier model to a Tier with full integration of the financial statements Table 1—Types of Earnings