9. Subjective Scoring
Model
Choice of Factors:
We reviewed the
predictive strength of all the factors for the in sample period to determine a
subjective scoring system used to weight each factor. These weights were then used to produce a
trading strategy for the out of sample years.
The factors we used are:
The best
performing factors were LTM EPS Yield and Forward year earnings yield. Both of these factors had good step functions
from the top quartile to the bottom, and performed well in a long-short trading
strategy. These factors are both
indicators of growth, and they are strongly correlated. It makes sense that if a stock has had solid
earnings yields over the past twelve months that analysts would estimate strong
yields for the next year as well.
Diagnostically FY1 Yield was slightly better so that is the factor we
chose.
The next factors
to show promise were those based on analysts revisions. We chose the FY1 up versus down revision ratio,
primarily because the other factor, 3 month revisions, had a very high turnover
which would lead to high transaction costs.
This factor would catch those stocks with improving expectations, that
don’t necessarily pay dividends and don’t (yet) have good earnings yields.
We chose
Dividend Yield as our third factor because it performs decently in predicting
the top portfolio and it is very uncorrelated with the
other two factors. Dividend yield could
capture more of the value stocks that wouldn’t be reflected in the other two.
Scoring:
Each portfolio
within each factor was scored on a scale of one to five based on our assessment
of its ability to predict returns, both good and bad, in sample. On the extremes, those with the best positive
predictions received a score of five, and those with the best negative
predictions would receive a score of -5.
If a factor showed no predictability for returns in a certain portfolio
it would not be scored. We distributed
the following scores:
Analysis
Equal
Weighted: In sample, implementing
a trading strategy based on the subjective scores did very well, with an
annualized average return of 33.16% for the top portfolio and 9.35% for the
bottom portfolio. The long-short return
would be 23.8%, and very well hedged with betas of 1.03 and .99 for the top and
bottom portfolios respectively. One
concern would be the percentage turnover for the portfolios, particularly the
bottom one, at 20.96% and 39.00% respectively.
Such turnovers would result in high transaction costs that could eat
into the returns.
Out of
sample this model performed well 4 out of the 5 years, with 1999 being the
negative year. By implementing this
strategy, we would have had a loss of 25% in 1999, but average gains of more
than 50% in each of the other years.
Value
Weighted: In sample, the value weighted portfolios performed similarly
to the equal weighted ones. The returns
are slightly lower, at 28.29% for portfolio 1 and 10.36% for portfolio 5. However, the standard deviations are also
lower than for equal weighted. The betas
are not as evenly matched, but still give a good hedge with a beta of 1.09 at
the top and .93 at the bottom.
Out of
sample, this strategy performs extremely well.
Similar to the equal weighted, in 1999 a long-short strategy based on
this model would lose money, but only 8%.
For the remaining years we would achieve an average of 36% returns
annually, with no disastrous years.
Conclusion:
Using a
combination of factors we developed a good model for a long-short trading strategy. The factors capture elements of growth and
value stocks and seem to do well in various market conditions, with decent
upside and limited downside. We would
probably implement this model using a value weighted portfolio because it seems
to hedge the downside better. The one
major concern with this model is the potential transactions costs due to
turnover.