Modeling Stock Return Volatility, a Comparative Approach
Open Access
- Author:
- Krimetz, Robert
- Area of Honors:
- Data Sciences
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- Matthew D Beckman, Thesis Supervisor
Anh Tuan Le, Thesis Honors Advisor
Jeremy Seeman, Faculty Reader - Keywords:
- Stock Volatility
time series models
Bayesian Models
Probabilistic Programming
Markov chain Monte Carlo
Hamiltonian Monte Carlo
Implied Volatility
Realized Volatility - Abstract:
- The application of machine learning and probabilistic programming methods on stock return prediction has grown in tandem with the availability of high frequency stock data. With well recorded heteroskedasticity in historical stock returns, modeling attempts have evolved from making general assumptions about the underlying data generating distribution to predicting changes in the underlying distribution of returns. The increase in popularity of ‘tradable volatility’ through derivative contacts and VIX futures over the past three decades has motivated research efforts to model the variance of daily returns. Along this line of research, three schools of thought have emerged to model return volatility; Time Series Models, Stochastic Models, and Bayesian Models. Given that the preliminary assumptions underlying these models differ, the nature of their results and the varying metrics used to calculate their respective accuracy makes it difficult to directly compare them. Accordingly, the currently available pool of research has diverged along these three separate paths making it unclear the advantages of each. Notably, Bayesian models have largely been neglected in the current pool of research due to their computational intensity. In this paper I derive ten time series and Bayesian models then provide a comprehensive comparative study of the results on real stock data. I found that Bayesian models with intractable posterior distributions significantly outperform time series models at predicting directional change in future volatility, while the GARCH and FIGARCH time series models generate the most accurate point predictions for future volatility. I hope the results outlined in this paper better contextualize different volatility predictions and motivate the creation of more accurate tradeable volatility models.