Module 10: Empirical Tests IF440 - Capital Markets and Uncertainty Ranik Raaen Wahlstrøm Overview • CAPM assumes only one thing matters when it comes to determining why different securities, offer different expected rates of return – beta! – Nearly impossible to conduct definitive test of CAPM given reliance on unobservable market portfolio • Recent research suggests the following: – Risk-return relationship might entail multiple sources of systematic risk – Non-risk-related considerations could lead to substantial variation in expected returns – Behavior-finance issues can lead to security mispricing 2 Testing CAPM • Early simple test followed three steps: 1. Setting up sample data 2. Estimating the security characteristic line (SCL) 3. Estimating the security market line (SML) 3 Testing CAPM: (1) Setting up sample data • Decide on: – – – – – a sample period (e.g., 10 years, 2011 - 2021) a holding period (e.g., monthly) a sampling size (e.g., 1000 stocks) a proxy for ๐ (e.g., DJIA) a proxy for ๐๐ (e.g., T-bills) • The example choices above gives us: – – – – – ๐ก = {1,2, … , 120} ๐ = {1,2, … , 1000} ๐๐๐ก = {1,2, … , 120,000} ๐๐๐ก = {1,2, … , 120} ๐๐๐ก = {1,2, … , 120} 4 Testing CAPM: (2) Estimating the security characteristic line (SCL) First-pass regression: • Use regression to solve for each stock ๐: ๐๐,๐ก − ๐๐,๐ก = ๐๐ + ๐๐ ๐๐๐ก − ๐๐๐ก + ๐๐,๐ก Will be fitted for 1000 stocks, each over 120 months, to obtain: • ๐๐ – estimate for the ๐ฝ๐ • ๐ 2 ๐๐ – estimate for nonsystematic risk for each stock • ๐๐ − ๐๐ – sample averages for each stock, over 120 months • ๐๐ − ๐๐ – sample average over 120 months 5 Testing CAPM: (3) Estimating the security market line (SML) Second-pass regression: • Use regression to solve: ๐๐ − ๐๐ = ๐พ0 + ๐พ1 ๐๐ + ๐พ2 ๐ 2 ๐๐ Will be fitted over 1000 observations, one for each stock. If CAPM is true ๐ธ ๐๐ − ๐๐ = ๐ผ๐ + ๐ฝ๐ ๐ธ ๐๐ − ๐๐ , we would expect: • ๐พ0 = 0 • ๐พ1 = ๐๐ − ๐๐ • ๐พ2 = 0 – Non-systematic risk should not be priced 6 Testing CAPM • Early tests (John Lintner, 1965, Merton Miller and Myron Scholes, 1972) was not promising. ๐พ0 = 12.7% ๐พ1 = 4.2% ๐พ2 = 31.0% ๐๐พ0 = 0.6% ๐๐พ1 = 0.6% ๐๐พ2 = 2.6% ๐๐ − ๐๐ = 16.5% • Highlights: – SML is “too flat” and intercept is “too large” – ๐พ0 − 0 = ๐๐๐๐พ0 – ๐พ1 − ๐๐ − ๐๐ = ๐๐๐๐พ1 – ๐พ2 − 0 = ๐๐๐๐พ2 7 Exercises solved with code 1. Perform the first-pass regressions for a single-index model and tabulate the summary statistics. 2. Specify the hypotheses for the second-pass regression used to test the SML and the CAPM. 3. Perform the second-pass SML regression by regressing the average excess return of each portfolio on its beta. 4. Summarize your test results and compare them to the results reported in the text. Code and data available at: https://github.com/ranikrw/IF440-Empirical-tests Note: coding is not part of the syllabus in this course. You will not be asked to either write or interpret code on the exam. 8 Difficulties with approach employed • Stock returns are extremely volatile, lessening the precision of any tests of average return • Fundamental concerns about the validity of the tests 1. The proxy for ๐ (for example DJIA) is not the real ๐ 2. Betas from first-stage are estimated with sampling error 3. Investors cannot borrow at ๐๐ 9 Concerns when testing CAPM: (1) The proxy for ๐ด is not the real ๐ด • Roll's Critique (Richard Roll, 1977): – The only testable hypothesis for testing CAPM: The market portfolio is mean-variance efficient – CAPM cannot be tested, because one cannot observe ๐ • Benchmark error: incorrect proxy for ๐ → incorrect testing of CAPM • ๐ might be mean-variance efficient, even if a proxy is not – Leads us to believe the CAPM is not true, even if it is • A proxy might be mean-variance efficient, even if ๐ is not – Leads us to believe the CAPM is true, even if it is not 10 Concerns when testing CAPM: (2) Large sampling errors in the first regression • Recall the first regression: ๐๐,๐ก − ๐๐,๐ก = ๐๐ + ๐๐ ๐๐๐ก − ๐๐๐ก + ๐๐,๐ก • Such regression (linear with error term on the right) gives a downward bias slope ๐๐ and upward bias intercept ๐๐ – Showed (early 1970’s): estimated the SCL using random numbers with averages exactly agreeing with CAPM – helps us explain the poor empirical results for CAPM, for early tests 11 Concerns when testing CAPM: (2) Large sampling errors in the first regression • Solution - Using portfolios instead of single stocks → diversifying away most firm-specific risk ๐๐๐ก – New Problem 1: less observations for the second regression – New Problem 2: likely very small betas – Still not much results in favor for the CAPM 12 Using portfolios instead of single stocks: Fama and MacBeth (1973) Procedure • Step 1: Portfolio formation – for the period 1926-1929 – In order to obtain portfolios with different betas we first compute betas of individual stocks ๐ Cov(๐๐ , ๐๐ ) ๐ฝ๐ = ๐ 2 ๐๐ – ๐๐ = observed returns for each security ๐ – ๐๐ = the market portfolio – Next, form 20 portfolios based on the sorted ๐ฝ๐ 13 Using portfolios instead of single stocks: Fama and MacBeth (1973) Procedure • Step 2: Initial Estimation Period (1930-1934) – For a later time period, recompute the ๐ฝ๐ for each security – These are averaged across securities within portfolios to obtain 20 initial portfolio ๐ฝ๐ – For the sample, estimate the variance of residuals ๐ 2 (๐๐ ) for each security ๐: ๐๐๐ก = ๐๐ + ๐๐ ๐๐๐ก + ๐๐๐ก – Get the portfolio's variance ๐ 2 (๐๐ ) by averaging over the corresponding securities’ variance ๐ 2 (๐๐ ) 14 Using portfolios instead of single stocks: Fama and MacBeth (1973) Procedure • Step 3: Testing period (1935-1968) – For each month ๐ก per year, perform cross-section regression: ๐๐๐ก = ๐พ0 + ๐พ1 ๐ฝ๐ + ๐พ2 ๐ฝ๐2 + ๐พ3 ๐ 2 ๐๐ + ๐๐๐ก – ๐๐๐ก is the dependent variable of the 20 portfolio returns – ๐ฝ๐ represents the 20 portfolios betas as explanatory variable – For each month in 1935 we use the portfolio betas from Step 2 • They are however updated yearly (starting at Jan 1936) – Calculate the time series averages of ๐พ0 and ๐พ1 – The CAPM predicts: ๐พ0 = ๐เดฅ๐ ๐พ1 = ๐๐ − ๐เดฅ๐ ๐พ2 = ๐พ3 = 0 15 Using portfolios instead of single stocks: Fama and MacBeth (1973) Procedure (all rates in basis points per month) Period Average γ0 t-statistic (testing γ0 = 0) Average rM − rf Average γ1 t-statistic (testing γ1 = rM − rf ) Average γ2 t-statistic (testing γ2 = 0) Average γ3 t-statistic (testing γ3 = 0) Average R-square 1935/6 to 1968 8 0.20 130 114 1.85 −26 −0.86 516 1.11 0.31 1935 to 1945 10 0.11 195 118 0.94 −9 −0.14 817 0.94 0.31 1946 to 1955 8 0.20 103 209 2.39 −76 −2.16 −378 −0.67 0.32 1955/6 to 1968 5 0.10 95 34 0.34 0 0 960 1.11 16 0.29 Exercises solved with code 5. Group the nine stocks into three portfolios, maximizing the dispersion of the betas of the three resultant portfolios. Repeat the tests of Exercises 1-4 and explain any changes Code and data available at: https://github.com/ranikrw/IF440-Empirical-tests in the results. Note: coding is not part of the syllabus in this course. You will not be asked to either write or interpret code on the exam. 17 Tests of the Multifactor Models • Three types of factors likely augment market risk factor in a multifactor SML: 1. Factors that hedge consumption against uncertainty in prices • For example, housing or energy 2. Factors that hedge future investment opportunities • For example, interest rates or market volatility 3. Factors that hedge assets missing from market index • For example, labor income or private business 18 Labour income: Jagannathan and Wang (1996) Second-pass regression: ๐ธ(๐ ๐๐ก ) = ๐0 + ๐size log(ME๐ ) + ๐vw βvw + ๐credit βcredit + ๐labor βlabor ๐ ๐ ๐ • • • • ME vw labor ๐๐๐๐๐๐ก – market value of equity – value-weighted stock index portfolio – rate of change in aggregate labor income – business cycle: spread of yields between high- and low-grade corporate bonds 19 Labour income: