Portfolio Sorts and Panel Regressions for Predicting Stock Returns
Summary
The document compares two approaches to empirical stock return prediction: estimating a panel regression with stock-level characteristics and sorting stocks into portfolios. Portfolio sorts are popular because their results are easy to explain and can be translated into factor portfolios for performance tracking and backtesting. Grouping stocks may also reduce noise from stock-specific effects relative to a single-stock predictive regression, while revealing nonlinear patterns at the ends of the characteristic distribution that a rigid linear specification could miss.
The response cautions that panel models need to account for sector and country effects, since characteristics such as valuation can have different meanings across industries and markets. It also notes that with weekly observations, size and valuation ratios may change little over time; ratios such as price-to-earnings and price-to-book can move mainly through their price component. The discussion gives methodological considerations, not a tested comparison or a complete specification, and both approaches can be useful depending on the research question.
Key ideas
- Portfolio sorts are intuitive and make it straightforward to form and track factor portfolios.
- Sorting can reduce the influence of noisy stock-specific effects in return prediction.
- Portfolio buckets can expose nonlinear behavior that a linear regression may obscure.
- Panel regressions should account for sector and country differences.
- Slowly changing characteristics can have limited predictive variation in weekly panel data.
Tags
Full text
# Predicting stock returns - in a panel data specification or by using portfolio formation strategies?
# Predicting stock returns - in a panel data specification or by using portfolio formation strategies?
I'm working on an empirical analysis where I try to predict stock returns using weekly data. Ideally, I would like to use a panel data model like the following: $$ Y_{it}=X_{it}'\beta+\varepsilon_{it} $$
(here presented in a very simple format - it will be more complex in the analysis)
Here $Y_{it}$ is a vector of weekly returns and $X_{it}$ is a vector of explanatory variables with coefficients vector $\beta$.
However, in much of the empirical literature this is not the standard approach. Standard approach involves the sorting of stocks into different portfolios and using portfolio formation strategies.
My questions are:
1) Why is the portfolio approach the standard approach?
2) What are the caveats from using a panel data model?
3) Can explanatory variables typically found in the literature on the prediction of stock returns (e.g. size, P/E ratio, P/B ratio and momentum for each company) also be used in a panel data model?
## Answer by Felix (score 3)
https://quant.stackexchange.com/a/14682
Both approaches can be useful. For stocks, sorting into quantiles is popular because
- it's easy to understand and explain
- it's a simple matter to build factor portfolios and track or backtest their performance, while the translation from expected returns to a portfolio is a bit more involved
- more robust than a single-stock regression, because it is less affected by stock-specific effects which make a predictive regression very noisy
- avoids the rigidity of a linear model, which is a good idea because there are sometimes interesting 'outlier effects' in the first and last quantiles
Regarding your second and third question: there are sector- and country-specific effects (e.g. the valuation of technology vs. finance stocks) which will be important. In a weekly model, the variables size, P/E and P/B will often only vary with 'P', which limits their use as independent variables.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.