You are on page 1of 2

ARCH-M MODELS

Economic theory holds that investors should be rewarded for taking risks. The ARCH-M (ARCH in mean) model provides an explicit link between the risk (conditional volatility) and the best forecast of a time series. No such relationship holds for the ARMA-ARCH models. For example, suppose that yt is AR (1) with ARCH (q ) errors. In this case, we have yt = ayt −1 + εt where εt is ARCH (q ). The εt are distributed as εt ⎥ ψt −1 ∼ N (0 , ht ), where ht = ω + Σ αi εt2−i and ψt −1 is the information set at time t − 1. Suppose
i =1 q

we want to forecast yt at time t −1. Regardless of the value of the conditional volatility (risk) ht , the best forecast of yt is just the conditional mean, μt = E [yt ⎥ ψt −1] = ayt −1 , so the reward (expected future value) does not depend on the risk. The ARCH-M model is

⎯ ⎯ yt = C + θ√ht + εt ,
where εt is ARCH (q ) with εt ⎥ ψt −1 ∼ N (0 , ht ). The best forecast of yt given ψt −1 is the conditional mean

⎯ ⎯ μt = E [yt ⎥ ψt −1] = C + θ√ht , ⎯ ⎯ which is an explicit function of the risk ht . Note that the model implies that μt is proportional to √ht
rather than ht . This is sensible, since doubling all observations should double (not quadruple) the forecast. If θ is positive, then μt increases with ht : the reward increases with the risk. Note, however, that the reward is not guaranteed: High volatility (large ht ) implies that yt is expected to be large, but not that yt is guaranteed to be large. So reward is measured here by the expected future value μt , not by the actual future value yt .

μt would represent the "risk premium" necessary to induce a risk-averse agent to hold the longer term asset. and x is the time series. Lilien and Robins (1977). The ARCH-M model has been used to investigate the term structure of interest rates. As long as {εt } is stationary. and hence {μt } and {yt } are stationary as well. q is the number of arch parameters. Engle.mean) x c" will carry out the estimation. Here. and εt is an ARCH (q ) shock which was completely unforecastable from time t −1. . {ht } must also be stationary. θ was statistically significant) for a series of excess returns on 6 month treasury bills compared to the return on two consecutive 3 month treasury bills. The {yt } are not white noise. The parameters of the ARCH-M model can be estimated from data using the maximum likelihood method. but the best possible forecast μt is a nonlinear one. The tsp command "arch(nar=q.-2- We have shown that the ARCH-M model can be written as yt = μt + εt . so they are linearly forecastable. On the other hand. the failure of the forecasts to depend directly on the autocorrelations may be a drawback of the model. Here. Note that μt is the best forecast of yt which could have been made at time t −1. showed that there were significant ARCH-M effects (that is. in their paper which introduced the model. where ⎯ ⎯ μt = C + θ√ht .