Jagannathan and Wang (1996) credit + ๐ labor ๐ธ(๐ ๐๐ก ) = ๐0 + ๐size log(ME๐ ) + ๐vw βvw + ๐ β β ๐ credit ๐ labor ๐ Coefficient Estimate t-statistic Estimate t-statistic Estimate t-statistic Estimate t-statistic ๐๐ 1.24 5.16 2.08 5.77 1.24 4.10 1.70 4.14 ๐vw −0.10 −0.28 −0.32 −0.94 −0.40 −0.88 −0.40 −1.06 ๐credit 0.34 1.73 0.20 2.72 ๐labor 0.22 2.31 0.10 2.09 ๐size R2 1.35 −0.11 −2.30 57.56 55.21 −0.07 −1.30 64.73 Findings: • Rejection of the CAPM: negative ๐vw (expected average returns fall with beta) • Human Capital is important in any version of the CAPM 20 Private (non-traded) business: Heaton and Lucas (2000) • Extend the equation of Jagannathan and Wang (1996) to also include the rate of change in proprietary-business wealth • Find that households with higher investments in private business reduce the fraction of total wealth invested in equity 21 Empirical Tests of the APT: Chen, Roll and Ross (1986) • Identify possible variables that might proxy for systematic factors: IP Growth rate in industrial production EI Changes in expected inflation UI Unexpected inflation CG Unexpected changes in risk premiums on bonds GB Unexpected changes in term premium on bonds ๐ = ๐ + ๐ฝ๐ ๐๐ + ๐ฝ๐ผ๐ ๐ผ๐ + ๐ฝ๐ธ๐ผ ๐ธ๐ผ + ๐ฝ๐๐ผ ๐๐ผ + ๐ฝ๐ถ๐บ ๐ถ๐บ + ๐ฝ๐บ๐ต ๐บ๐ต + ๐ 22 Empirical Tests of the APT: Chen, Roll and Ross (1986) Used the Fama and MacBeth (1973) Procedure ๐ = ๐พ0 + ๐พ๐ ๐ฝ๐ + ๐พ๐ผ๐ ๐ฝ๐ผ๐ + ๐พ๐ธ๐ผ ๐ฝ๐ธ๐ผ + ๐พ๐๐ผ ๐ฝ๐๐ผ + ๐พ๐ถ๐บ ๐ฝ๐ถ๐บ + ๐พ๐บ๐ต ๐ฝ๐บ๐ต + ๐ – ๐พj is the price of risk associated with factor ๐ – The APT is valid if ๐พj is statistically significant A B EWNY IP EI UI CG GB Constant 5.021 14.009 −0.128 −0.848 0.130 −5.017 6.409 (1.218) (3.774) (−1.666) (−2.541) (2.855) (−1.576) (1.848) VWNY IP EI UI CG GB Constant −2.403 11.756 −0.123 −0.795 8.274 −5.905 10.713 (−0.633) (3.054) (−1.600) (−2.376) (2.972) (−1.879) (2.755) VWNY = Return on the value-weighted NYSE index; EWNY = Return on the equally weighted NYSE index 23 Exercises solved with code Suppose that a second factor is considered. The values of this factor for years 1 to 12 were as follows 8. Perform the first-pass regressions as did Chen, Roll and Ross and tabulate the relevant summary statistics. 9. Specify the hypothesis for a test of a second-pass regression for the two-factor SML 10. Do the data suggest a two-factor economy? Code and data available at: https://github.com/ranikrw/IF440-Empirical-tests Note: coding is not part of the syllabus in this course. You will not be asked to either write or interpret code on the exam. Year % change in factor value 1 -9.84 2 6.46 3 16.12 4 -16.51 5 17.82 6 -13.31 7 -3.52 8 8.43 9 8.23 10 7.06 11 -15.74 12 2.03 24 Empirical Tests of the APT: Fama and French (1993) Break down stocks based on book-to-market ratio and size (total assets) Size (total assets) Small (50%) Book to market ratio Big (50%) High (30%) S/H B/H Medium (40%) S/M B/M Low (30%) S/L B/L 25 Empirical Tests of the APT: Fama and French (1993) • High book-to-market firms experience higher returns • Smaller firms experience higher returns ๐ธ ๐๐ − ๐๐ = ๐๐ + ๐๐ ๐ธ ๐๐ − ๐๐ + ๐ ๐ ๐ธ ๐ ๐๐๐ต + โ๐ ๐ธ ๐ ๐ป๐๐ฟ • ๐ธ ๐๐ − ๐๐ – Excess return • SMB – Small Minus Big: Difference in returns on portfolios of small and big stocks – High Minus Low: Difference in returns on portfolios of high (value) and low (growth) book-to-market stocks • HML 26 CAPM versus FF Model Goyal (2012): CAPM versus the Fama and French model. The figure plots the average actual returns versus returns predicted by CAPM and the FF model for 25 size and book-to-market double-sorted portfolios. US stocks, 1946-2010. 27 • Firm size (SMB) and book-to-market (HML) are firm characteristics. Why do these predict stock returns? • Two suggested possibilities: 1. Capture some aspects of the business cycle 2. Behavioral explanations 28 Risk-Based Interpretations • Liew and Vassalou (2000): Returns on style portfolios (HML and SMB) seems to predict GDP growth, and thus may capture some aspects of business cycle risk Difference in return to factor portfolios in year prior to above-average versus below-average GDP growth. Both SMB and HML portfolio returns tend to be higher in years preceding better GDP growth. 