The efficient market hypothesis Stats 413: Applied Regression Analysis Winter 2017 1 The US Treasury bills market • US Treasury bills are essentially IOU’s by the US government that promises the pay the holder of the bill a specified amount (the face value) on a future date (the maturity date). Investors make a profit by buying the bill at a discount. For example, if you purchased a $1000 T-bill that matures at the end of month t at the start of the month for $990, you can redeem the T-bill at the end of the month for $1000, making a $10 profit. • The US government sells T-bills to raise money to fund its operations. They are attractive to investors because they are a very low-risk investment. Unless the US government defaults on its debts, investors will be able to redeem their T-bills at maturity. The flip side is the rate of return on T-bills is lower than that of many other higher-risk investments. • Notation: – vt is the price of a T-bill at the beginning of month t. The price is at most the face value of the bill. – Rt is the nominal rate of return on the bill over month t. By nominal, we mean unadjusted for inflation. For example, if you purchased a $1000 T-bill that matures at the end of month t at the start of the month for $990, your rate of return over month t is 1000 − 990 Rt = ≈ 0.01. 990 – Pt is the Consumer Price Index (CPI) at the beginning of month t. The change in the CPI is a measure of inflation. 1 – πt+1 is the inflation rate over month t: πt+1 = Pt+1 − Pt . Pt – π bt+1 is the forecast of the inflation rate over month t. – t+1 is the forecast error: πt+1 − π bt+1 . – rt+1 is the real rate of return over month t: rt+1 = vt 1 Pt+1 − Pt vt Pt = 1 + Rt − 1 ≈ Rt − πt+1 . 1 + πt+1 The nominal rate of return Rt is not adjusted for inflation that occurs over the maturity term. The real rate of return rt+1 adjusts for inflation. – b rt+1 is the forecast of the real rate of return: b rt+1 = 1 + Rt − 1 ≈ Rt − π bt+1 . 1+π bt+1 The subscript on non-hat variables indicates the first month in which the variable is known. For example, the inflation rate πt+1 is only known at the beginning of month t + 1 because it depends on Pt+1 . On the other hand, the rate of return on the bill Rt is known at the beginning of month t because it only depends on the price vt and face value of the bill. • To keep things simple, we only consider T-bills whose maturity term coincides with the sampling frequency of the data (monthly). 2 Fama’s efficient market hypothesis • Fama’s efficient market hypothesis posits: – rational expectations: the forecast of the inflation rate assimilates all available information: π bt+1 = E πt+1 | It , It := {Rt , Rt−1 , . . . , πt , πt−1 , . . . }, where the information set It includes all the variables known at the beginning of month t. 2 – constant real interest rate forecasts: the forecasts of the real rate of return is constant: b rt+1 = r0 for all t. • The first condition is a formal statement of the hypothesis that the market “fully assimilates” the information at time t: the conditional expectation πt+1 is the minimum MSE predictor that depends only on the information in It . • Under Fama’s hypothesis, the forecast error sequence {t } is a MDS. Indeed, we have E t+1 | It = E πt+1 − π bt+1 | It = E πt+1 | It − E π bt+1 | It = E πt+1 | It − E πt+1 | It (rational expectations) = 0. (2.1) At this point, we wish to appeal to the tower property to conclude E t+1 | t , t−1 , . . . = E E t+1 | It | t , t−1 , . . . = 0. To justify the first step, we need to show that It contains at least as much information as {t , t−1 , . . . }. The formal way of establishing this is to show s , s ≤ t is a function of the variables in It . Indeed, we have s = πs − π bs = πs − E πs | Is , s ≤ t which is a function of the variables in It . We conclude {t } is a MDS. • The intuitive interpretation of (2.1) is the forecast error t+1 does not depend on It . After all, if the forecast assimilates all the information in It , then forecast error should not depend on It . • Since MDS’s are serially uncorrelated, one possible test of Fama’s hypothesis is to test whether the forecast errors are in fact serially uncorrelated. Testing for serial correlation is a topic that is taught in courses on time series analysis, and we skip it here. • The real interest rate has a constant expected value and is serially 3 uncorrelated. We have rt+1 = Rt − πt+1 = Rt − π bt+1 + π bt+1 − πt+1 = r̂t+1 − t+1 (definition of r̂t+1 and π̂t+1 ) = r − t+1 . (constant real interest rates) (2.2) Since {t } is a MDS and a MDS is centered and serially uncorrelated, we have E rt = r + E t = r, cov(rs , rt ) = E (rs − r)(rt − r) = E s t = 0, s 6= t. • To derive a testable implication from Fama’s