The secret of 'timing' success
Market timing is certainly a perilous activity and perhaps a foolhardy one. However, varying market class weights in response to changing forecasts for asset class returns covers a wide range of strategies – from aggressive short-term tactical asset allocation programmes through to multi-year reviews of strategic policy. These are very distinct in aim and execution, and yet formally all fall under the rubric of timing. It may be that they all deserve equally to be damned, but a formal analysis of what is required for success suggests otherwise.
An important part of the case against any form of market timing is that diversification across time is not the same as diversification across asset classes. A policy that involves timing will be more volatile than a policy that does not, even if both have the same average asset allocation. The more volatile policy must have a lower geometric mean return and the difference will be directly proportional to the amount of additional volatility induced by timing. The cumulative return of a strategy is directly related to the geometric mean and so, other things being equal, a timing strategy by inducing volatility starts at a disadvantage even if there are no transactions costs. A timing policy will underperform the benchmark selected for comparison purposes, and a certain amount of skill is needed just to break even. This is a result that holds even if the returns on the asset classes are assumed (unrealistically, of course) to be fixed. While such return streams would violate even the most forgiving asset pricing model, the point is that what drives this result is the volatility that switching induces in the portfolio, not the volatility of the underlying asset class as such. Given this, the larger the asset class swings, the greater the volatility of the strategy and therefore the larger the asset switching premium.
More prosaically, a timing strategy may well involve higher transactions costs than a simple rebalancing strategy. This is not inevitable. A simple strategy of letting the weights drift, assuming that returns are positively serially correlated would be a timing strategy and would involve no transaction costs. Similarly relative to a policy of monthly periodic rebalancing, an annual review might – depending on the size of switches – have lower transaction costs. But most of the time a timing portfolio will have higher transaction costs.
A natural benchmark for a market timing strategy is a static asset allocation. Since a static allocation cannot be implemented without some form of rebalancing strategy, this suggests that a fixed-weight benchmark rebalanced at the start of each time period is a reasonable starting point. Thinking through the impact of the asset switching premium and the impact of transaction costs is best done with minimal assumptions about the model used to predict timing. The idea is to figure out the breakeven level of skill required to make market timing work under a given set of assumptions about the risk structure of market returns. Repeating such an analysis over a range of assumed frequency and size of switch assumptions generates insight into how high the hurdles are for differing policies.
The analysis here is based on 50 years of monthly data on the S&P and US Treasury returns. The results are based on a simple strategy that takes positions of a given size relative to a 50% equity/50% bond benchmark. By varying the frequency with which the positions are adjusted and by also varying the size of the positions, some insight into the impact of timing on a portfolio can be generated. The actual mean returns and correlations were used as inputs to a Monte Carlo simulation. This simulation used 1,250 years of simulated monthly data. The Monte Carlo methodology is needed here as we want time series that actually follow a random walk. Results generated on the assumption that markets follow a random walk make a good baseline for measurement. Of course, if it turns out that markets really do follow a random walk at a given investment horizon then these results are also a good guide as to what to expect in practice. The results are expressed as a ‘gross’ timing premium which would need to be earned back even if transactions costs were zero, and a net premium which additionally incorporates the impact of transaction costs, assumed to be a 7.5 basis points spread. So these spreads show how much the timing strategy can be expected to underperform a simple rebalancing strategy on an annualised basis if the asset returns follow a random walk. Figure 1 shows that the relationship is non-linear with respect to the time horizon. As the time horizon increases, the transaction cost component falls and the gross and net numbers get closer. The pure timing premium peaks at an intermediate time horizon before falling very markedly as time horizons change to years rather than months.
This analysis assumes a random 24% switch in weights away from the benchmark at every rebalancing point. As would be expected, larger switches lead to a higher asset switching premium. Table 1 shows how the asset switching premium varies with the size of asset class switch assuming annual rebalancing.
The premium increases more than linearly with the size of the switch at a given time horizon. The conclusions to be drawn from this simple Monte Carlo experiment are conditional in some details of the switching policy but the key points are robust to the form of policy: a timing policy starts with a deadweight loss that is a function of the volatility induced by switching. The size of this deadweight loss increases with the size of switch taken and tends to peak at an intermediate time horizon for a given size of switch.
This information is one half of a trade-off. Ideally, it would be nice to know how predictability varies with time horizon. Obviously, if short-term equity returns were easy to forecast then no one would care about the high net of transaction cost asset switching premium. Conversely, if longer horizon returns were clearly random, then the lower asset switching premium at long horizons would be cold comfort.
In practice, of course, this cannot be answered for every possible model that might be used to try and market-time. However, there are a few portmanteau techniques that can help shed light on the challenges facing any model at a given forecasting horizon. One technique that has been extensively used is the variance ratio test. If markets are truly efficient then asset prices should follow a random walk or something close to it. It does not of course follow that the market is efficient if an asset price follows a random walk, but it is a fair working assumption. The variance ratio test exploits the fact that if a price series is a random walk then the variance of six-month terms should be six times the variance of one-month returns and so on. If the price process really is a drift then it should be a drift no matter what horizon it is measured over and the spread of prices five years from now should look the same if predicted using monthly annual or any other frequency of returns. On the other hand, if the variance ratio is below one at longer horizons then the inference would be that there is mean reversion in returns at these horizons. This mean reversion would reduce the spread of prices and thus produce a lower estimate of variance when longer horizon returns are used.
Figure 2 shows the results of applying the variance ratio test to the 50 years of S&P data. At longer horizons the variance ratio does fall well below one. Just as importantly at these same horizons the likelihood that the returns are a random walk falls very markedly. So while at short horizons it is not possible to reject the random walk hypothesis at any reasonable level of significance, at two and a half years and above the random walk can be rejected at 10% or in some cases 5% significance levels. Simply put, at longer time horizons there is strong evidence of mean reversion in stock returns.
This is useful to know because it suggests that at shorter time horizons where returns are close to a random walk the asset switching premium analysis is not only a reasonable theoretical guide but a fair forecast of what one could actually expect to achieve by market timing: a loss. At longer horizons, though, the increased predictability of stock markets suggests that it may be easier to beat the asset switching premium, particularly when the switches are in the lower half of the size range analysed here, because returns are easier to forecast and the asset switching premium is much lower in any case. These results may understate the case for long-horizon market timing. At longer horizons, there is more chance of being able to implement the more modest changes in weights by the allocation of new money or judicious sales if money is net leaving the fund or reinvestment of income, bringing down the transactions costs.
So this suggests that in trying to determine where success is most likely to be achieved in market timing that, at least until everyone catches on to this, the longer horizons are the place to look. The secret of success in market timing is – well – timing.
Anthony Foley is managing director of the Advanced Research Center at State Street Global Advisors in Boston