Does volatility targeting improve Sharpe ratios and provide crash protection across asset classes? In their May 2018 paper entitled “Working Your Tail Off: The Impact of Volatility Targeting”, Campbell Harvey, Edward Hoyle, Russell Korgaonkar, Sandy Rattray, Matthew Sargaison, and Otto Van Hemert examine return and risk effects of long-only volatility targeting, which scales asset and/or portfolio exposure higher (lower) when its recent volatility is low (high). They consider over 60 assets spanning stocks, bonds, credit, commodities and currencies and two multi-asset portfolios (60-40 stocks-bonds and 25-25-25-25 stocks-bonds-credit-commodities). They focus on excess returns (relative to U.S. Treasury bill yield). They forecast volatility using realized daily volatility with exponentially decaying weights of varying half-lives to assess sensitivity to the recency of inputs. For most analyses, they employ daily return data to forecast volatility. For S&P 500 Index and 10-year U.S. Treasury note (T-note) futures, they also test high-frequency (5-minute) returns transformed to daily returns. They scale asset exposure inversely to forecasted volatility known 24 hours in advance, applying a retroactively determined constant that generates 10% annualized actual volatility to facilitate comparison across assets and sample periods. Using daily returns for U.S. stocks and industries since 1927, for U.S. bonds (estimated from yields) since 1962, for a credit index and an array of futures/forwards since 1988, and high-frequency returns for S&P 500 Index and 10-year U.S. Treasury note futures since 1988, all through 2017, they find that:
- For U.S. equities:
- Gross annualized Sharpe ratio for a value-weighted portfolio of all U.S. stocks improves from 0.40 for unscaled holdings to between 0.48 and 0.51 with volatility scaling across different volatility forecast decay half-lives.
- These improvements in Sharpe ratio translate to higher gross cumulative returns.
- Turnovers range from zero for unscaled to about one time a year for the longest half-life and five times a year for the shortest half-life.
- Volatility scaling truncates both the left and right return tails, suppressing both crashes and surges.
- Volatility scaling improves Sharpe ratios for all 10 industries and two of three subperiods (not 1957-1987). Crash losses are shallower for all industries and subperiods.
- For S&P 500 Index futures, high-frequency (5-minute) volatility forecasting offers modest improvements in gross Sharpe ratio and crash protection over daily data.
- For U.S. bonds:
- Volatility scaling is less effective than for U.S. stocks, with Sharpe ratios and crash risks similar to that of an unscaled position.
- Using intraday data for the volatility forecast offers a slight improvement.
- For the credit index when using rapid-decay (slow-decay) volatility forecasts, volatility scaling: (1) substantially increases (decreases) gross Sharpe ratio; and, (2) has little effect on (worsens) crash risk.
- Among futures and forwards, volatility scaling improves gross Sharpe ratio only for equity index futures but reduces crash risk for nearly all.
- For both the 60-40 stocks-bonds portfolio (based on S&P 500 Index futures and T-note futures) and the 25-25-25-25 stocks-bonds-credit-commodities portfolio:
- Volatility scaling at the asset level improves gross Sharpe ratio and suppresses crash risk.
- Further volatility scaling at the portfolio level (adjusting for time variation in correlations between assets) generates further improvements.
In summary, evidence indicates that volatility scaling at both asset and portfolio levels selectively improves Sharpe ratios but pervasively suppresses crash risks across asset classes.
Cautions regarding findings include:
- Reported results are gross, not net. Accounting for daily maintenance frictions, which may be cumulatively onerous, would reduce returns for volatility scaling. Net findings may differ materially from gross findings.
- Some simplifying assumptions, such as using indexes rather than investable funds, modeling of bond returns from yields and ignoring futures rolling costs, may be materially unrealistic.
- Testing a model across different assets and testing permutations of the model (volatility forecasting decay rates) on the same data introduces data snooping bias, such that the best-performing combinations overstate expectations.
- As noted in the paper, the most detailed analyses are for the U.S. stock and bond markets, which may impound survivorship bias (have lucky histories) and therefore not be representative of future U.S. markets and other markets.
- Also as noted in the paper, other ways to manage long portfolio risk, including trend-following, may work as well as or better than volatility scaling.