29 Behavioral Explanations • Value premiums is a manifestation of market irrationality • “Glamour firms” – Recent good performance – High prices – Lower book-to-market ratios • Overreaction – High past growth is extrapolated and then impounded in price • Extrapolation error – Market ignores evidence that past growth cannot be extrapolated far into the future 30 Momentum: A Fourth Factor • Carhart (1997) adds a momentum factor to the FF 3-factor model: ๐ธ ๐๐ − ๐๐ = ๐๐ + ๐๐ ๐ธ ๐๐ − ๐๐ + ๐ ๐ ๐ธ ๐ ๐๐๐ต + โ๐ ๐ธ ๐ ๐ป๐๐ฟ + ๐ค๐ ๐ธ ๐ ๐๐๐ฟ • WML: Winners Minus Loosers • Motivation: – Empirical evidence shows that stocks that have been doing well in the recent past (about 12 months) continue to do so – Stocks that have been doing poorly in the recent past also continue 31 to do so Characteristics versus Factor Sensitivities These are factors that have joined as additions to the CAPM: – Size – Asset growth – Book-to-market – Dividend yield – Momentum – Volatility – Stock issues – Turnover – Accruals – Sales – Profitability Liquidity and Asset Pricing • Liquidity involves the following: – – – – – Trading costs Ease of sale Necessary price concessions to effect a quick transaction Market depth Price predictability 33 Liquidity and Asset Pricing • Pástor and Stambaugh (2003) look for evidence of price reversals, especially following large trades – If price reversed on the following day, then part of the original price change the previous day was due to the trades – That is, concession on the part of trade initiators who needed to offer higher purchase prices (or accept lower selling prices) to complete trade in a timely manner 34 Liquidity and Asset Pricing • Pástor and Stambaugh (2003) used models that ignore liquidity (CAPM and FF) – Find the same if also including momentum 35 The Equity Premium Puzzle • Recall from the derivation of the CAPM that the equilibrium risk premium is given by: 2 าง ๐ ๐ธ ๐ ๐ = ๐ธ ๐๐ − ๐๐ = ๐ด๐ – ๐ดาง is the average investor risk aversion Historically, market risk premiums on risky assets in the U.S. are too large to be consistent with economic theory • unless we assume implausibly high levels of risk aversion 36 The Equity Premium Puzzle Possibly reason 1: Actual return > expected returns • The Equity Premium Puzzle is not because economic theory and suggested levels of risk aversion is wrong • Rather, it comes from extraordinary good times after 1949 →Actual returns turned out to be higher than what the market expected 37 The Equity Premium Puzzle Possibly reason 2: Survivorship Bias • The Equity Premium Puzzle comes from studying the U.S. economy, the most successful economy after WW2 – Ignoring the evidence from stock markets that did not survive for the full sample period • This could not have been expected. So, when considering expected returns beforehand, one must have considered the whole world economy → then the Equity Premium Puzzle would probably not be present. 38 The Equity Premium Puzzle Possibly reason 3: Illiquidity The calculations behind the Equity Premium Puzzle do not account for illiquidity → Part of the Equity Premium Puzzle may come from a compensation for illiquidity 39 The Equity Premium Puzzle Possibly reason 4: Irrational investor behavior • Narrow framing: Investors consider every risk in isolation, rather than their total risk exposure → Investors focus on total volatility rather than the low correlation of a stock portfolio with other components of wealth → Require higher risk premiums than rational models would predict 40 The Equity Premium Puzzle Possibly reason 4: Irrational investor behavior • Loss aversion: Investors prefer avoiding losses more than receiving gains 41