Handling Skewed Fundamental Survey Scores in Factor Models

Summary

The document describes a practical problem in combining analysts’ ordinal ratings with quantitative stock factors. The ratings run from very good to very bad, but the watchlist excludes the worst stocks, so the observed scores cluster near the favorable end. The author also notes that the rare poor ratings may be influenced by unusual cases, and that ordinary least squares factor-premium estimates have low explanatory power.

It raises the question of which statistical technique can accommodate this skew, but gives no proposed estimator, diagnostic, or empirical comparison. The stated low R-squared is a result for the author’s current analysis, not evidence that a particular alternative would improve the model. The selection of stocks into the watchlist also means the score distribution is truncated, which may matter alongside skew when interpreting ratings. The document is useful as a problem statement for factor research, but it does not provide a method or establish whether the ratings contain predictive information.

Key ideas

• Analyst ratings are ordinal scores that cluster toward the favorable end of their scale.
• The watchlist excludes the most negatively rated stocks, truncating the observed score distribution.
• Rare poor ratings may be unusual cases, making their influence on estimates worth examining.
• The reported OLS factor-premium analysis has low R-squared, but no alternative method is tested.

Tags

# Factor models based on fundamental surveys: how to deal with the pointy end?


# Factor models based on fundamental surveys: how to deal with the pointy end?












I'm a quant working in a mainly fundamental shop. Analysts are asked to score things like management or industry trends of stocks in their "watchlist", and I am now trying to weave the results into a factor model. The model includes other factors such as earnings yield and some financial quality metrics.

The scores are simple 1 to 5 scores, where 1 is very good and 5 is very bad. A lot of thought is put into them, and they derive from team consensus (and hours of arguments).

The problem is that they are heavily skewed towards the better end of the spectrum. Stocks which would score a 5 would be considered so bad we wouldn't bother watching them (and thus are not even in our watchlist). Stocks which score a 4 are rare: only a couple of stocks, one of which has had substantial volatility.

Which statistical technique best deals with the skew in score distributions?

When I compute factor premiums with OLS, I get a very low R-squared.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.