Reasonable intervals for financial data

[Hubert C]

Hi people,

I am doing a Fama-MacBeth style regression in order to judge whether one type of portfolio has outperformed another over time.

The problem is that the dataset I am working with contains a lot of errors (I use Datastream). For example, some companies turn out market-to-book ratios into the millions, and betas of values from 100 into the 1000s, both positive and negative. Seeing as there is no form of value-weighting in this type of regression, it really wrecks havoc with my results, even though the instances themselves are very few. I have already attempted to control for it by setting a cutoff in terms of market value (the penny stocks are the worst), but I realize I have to screen the data for wild outliers.

My question to the people of this good forum is as follows: What are your thoughts on reasonable intervals for a stock beta, market-to-book value, dividend yield (One company had a "yield" of over 1000,000.00 %), momentum and turnover/shares outstanding (one company evidently had 10,000,000.00 % of its market cap traded in a given month). The ideal thing would be if somebody knew of any research on the field (maybe 99% confidence intervals or something), but I'd be happy just for some reasonable suggestions.

Thanks for taking the time to help me out!

-Hubert

 

[Hubert C]

Alternatively, if anyone know of a good paper/book detailing winsorizing/trimming data that would be interesting too. It seems fairly common practice to do this, but it would be nice to have a big name researcher/professor to lend some legitimacy to the method for the report I am doing, hehe.

 

[joel_b]

This paper revisits FM regression, points out the negatives of the model and proposes a few things.

http://webuser.bus.umich.edu/ppasquar/shortpaper4.pdf

In this paper the datasets are broken into 9 time periods, then the securities are ranked by beta, then 20 portfolios are created making ranges for the beta ranks, to give better precision on beta estimates. I think this is always to be done in a FM regression, but I don't know enough about it.

You will also notice that in table 1, not every available security meets what Pasquariello calls 'data requirements', however I don't think he elaborates on the conditions in the short paper, which may be of interest to you for screening. That being said I'm sure you could get your hands on the Long Paper from NYU, or contact him to get an opinion.

Pasquariello: Research Work

 

[Hubert C]

Thank you, joelb! That was just the kind of article I was searching for.

-Hubert