Forecasting Stock Price Volatility (2016)
Forecasting Stock Price Volatility Dissertation – This dissertation addresses the question of whether there might exist a more accurate method of forecasting volatility that those in common use today. More specifically the project restricts its scope to stock prices since historical data, including pricing, for these assets are transparent, readily available and voluminous. Before moving on to the research hypothesis, however, a final general observation about volatility is, perhaps, in order.
Volatility is a useful proxy for risk and it is commonly treated as such in the financial literature and mathematical models. However there is a countervailing viewpoint that volatility is not exactly the same thing as risk. A stock that rises exhibits high volatility, but if that rise is based on business fundamentals then the underlying risk may not have moved. The research hypothesis for this dissertation is that recent advances in machine learning and deep neural nets allow the construction of a high performance volatility forecasting model that is easier to configure correctly and more stable against changes in model hyper-parameters or inputs. An LSTM based recurrent neural network architecture is proposed as suitable for time series forecasting.
Data from the CRSP US Stock Database and Google Trends is used to construct training patterns and testing patterns and the model is subjected to fifteen different training scenarios. These scenarios vary the model hyper-parameters and inputs, and compare the consequent performance against three common benchmark forecasting models. The model is found to perform well in most scenarios, be easy to configure and demonstrates good resilience to hyper-parameter and input changes. In the highest performing configurations, the model demonstrated an RMS error rate less than 50% of the next best performing benchmark.
- 16,000 words – 68 pages in length
- Excellent use of literature
- Excellent analysis of subject area
- Well written throughout
- Ideal for finance students
1 – Introduction
Hypothesis
2 – Research Goal
3 – Theory
Stochastic Volatility
Volatility Forecasting / Forecasting Stock Price Volatility
Artificial Neural Networks
Recurrent Neural Networks
Long Short-Term Memory
Tensors
4 – Data Sources
5 – Model Design
Basic Architecture
Input / Output Transformation
Training
Input Features
CRSP Features
Google Trends Features
Input Tensor
Output Tensor
Implementation Outline
Model Visualisation
6 – Experimental Results
Analysis of Scenarios
7 – Analysis
8 – Conclusions and Future Directions
References
