Architecture Of AMSI
Figure 3 provides a summary of the five factors that comprise the AMSI. The weights of the individual AMSI factors are proprietary and dynamic, adjusting each time the index is updated. To facilitate the interpretation of the AMSI, we converted the indicator to a percentile value, ranging from 0 (extreme fear) to 100 (extreme greed). The importance of each element with respect to the significance of their weights in the AMSI, in descending order, are P/E, momentum, realized volatility, high-yield returns and TED spread. AMSI spans the period from January 1986 to the present, i.e., the longest common period over which data for all component elements is available. The absence of data prior to the mid-1980s for high-yield bonds and Libor, which is a component of the TED spread, precludes AMSI from starting at an earlier date.
Figure 4 is a table showing the weights of the different factors at different times, including maximum and minimum weightings. To determine the dynamic weights, we find the percentage ranks of the individual components’ historical data series and run correlations against rolling three-month, six-month, nine-month and 12-month trailing and forward-looking S&P returns. For each rolling data set, we calculate a weight for each component by dividing the absolute value of its correlation by the sum of all of the components’ correlations to that data set. We continue this process for each set of rolling returns. The final component weight is a simple average of all of these weights across the four periods.
A graph of AMSI and of the AMSI Six-Month Moving Average from AMSI’s January 1986 inception through November 2013 is shown in Figure 5. The dynamic and cyclical nature evokes the ocean wavelike movement described by Charles Dow and the pendulum analogy used by Benjamin Graham. Figure 5 also calls to mind the comments made by prominent present-day investor Howard Marks6 of Oaktree Capital Management. Marks said, “I believe strongly that (a) most key phenomena in the investment world are inherently cyclical, (b) these cycles repeat, reflecting consistent patterns of behavior, and (c) the results of that behavior are predictable.”
A moving average is designed to smooth volatile data series, thereby providing a more stable pattern. Moving averages of data series often enable trends to come into clearer focus, while original data series may often “flip flop” in its direction, providing mixed signals. The weakness of moving averages is they are lagged and may be delayed in identifying the turning points of an indicator. We tested a number of moving averages and found the six-month moving average provided the best combination of trend exposition and timeliness. In Figure 5, both line graphs illustrate a similar pattern, but the moving average line is more stable.