Market timing and trading strategies using asset rotation: non-neutral market positioning for exploiting arbitrage opportunities

Abstract

We present empirical results on the statistical and economic viability of a market timing and trading strategy that is based on a pairwise rotation between two risky assets. Using data on equity exchange traded funds, and models for both the returns and the volatility of the underlying assets, we compare the performance of the suggested models with the standard benchmarks of a buy-and-hold strategy and an equally weighted portfolio. The underlying intuition for the use of such a strategy rests with literature on sign and volatility predictability. The rotation strategy, as we apply it in this paper, is not risk-neutral and assumes the presence of arbitrage opportunities in the markets and short-term trends. Furthermore, the model specification uses the interplay between relative returns and relative volatilities in picking-up the asset with the highest return. Our results show that even a naive model that is based on a moving average of relative returns can outperform both benchmarks and that more elaborate specifications for the rotation model may yield additional performance gains. We also find that, in many cases, the rotation strategy yields statistically significant sign predictions of the relative returns and volatility. While our results are conditional on the data that we have used in our analysis they, nevertheless, support the market-timing literature and show that an active trading strategy can be based on the concept of rotation.

Keywords:

• Exchange-traded funds
• Forecasting
• Market timing
• Sign predictions
• Statistical arbitrage
• Technical trading
• Volatility timing

• JEL Classification:
• C51
• C53
• G10

Notes

¹ A brief review of the market timing and related literature is given in the next section.

² The proposed strategy is similar to what is called a ‘quantitative directional equity trading’ in the industry. There is a growing interest for this kind of strategy.

³ Here it’s important to note that sign and directional predictions have been found to heavily depend on volatility forecastability for which there is a rather large literature which we will not review here.

⁴ According to the 2008 Morgan Stanley ETF Global Industry Review.

⁵ According to Morgan Stanley in the US there are 288 options, which means that 47% of US listed ETFs have options.

⁶ The window size is n₀ = 104 weeks for the results presented in the next section.

⁷ Note that there is an implicit assumption here, namely that sign generation and trade execution occurs within a time interval adjacent to the closing price at time t. This is a usual way of backtesting strategies and is implementable in a straightforward way in electronic after-hours trading or via a limit order within a radius of the closing price used for sign generation.

⁸ Note that the strategy’s return depends on the realized values of the individual asset returns, denoted with small case letters, and its uncertainty comes only from the uncertainty of the forecast.

⁹ Results are given in ¢ for an investment of $1; a positive difference shows that the rotation strategy is better; compare across all three trading days.

¹⁰ See, among others, Gibbons and Hess (1981), French (1980), Conrad and Kaul (1988), Rogalski (1984) and Chordia et al. (2001) for a discussion on the day-of-the-week effect.

¹¹ Manganelli (2005) proposed a framework to model duration, volume and returns simultaneously, obtaining an econometric modelling which incorporates the interaction among these variables.

¹² The mean trading time is the proportion of total observations that a signal for a rotation was given.

¹³ Leitch and Tanner (1991) argued that the ranking of forecasts based on sign tests is closely related to their ranking of correct predictions in simple trading strategies. Pesaran and Timmerman (2005) also argued about the importance of having correct sign predictions in the context of a trading methodology. Lam and Lee (2004) suggested that a correct prediction probability should be around 60% in order for a trading strategy to be economically significant with a 0.1% transaction cost.

¹⁴ Kavajecz and Odders-White (2001) examined volatility within three related intra-day series, transaction returns, quote midpoint returns, and limit order book midpoint returns using as data span NYSE listed stocks using GARCH methodology.

¹⁵ Bootstrap-based p-values are being reported in table 8.