hypothesis, we combine the definition of πt+1 with (2.2) and (2.1) to obtain E πt+1 | It = E Rt − rt+1 | It (definition of rt+1 ) = E Rt − r + t+1 | It (2.2) (2.3) = Rt − r + E t+1 | It = Rt − r. (2.1) Since Rt ∈ It , we have E πt+1 | Rt = E E πt+1 | It | Rt = Rt − r, which suggests a one-for-one relationship between inflation and the nominal interest rate. 3 Is the US T-bills market efficient? • In summary, Fama’s hypothesis suggests the regression coefficient of Rt is one. Before we test this restriction, we check the conditions of the linear model are implied by Fama’s hypothesis. T • Linearity: Let yt = πt+1 and xt = 1 Rt . By (2.3), we have E πt+1 | It = xTt β ∗ + t , T β ∗ = −r 1 . • Predetermined predictors: The predetermined condition is 1 t+1 E t+1 = E = 0. Rt Rt t+1 4 Thus we need to show E t = 0 and E Rt t = 0. The former holds because {t } is a MDS. The latter holds because E Rt t = E E Rt t | It = E Rt E t | It (Rt ∈ It ) =0 (2.1) • Ergodicity and stationarity: The stationarity condition is problematic because the nominal interest rate Rt , as shown in Figure 1, has an upward trend. For now, we ignore this empirical evidence to the contrary and proceed. It is hard to check the ergodicity condition in practice, and we just have to assume it. • The second moment of the predictors is " # T 1 1 1 Rt E =E . Rt Rt Rt R2t Its determinant is 2 E R2t − E Rt = var Rt , which is non-zero. • We know {t , t−1 , . . . } ⊆ It . Thus {xt t , xt−1 t−1 , . . . } ⊂ {xt , xt−1 , . . . , t , t−1 , . . . } ⊆ It . By the tower property, we have E xt+1 t+1 | xt t , xt−1 t−1 , . . . = E E xt+1 t+1 | It | xt t , xt−1 t−1 , . . . (It has more information than {xt t , xt−1 t−1 , . . . }) = E xt+1 E t+1 | It | xt t , xt−1 t−1 , . . . (Rt ∈ It ) = 0, which implies {xt t } is a MDS. • Having checked that the conditions of the linear model are plausible, we proceed to test H0 : β2∗ = 1. We use the dataset in [?] which includes the one-month T-bill rate and CPI from February 1950 to December 1990. Figure 1 shows the data. 5 Figure 1: Inflation πt and interest rates Rt (Figure 2.3 in [Hay00]) • To replicate Fama’s results, we restrict to the data from January 1953 to July 1971 (n = 223). We use the heteroskedasticity-robust estimator of the asymptotic variance −1 P P −1 1 1 T 2 x xT T d βbn = 1 P ˆ Avar x x x x i i i i i i i i∈[n] i∈[n] i∈[n] n n n = n(XT X)−1 XT diag(ˆ 2 )X(XT X)−1 , which leads to heteroskedasticity-robust standard errors. By default, R does not return heteroskedasticity-robust standard errors, and we must ask for them specially. lm(formula = inf ~ TB1, data = fama) Residuals: Min 1Q Median 3Q Max -7.1928 -1.9386 -0.2488 1.8967 8.4651 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) -0.8678 0.4332 -2.003 0.0464 * TB1 1.0147 0.1127 9.003 <2e-16 *** --Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1 6 Figure 2: Real interest rates rt (Figure 2.4 in [Hay00]) Residual standard error: 2.843 on 221 degrees of freedom Multiple R-squared: 0.2364, Adjusted R-squared: 0.2329 F-statistic: 81.79 on 1 and 221 DF, p-value: < 2.2e-16 • We see that 1 is well within two standard errors of βb2 , which implies the two-sided z-test of H0 easily accepts. 4 Subsequent developments • Studies after the publication of Fama’s paper showed that the one-forone relationship between inflation and interest rates is limited to the period between the end of World War II and 1979. Further, Figure 2 shows that the second part of Fama’s hypothesis, that the real interest rate is constant, is implausible post-1979. • If we restrict to the data post-October 1979, the interest rate coefficient in (2.3) is significantly smaller than one. lm(formula = inf ~ TB1, data = afterFama) 7 Residuals: Min 1Q Median -11.4406 -2.8608 -0.6087 3Q Max 2.6677 16.8705 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 1.57500 0.69340 2.271 0.024 * TB1 0.66478 0.09997 6.650 2.1e-10 *** --Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1 Residual standard error: 4.347 on 231 degrees of freedom Multiple R-squared: 0.1391, Adjusted R-squared: 0.1354 F-statistic: 44.6 on 1 and 231 DF, p-value: 1.782e-10 • This raises the question of why the one-for-one relationship between inflation and interest rates vanishes in the post-1979 data. One possible explanation by Mishkin is the inflation rate affects interest rate gradually. Thus transient fluctuations of inflation have little effect on interest rates, but sustained trends in inflation affect interest rates. • We see in Figure 1 that there is a sustained upward trend from 1950 to 1979, which is also the period when there is strong correlation between inflation and interest rates. References [Hay00] Fumio Hayashi. Econometrics. Princeton University Press, 2000. 8