Anomalies in option pricing: the Black-Scholes model revisited

New England Economic Review, March-April, 1996 by Peter Fortune

The second measure of volatility is the implied volatility. This was constructed as follows, using the data described above. For each trading day, the implied volatility on call options meeting two criteria was computed. The criteria were that the option had 45 to 75 calendar days to expiration (the average was 61 days) and that it be near the money (defined as a spread between S&P 500 and strike price no more than 2.5 percent of the S&P 500). The first criterion was adopted to match the term of the implied volatility with the 60-day term of the actual volatility. The second criterion was chosen because, as we shall see later, near-the-money options are most likely to conform to Black-Scholes predictions.

The Black-Scholes model assumes that an option's implied volatility is an optimal forecast of the volatility in SPX returns observed over the term of the option. Figure 3 does not provide visual support for the idea that implied volatilities deviate randomly from actual volatility, a characteristic of optimal forecasting. While the two volatility measures appear to have roughly the same average, extended periods of significant differences are seen. For example, in the last half of 1992 the implied volatility remained well above the actual volatility, and after the two came together in the first half of 1993, they once again diverged for an extended period. It is clear from this visual record that implied volatility does not track actual volatility well. However, this does not mean that implied volatility provides an inferior forecast of actual volatility: It could be that implied volatility satisfies all the scientific requirements of a good forecast in the sense that no other forecasts of actual volatility are better.

In order to pursue the question of the informational content of implied volatility, several simple tests of the hypothesis that implied volatility is an optimal forecast of actual volatility can be applied. One characteristic of an optimal forecast is that the forecast should be unbiased, that is, the forecast error (actual volatility less implied volatility) should have a zero mean. The average forecast error for the data shown in Figure 3 is -0.7283, with a t-statistic of -8.22. This indicates that implied volatility is a biased forecast of actual volatility.

A second characteristic of an optimal forecast is that the forecast error should not depend on any information available at the time the forecast is made. If information were available that would improve the forecast, the forecaster should have already included it in making his forecast. Any remaining forecasting errors should be random and uncorrelated with information available before the day of the forecast. To implement this "residual information test," the forecast error was regressed on the lagged values of the S&P 500 in the three days prior to the forecast.(11) The F-statistic for the significance of the regression coefficients was 4.20, with a significance level of 0.2 percent. This is strong evidence of a statistically significant violation of the residual information test.