State Space Models

All state space models are written and estimated in the R programming language. The models are available here with instructions and R procedures for manipulating the models here here.

Friday, October 19, 2012

What Caused The 1987 Flash Crash?



Today is the 25 Year Anniversary of the October 1987 Flash Crash (referred to as Black Monday) when the Dow Jones Industrial (DJI) Average lost 22.6% in one trading session. Since there was also a 2010 Flash Crash, there is some concern that flash crashes are becoming a repeating feature of Wall Street investing.

In the video above, David Blitzer, managing director of the S&P 500 committee, thinks flash crashes can happen again when everyone decides that the world is over-valued and rushes for the exits simultaneously, selling off all their positions. Today, there are circuit breakers put in place as a result of the 1987 Flash Crash, but the driving bubble-panic dynamic is still part of stock market psychology.

These comments are of interest to the Random Stock Walker for a number of reasons. If the DJI is a random walk, then anything can happen at any time in response to some shock. If the DJI is, on the other hand, attractor driven (a random walk has no attractor) then was the flash crash the result of movement off the attractor, that is, a stock market bubble?

To test these ideas I ran four models through the Random Walk estimator (using the rw package, which I will describe in a future post). The four models estimated were a random walk (RW), a business-as-usual (BAU) model, a model where the DJI is driven by the US economy and a model where the DJI is driven by the World economy. The US economy model was identified as the best model (using the AIC criterion) and from the predicted vs. actual plot above, it did a good job of predicting the DJI over the late 20th century.
Using this model, I then ran an attractor simulation. An attractor simulation forecasts the model starting from the initial conditions in 1950 rather than the step-ahead predictions form last period's values displayed in the first graphic. The attractor simulation, displayed above, shows that the 1987 Flash Crash occurred after a multi-period movement away from the DJI attractor (dotted red line). What is unusual is not the movement away (the same occurred in 1983) but the sharp "crash" in October of 1987. What is also unusual was the more than two years necessary for the DJI to return to its attractor path after 1990.

The Random Walker models do not provide an answer to the last two questions, except to point out that movement away from the attractor path eventually results in a return to that path. The causes of the 1987 Flash Crash (here) include program trading, overvaluation, illiquidity and market psychology. Since the crash is thought to have started in Hong Kong, spread to Europe and finally hit the US, macroeconomic causes have also been investigated. 

We need to distinguish between the causes of the bubble and the causes of the collapse. Since the best model for the DJI is driven by the US economy, internal US market dynamics are the best explanation for the movement away from the attractor. The shock for the crash itself may well have been transmitted from the World economy. The rapid crash seems likely to have been caused by programmed trading (algorithms rushing for the exists), however, psychology must have also been hurt given the prolonged period after 1988 when the market could not get back to its attractor path.

From the perspective of the Random Walker, the best solution for flash crashes would be to limit movement away from the attractor path. Circuit breakers will not be enough unless they work more aggressively on the way up. The counterfactual question here is whether strong circuit breakers triggered by bubbles would ultimately have kept the DJI closer to its attractor path?

For the investor, selling as DJI stocks moved away from the attractor path is the right strategy. However, how much you suffered from the Flash Crash would depend on how much of your position you sold off before the crash hit. Since the timing of a crash based on exogenous shocks is impossible to predict, it is unlikely that the strategy would prevent you from taking losses.




No comments:

Post a Comment