Stock Return Predictability and Model Uncertainty: A Frequentist Model Averaging Approach
|Author:||Mgr. Vojtěch Pacák|
|Year:||2019 - summer|
|Leaders:|| doc. PhDr. Tomáš Havránek Ph.D.
|Work type:|| Finance, Financial Markets and Banking
|Awards and prizes:|
|Abstract:||The model uncertainty is a phenomenon where general consensus about the form of specific
model is unclear. Stock returns perfectly meet this condition, as extensive literature offers
diverse methods and potential drivers without a clear winner among them. Relatively
recently, averaging techniques emerged as a possible solution to such scenarios. The two
major averaging branches, Bayesian (BMA) and Frequentist (FMA) averaging, naturally deal
with uncertainty by averaging over all model candidates rather than choosing the "best" one
of them. We focus on FMA and apply this method to our data from U.S. market about S&P
500 index, that I help to explain with the set of eleven explanatory variables chosen in
accordance with related literature. To preserve a real-world applicability, I use rolling
window scheme to regularly update data in the fitting model for quarterly based reestimation. Consequently, predictions are obtained with the use of most recent data.
Firstly, we find out that simple historical average model can be beaten with a standard
model selection approach based on AIC value, with variables as Dividend Yield, Earnings
ratio, and Book-to-Market value proving consistently as most significant across quarterly
models. With FMA techniques, I was not able to consistently beat the benchmark from
model selection, as results are rather disappointing. I discuss impact of window size choice in
the rolling scheme and eventually re-run analysis with single window shift and random split.
Observed results are much more promising, as all FMA methods are outperforming