316  Estimating the Cost of Capital We find that individual company betas can be heavily influenced by nonre- peatable events, so we recommend using an industry peer median rather than the historically measured beta for the company in question. Betas can also be affected by unusual events in the stock market, such as the dot-com bubble of the early 2000s or the financial crisis of 2007–2009. By examining how industry betas have changed over time, you can apply judgment about whether betas will revert to their long-term level if they are currently not there. The remainder of this section describes how to estimate a company’s beta step-by-step. First, use regression to estimate the beta for each company in the peer group. Then convert each company’s observed beta into an unlevered beta—that is, what the beta would be if the company had no debt. Once you have a collection of betas, examine the sample for a representative beta, such as the median beta. To ensure that the current beta is representative of risk and not an artifact of unusual data, do not rely on a point estimate. Instead, examine the trend over time. We discuss each step next. Estimating Beta for Each Company in the Industry Sample Set  To develop an industry beta, you first need the betas of the company’s peer set. Since beta cannot be observed directly, you must estimate its value. The most common regression used to estimate a company’s raw beta is the market model: R R i m = + + α β ε In the market model, the stock’s return (Ri), not price, is regressed against the market’s return. Exhibit 15.5 plots 60 months of Costco stock returns versus Morgan Stan- ley Capital International (MSCI) World Index returns between September 2015 EXHIBIT 15.4  Cost of Equity Using the Capital Asset Pricing Model (CAPM) 0 4 2 6 10 8 12 14 0.0 0.5 1.0 Beta (systematic risk) Expected return, % 1.5 2.0 General Mills Market portfolio Micron Technologies Source: Refinitiv Thomson One. Estimating the Cost of Equity  317 and August 2019. The solid line represents the “best fit” relationship between Costco’s stock returns and the stock market. The slope of this line is commonly denoted as beta. For Costco, the company’s raw regression beta (slope) is 0.85. But why did we choose to measure Costco returns in months? Why did we use five years of data? And how precise is this measurement? The CAPM is a one-period model and provides little guidance on how to use it for valu- ation. Yet following certain market characteristics and the results of a variety of empirical tests leads to several guiding conclusions: • The measurement period for raw regressions should include at least 60 data points (e.g., five years of monthly returns). Rolling betas should be graphed to search for any patterns or systematic changes in a stock’s risk. • Raw regressions should be based on monthly returns. Using more frequent return periods, such as daily and weekly returns, leads to systematic biases.17 • Company stock returns should be regressed against a value-weighted, well-di- versified market portfolio, such as the MSCI World Index, bearing in mind that this portfolio’s value may be distorted if measured during a market bubble. In the CAPM, the market portfolio equals the portfolio of all assets, both traded (such as stocks and bonds) and untraded (such as private companies and human capital). Since the true market portfolio is unobservable, a proxy is neces- sary. For U.S. stocks, the most common proxy is the S&P 500, a value-weighted EXHIBIT 15.5  Costco: Stock Returns, 2015–2019 15 10 5 0 –5 –10 –15 –15 –10 –5 5 10 15 Regression beta = 0.85 Morgan Stanley Capital International (MSCI) World Index monthly index returns Costco monthly stock returns 0 Source: Refinitiv Thomson One. 17 Using daily or even weekly returns is especially problematic when the stock is rarely traded. An il- liquid stock will have many reported returns equal to zero, not because the stock’s value is constant but because it hasn’t traded (only the last trade is recorded). Consequently, estimates of beta on illiquid stocks are biased downward. Using longer-dated returns, such as monthly returns, lessens this effect.