Is It Possible to Forecast Large Market Selloffs?

How do you know when a selloff ends?

My definition of selloff is a market drop of 15% or more from a prior peak. The selloff ends once a 10% retracement has taken place. Any subsequent selloff constitutes a brand-new selloff episode.

The blue line in the top panel of the next chart shows the cumulative return on a dollar invested in the S&P 500 index in 1960. The orange line shows the life-to-date maximum return earned by the initial $1 investment. When the blue line dips below the orange line, we call that a market drawdown. This drawdown, in percent, is shown in the bottom panel of the figure.

Selloff starts are indicated by the rightward pointing green triangles. Selloff troughs are identified by downward pointing purple triangles. Keep in mind that you would not have been aware of a selloff trough in the month in which it took place because you would not have known whether the market would sell off in the next month or not.

The leftward pointing red arrows indicate the ends of selloff episodes, which occur following a 10% or higher retracement from the trough. Because of this 10% retracement rule you would have known about the end of the selloff in the month in which it took place. When I analyze post-selloff returns, I only start counting after the first 10% of the post-trough bounce has already taken place.

AI and large market selloffs

To find the association of current economic and market conditions with future selloffs, I first identify 12-month periods with a start-to-trough selloff of at least 20% (this is a slightly different definition of selloff than in the prior section, but it generates many more selloff episodes to train the model).

The blue line in the above chart shows past occurrences of 20% or larger selloffs in the subsequent 12 months, and the red and orange series show the model probabilities of such selloffs. The orange line shows the forecast from a simple, one-layer neural network designed to predict such selloffs. The model takes as inputs several market and macroeconomic variables, and identifies an association between these and future large selloffs.

While the simple neural network captures several of the selloff events, it does a poor job with the 1987 market crash. It also predicts some historical selloffs which never took place, such as its large selloff forecast just prior to 2000.

One of the key advances in machine learning over the last several decades has been the advent of deep learning, which involves training neural networks with multiple layers, to better capture complex, real-world relationships. The red line in the above figure shows the large selloff forecast resulting from a multilayer neural network model. While this also is not perfect, it does a (slightly) better job with the 1987 crash, and shows considerably fewer false positives, i.e., forecasts of large selloffs that never materialized.

Of course, neither the simple neural network model, nor its deep learning cousin, is perfect. All these models can do is find a relationship between forecasting variables and past selloff episodes. It is possible – indeed, likely – that future selloffs will happen for very different reasons, and that our market and macroeconomic forecasting variables may not be rich enough to capture the relevant state of the world for forecasting future selloffs. That being said, the models can answer the question of how likely it is for a large selloff to occur, if it were to occur for a reason that was observed in our training sample.

With these caveats in mind, both of our forecasting models are currently generating benign large selloff probabilities. The simple model forecasts an 11.4% percent chance of a large selloff, while the deep neural network model — my preferred model — forecasts only a 0.9% percent chance of a large selloff. The model is sanguine because there are offsetting effects from the forecasting variables. Despite the fact that some dimensions of the market environment are suggestive of large future selloffs, other dimensions are not.

For example, I find that high two-year stock returns are associated with an elevated probability of large future selloffs. And while lagged two-year stock returns are indeed high, there are offsetting effects. For example, the 2s10s curve (the difference in yields between 10-year and 2-year Treasury bonds) was often very steep (i.e., 10-year rates high relative to 2-year rates) prior to past selloffs, and this is not currently the case.

Similarly, economic news sentiment prior to past selloffs has typically been negative, and economic sentiment today is benign. The net result of these offsetting effects is that the model’s large selloff forecast is currently quite low.

What happens next?

The next table shows the return on U.S. stocks in 12-month periods which followed the ends of prior large selloffs. Keep in mind that large selloffs end strictly after the first 10% of the post-trough bounce has already taken place. Thus, the ends of large selloffs are identifiable in real time.

Of the 13 large selloffs seen in U.S. markets in the prior 60 years, 12 of them were followed by strongly positive returns over the subsequent 12 months. Only once – during the bursting of the dot-com bubble – was a large selloff followed by another large selloff. In 12 of the 13 cases, large selloffs were followed by large rebounds.

The average post-selloff 12-month return is 21.5%. And this is after leaving the first 10% of the rally on the table. Those finding themselves in possession of investable assets in the aftermath of large selloffs are likely to be investing into a good future environment for stocks.

Our conclusion is that it is better to manage the risk in your investment portfolio in a way so that, when the inevitable selloff does come, you are not left scrambling to liquidate your positions and being forced to move to cash. Instead, if you were able to properly size your risk taking prior to the selloff, you may be able to avoid liquidating and instead add to your stock exposure at a time that is historically associated with high future returns.

For our part, we will continue to regularly update our large selloff forecasting model and will share the results with our clients. No model is guaranteed to be able to forecast future selloffs in real time, because these often happen for reasons that have never before been seen. Nevertheless, it is useful to have a mapping from market states to future selloffs, and to use this lens to monitor the current state of markets.

Harry Mamaysky is a professor at Columbia Business School and a partner at QuantStreet Capital.

QuantStreet is a registered investment advisor. It offers wealth planning, separately managed accounts, model portfolios and portfolio analytics, as well as consulting services. The firm’s approach is systematic and data-driven, but also shaped by years of investing experience. To work with or learn more about QuantStreet, contact the firm at [email protected].

---

A message from Advisor Perspectives and VettaFi: To learn more about this and other topics, check out some of our webcasts.

More Separately Managed Accounts Topics >