Methodology
  • We currently run our models on equity and bond funds, specifically on all major, primary funds in a country. Our methodology applies to country allocation as well as sector rotation. We are looking for funds that are benchmarked against a country’s major benchmark. For this reason, we exclude sector funds and specialized funds that do not really match the broad country benchmark.
  • We also focus exclusively on funds that exhibit a high correlation to the benchmark. In the U.S., we put our cut-off at a R2 of 0.85. This number was determined by looking at the correlation of the S&P Smallcap index vs. the S&P 500. Over the past seven years, the R2 of weekly returns between the two indexes stood at 0.82 However, when looking at the performance of those indexes, one has to conclude that they are two very different animals. By putting our R2 cut-off at 0.85, we allow for some difference in fund returns with the benchmark while maintaining comparability.

USL GN 00005

  • At the other end of the spectrum, we also exclude index funds that have a R2 of 1 with the index. These funds do not add anything to our data. We already know their sector weightings.
  • Our dynamic model searches for the most likely weight distribution that explains recent performance. Our model allows for weights to move from day to day but favors weight stability over time.
  • Our tracking model is not a rolling regression. The problem with this type of analysis is that out-of-sample weights can jump significantly from one day to another. Also, regressions result in a fixed weight estimate for over several days, which is unlikely in reality.
  • Once we have calculated daily weight estimates for each fund, we aggregate those numbers at the national level. For weights by quartile or decile, we determine percentile using total returns over the previous three months, on a rolling basis. For style funds, we use the self-reported style classification and apply an average by style.
  • Because, sometimes, funds miss the deadline and do not report their pricing until the next day, we make sure to recalculate weights retroactively to include those laggers when their NAV data is available. To minimize data volatility from this phenomenon, we apply minimal smoothing in the form of a five-day moving average to our data.
  • Improving the accuracy is an ongoing process for us. We are constantly trying to minimize our tracking error with actual fund performance. However, we also make sure to minimize the probability overfitting with a cross-validation process. New models have to improve results on a small portion of our data before we try it on a larger portion. Only then will we include a new model in our production data.