Looking Forward Not Backward When Estimating Volatility

When you drive a car, you need to look out your front window and not the rear-view mirror. The same should be true for estimating risk in financial markets. Ironically, most of the “low volatility” products use backward looking information regardless of whether they emphasize low beta or low historical volatility. Typically, this estimation window is contained within a range of 6 to 12 months. For products that use factor models that estimate risk using fundamental factors this historical lookback might be as long as 36 months. This means that these “low volatility” portfolios tend to own stocks that had low volatility in the past rather than stocks that are expected to have low volatility in the future. The consequence is that these methods are driving using the rear-view mirror and as a result are inherently slower to react to changes in volatility. This can cause their portfolios to actually have higher volatility than their cap-weighted benchmark (which does not minimize volatility) from time to time.

If we isolate the two most popular low volatility ETFs (USMV and SPLV) we can see this trend clearly over the last six months:

Both funds do an excellent job of providing exposure to the low volatility factor at a low cost. However, by design, they directly or indirectly capture stocks that have low business risk which is proxied by long-term volatility rather than stocks that are expected to have the lowest volatility next month.  Looking backward—especially very far backward—is not optimal from a forecasting standpoint.

So how does one look forward to create volatility estimates?  After all, no one has the proverbial crystal ball.

The only way to look forward is to get the market’s forward expectation of future volatility. That can only be derived by looking at options or futures on the same underlying market. For example, using option prices we can derive the “implied volatility” for any stock and also for many different markets or indices. The best example of this method is the VIX which is the implied or forward looking estimate of the volatility for the S&P500. While the VIX itself is biased upwards due to the volatility premium for hedging, the VIX is still a better estimator of next month’s volatility than historical standard deviation.

It is well known that volatility is forecastable while returns are much more difficult to forecast. If that is the case, then what method should we use to forecast volatility?  Here are the methods we will look at below:

1. Use historical volatility;
2. Use a mathematical forecasting method like GARCH;
3. Use forward or market implied volatility from the options market.

One of the classic academic papers is a meta-analysis of all relevant volatility forecasting articles in the academic literature by Poon and Granger called: “Forecasting Volatility in Financial Markets: A Review” (2003).  The authors found that forward/market implied volatility beat the fancier GARCH family of forecasting models an incredible 94% of the time across a range of academic studies. They conclude that: “The option implied volatility being a market-based volatility forecast has been shown to contain the most information about future volatility. “This should not be surprising since the option implied forecast can use any source of information to derive a volatility estimate including GARCH models, high frequency data, or any other relevant inputs such as judgement by informed institutional traders and even inside information.

Individual stocks are even more interesting as a testing ground since equity options are even more likely than the broader indices to include other sources of information. Option implied volatility, for example, incorporates a company’s exposure to different risk factors such as value, momentum and size or quality, in addition to historical volatility and GARCH forecasting models. But perhaps the more important source of additional information comes from forward-looking expectations for earnings. This makes implied volatility like the Vegas odds for sporting events – they already incorporate a wide range of information and are therefore much more efficient than trying to forecast the odds yourself using a simple model.

For comparison let’s look at forming concentrated portfolios (analogous to minimum variance portfolios) using the different volatility metrics below.  We will take the top 20 lowest volatility stocks from the survivorship-free S&P500 index using each method and rebalance monthly.

1. Forward/implied volatility using 30-day at the money options;
2. Backward – fancy volatility forecasting model using GARCH;
3. Backward – simple trailing 1-year historical volatility.

We can see that forward-looking volatility using the implied volatility forecasts from the option market outperform backward-looking volatility across a variety of relevant metrics. Most importantly, we would want to see that volatility is lower which indicates superior forecasting ability. In this case, the forward-looking measure is the minimum volatility portfolio, having a volatility that is roughly 5% lower than the backward-looking methods. We also see other benefits to forward looking volatility:  returns and risk adjusted returns are higher, while tail risk measures like conditional value at risk (95^th percentile of returns) are lower. From a factor perspective, the Fama-French model shows that it has the highest alpha which indicates that forward looking measures are capturing more unique sources of risk information than the backward-looking models.  Finally, the information ratio – which measures return relative to tracking error to the S&P500 – is also highest for the forward-looking volatility portfolio. The frequent challenge with low volatility portfolios is that they tend to have higher tracking error – in other words they don’t often move in sync with the index. The forward-looking low volatility portfolios have lower tracking error and higher returns (and hence a higher information ratio), and therefore are a better value for someone that is looking for an S&P500 replacement.

 Conclusion

 It pays to look forward instead of backward when estimating volatility.

The evidence is supported by academic research and our own research using S&P500 stocks. The caveat is that you need to concentrate (much like other factor studies) by holding fewer stocks to get the lion’s share of the benefit of using this information versus using historical measures. Forming concentrated low volatility portfolios may enhance a variety of portfolio metrics, but we believe that the added transaction costs of managing such a portfolio would reduce many of these benefits on larger portfolios when traded using large pools of capital. That is why traditional low-volatility ETFs that are more diversified like SPLV and USMV are still valuable and practical ways to access the low volatility anomaly despite using backward-looking information. They remain excellent core low volatility ETFs because of their low management fees and portfolio design.

The problem with low volatility portfolios is that they constrain both your downside and your upside by design. Philosophically, we believe in creating greater asymmetry in performance by maximizing upside versus downside. The real value of the forward-looking measures is to extract the unique information provided by the “volatility surface” of option data. By looking at this information you can form more intelligent portfolios by differentiating between forward estimates of good (up) versus bad (down) volatility. You can use this information in different ways to design both better low volatility portfolios also to create higher return portfolios. In the next post we will demonstrate some applications.