An MBA dissertation by Michael Paddon
Volatility is an intrinsic property of risk bearing assets and accurate forecasting of this variable is fundamental to effectively managing risk. There exists a rich literature in this domain and many forecasting models are extant. Machine learning techniques have been applied successfully to this problem. In particular, neural nets have been constructed and shown to outperform historical, implied and autoregressive conditionally heteroscedastic volatility models. These neural nets solutions, however, have proven complex in practice to configure optimally. The process has been characterised as “more of an art than science” (Hamid and Iqbal, 2004).
The research hypothesis for this project 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 hyperparameters 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 hyperparameters 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 hyperparameter and input changes. In the highest performing configurations, the model demonstrated an RMS error rate less than 50% of the next best performing benchmark.