No-Arbitrage Testing with Single Factor Presented by Meg Cheng Motivation • No-arbitrage condition is one of the most popular assumptions in the area of asset pricing. • If changes in the price of the asset are driven by some underlying factors, the excess return should be consist of all the prices of the associated underlying factors risks. • To be free from any model specification, nonparametric estimation method is adopted to recover the embedded information directly from the data. Bond Pricing Theory Suppose the economy can be driven by a state variable X, defined in a stochastic differential equation: dX (t ) x ( x)dt x ( x)dW Consider the dynamics of an asset price at current time t of a claim to terminal payoff P(T) at some future date T as follows: dP(t ) ( x)dt ( x)dW By Ito’s lemma, 1 ( x) Pt Px x Pxx x2 2 ( x) Px x Now, consider another asset price w/ terminal payoff as P*(T), within the same framework, we can write this dynamic process as dP* * ( X )dt * ( X )dW 1 * 2 * * * () Pt Px x Pxx x 2 * () Px* x If we want to make a portfolio Z to hedge against the factor risk with these two assets, then portfolio Z can be represented as follows: Z P * P * dZ dP *dP * [ * * ]dt [ * * ]dW where and * are portfolio weights on each asset Let * * , * * Then portfolio Z becomes a risk free asset. Hence, the drift term of dZ should be equal to risk-free rate r. i.e. * * r * r * r ( x) * Notice that if the underlying market is arbitrage free, this relationship holds for any arbitrary asset. This addresses our null hypothesis. i.e. 1(x)= 2(x)= …=P(x) Hypothesis testing We choose short rate as the factor, and use the three-month yield to maturity as the proxy for the short rate. Suppose we have P different assets to be estimated, and each one of them follows: dPi Xdt ai ( x)dt bi ( x)dW (t ) i 1, 2,..., P Pi ai (.) and bi (.) are local constants. Given each x, we can use gaussian process to describe the above diffusion process. Under the alternative hypothesis: L1 (a( x), b( x)) log( f i [ai ( x), bi ( x)]) K h ( X x) n i (a, b) arg max L1 (a( x), b( x)) a ,b where a( x) ai ( x) i 1,2,..., P b( x) {bi ( x) i 1,2,..., P} f i () is Gaussian density function K h () is kernel function Under the null hypothesis: If no-arbitrage restriction holds, the expression below is true for any arbitrary asset: r * r ... ( x) * So that given each x, the risk premium of each asset should be proportional to its diffusion term with a constant term across all assets. dPi Xdt [c( x)bi( x)]dt bi( x)dW , i 1, 2,..., P Pi Hence, the likelihood function under the null: L0 (b( x), c( x)) F (b( x), c( x)) K h ( X x) n (b, c) arg max L0 (b( x), c( x)) c(x) is a constant across all the assets F(.) is multivariate gaussian density function b( x) {bi( x), i 1,2,..., P} Data • We use weekly values for the annualized zero-coupon yields with six different maturities (0.5, 1, 2, 3, 5, 10 years). • Generally speaking, almost each bond/security comes w/ coupons or dividends, except treasury bills. • Since there is no generally accepted “best” practice for extracting zero coupon prices from coupon bonds, we construct our data by four methods: 1. Smoothed Fama-Bliss 2. Unsmoothed Fama-Bliss 3. McCulloch-Kwon 4. Nelson-Siegel •To test our null hypothesis, we propose to use empirical Likelihood Ratio (LR) test, since we’ve already constructed likelihood both under the alternative and the null. •We interpolate all the estimates associated with the chosen grids to compute the likelihood at each observation. •To get LR test statistics distribution under the null hypothesis, We adopt stationary bootstrap method proposed by Romano (1994). The procedures are described as follows: 1. We use first order Euler approximation to fit the model under the null, i.e. dPi [ X cˆ( X )bˆi( X )] bˆi( X )ˆti i 1,2,..., P Pi Since we don’t literally have maximum likelihood estimated on every data point, there still exists some dependence in time in the residuals extracted from the above. 2. Have all the residuals estimated from each asset into a matrix by columns and denote it by Y (NxP). 3. Let i be i.i.d. random variable generated from Uniform Distribution U(N). 4. Generate Bi,m={Yi, Yi+1, …,Yi+m-1}’ , the block consisting of m rows starting from Yi, and the r.v. m is drawn from geometric distribution (1-q)m-1q for m=1,2,…N. where qR(0,1). 5. Repeat step 3 and 4, stack each block matrix end to end, till the number of columns and rows of the newly generated Y* are equal to Y. 6. Put Y* back to the Euler euqation to get the new LHS. 7. Implement local MLE both under the null and the alternative. 8. Replicate step3~step7 for sufficiently enough time (around 1,000), then the statistics distribution will then be constructed. Nelson-Siegel 3 3 2 2 1 1 lamda lamda Unsmooth Fama-Bliss 0 -1 -2 -1 0 5 10 15 short rate % Mculloch-Kwon -2 20 6 0 5 10 15 short rate % Smooth Fama-Bliss 20 4 4 2 2 lamda lamda 0 0 0 -2 -2 -4 -4 0 5 10 15 short rate % 20 0 5 10 15 short rate % 20 0.8 autocovariance autocovariance 1 0.5 0 -0.5 0 100 200 lags 300 0.4 0.2 0 -0.2 400 0 100 200 lags 300 400 0 100 200 lags 300 400 0.8 autocovariance 1.5 autocovariance 0.6 1 0.5 0 -0.5 0.6 0.4 0.2 0 -0.2 0 100 200 lags 300 400 Nelson-Siegel 10 50 8 40 frequency frequency Unsmooth Fama-Bliss 6 4 2 0 30 20 10 0 500 statistics Mculloch-Kwon 0 1000 60 0 500 statistics Smooth Fama-Bliss 1000 50 40 frequency frequency 40 20 30 20 10 0 0 200 400 600 statistics 800 0 0 500 1000 statistics 1500 Unsmooth Fama-Bliss Nelson-Siegel L(1) -8345 -8162 L(0) -8474 -8318 T*=L(1)-L(0) 129 156 No. of replications 100 500 Prob.>T* 67% 73% Mculloch-Kwon Smooth Fama-Bliss L(1) -6035 -8138 L(0) -6289 -8228 T*=L(1)-L(0) 254 90 No. of replications 600 500 Prob.>T* 33% 69% Conclusion: So far, based on our result, the hypothesis of no-arbitrage condition tested with six different yields to maturities is not rejected. Put in another way, the no-arbitrage restriction may still holds in U.S. Treasury Bill and Bond market.