Limitations and Possible Extensions
The finding that even a simple optimization program can yield returns significantly
greater than simple buy-and-hold strategies is, and should be, of relevance
to professional portfolio managers. The accounting academic and accounting professional
should be gratified that the incorporation of accounting characteristics
enhances portfolio returns. The decline in performance in the presence of transactions
costs and short-sales constraints diminishes the practical value of the paper.
Limiting optimal strategies to the linear class simplifies the problem, but is
also restrictive. Because the problem is subsequently transformed to a statistical
estimation problem, it may be possible to allow for a larger strategy space. This
certainly would generalize the approach. In addition, it seems possible to allow
36 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
for nonconstant weights, with some kind of adaptive rule to arrive at the optimal
weights. These observations generally pertain to the overall BSCV (2009) procedure
and by extension to the Hand and Green (2011) methodology.
Specific to the Hand and Green (2011) paper, the authors could have considered
incorporating other variables instead of simply adding the well-known
accounting variables that already have been shown to be useful in producing higher
returns. They then could have used the estimation procedure to find the optimal
weights for the different characteristics. In addition, even with the existing variables,
it may have been interesting to see whether splitting accruals into components
such as discretionary and nondiscretionary would have produced better results.
One possible problem in just adding accounting variables to the FFC variables
is that it places these accounting variables on par with market-based variables.
This ignores the distinction that unlike the market variables, accounting
variables may be chosen endogenously. Finally, the authors could have considered
using moment restrictions on the accounting variables using some theoretical
arguments. I would think that the GMM or other statistical estimation
procedures are broad enough to accommodate more general moment settings.