Forecasting realized volatility through traditional time series and machine learning models

In quantitative finance, time series of assets are not directly forecasted due to their random walk behavior, as they change with high intraday frequency, making them unpredictable. For this reason, but also in the interest of being aware of an asset’s risk exposure, it is more common to forecast the volatility of prices. Among the different measures of volatility, such as historical and implied, I will consider the realized volatility because it is model free and it is based on high frequency intraday past returns. It is then an objective and direct measure of market tendencies and it is also useful to cluster periods of higher and lower volatility. The dataset regards the cryptocurrencies because among all financial assets, cryptos have high price fluctuations, i.e. more significant volatility. After an introduction of the realized volatility in finance, there will be a presentation of two groups of methods that will be compared to predict this measure: the econometric traditional and the machine learning models. The former group considers methods that are specific for financial time series like HAR, HAR-CJ, LevHAR. The latter group includes Neuronal Networks’ machine learning methods like LSTM and TCN. All models were trained on the first 70\% of each time series, validated on the subsequent 10\%, and tested on the final 20\% trough MSE loss function. Forecasting performance of models are compared through Diebold–Mariano test to assess whether differences between models were statistically significant. All of these steps are implemented in python. The main objective of this thesis is to compare the performance of the models across five cryptocurrency datasets.