Abstract
This article explores the role of trading volume in making out-of-sample forecasts of stock market volatility around the time of the 24 October 1929 crash. Following the recent literature on volatility forecasting, we compare the performance of symmetric and asymmetric GARCH-class models. Moreover, as the volume–volatility relationship is now well established for modern day markets, we also consider the performance of these models when volume is allowed to enter the conditional variance equation. Given the institutional evidence that trading volume was beginning to take on an increasingly important role in the eyes of investors and market regulators during the last part of the 1920s, this is a particularly insightful endeavour. Generally speaking, the volatility models with trading volume provided the best volatility forecasts after ‘Black Thursday’.
Acknowledgements
The authors thank session participants at the 2003 Southern Economic Association meeting for many useful comments on an earlier version of this article.
Notes
1 In 1907, the most important nonbank intermediaries were trust companies. Although trust companies continued to play an important role in leveraging stock purchases during the 1920s, broker's loans emerged to replace trust companies as the dominant form of nonbank financial institution.
2 Using the structural break found by Kim et al. (Citation1991), Hiemstra and Jones (Citation1994) examined two distinct periods, 1915–1946 and 1947–1990. We consider the 1920s to be a subset of Hiemstra and Jones's early period.
3 The stock market peaked on 3 September 1929, and then began to slide, reaching its trough on 13 November 1929. The most dramatic decline occurred during the month of October, culminating in the massive sell-off that took place between 24 and 29 October.
4 In studies of equity markets this is often referred to as the ‘leverage’ effect since negative shocks to returns (i.e. bad stock market news) may impact highly leveraged firms more than others.
5 See the time series econometrics texts of Harvey (Citation1994) for an overview of the Box–Jenkins method for specifying an ARMA model.
6 We computed the quasi-maximum likelihood covariances and SEs as described in Bollerslev and Wooldridge (Citation1992). The model is estimated under the assumption that the errors are conditionally normally distributed.