Neural network models for conditional quantiles of financial returns and volatility

Thesis title in Czech: | Modely neuronových sítí pro podmíněné kvantily finančních výnosů a volatility |
---|---|

Thesis title in English: | Neural network models for conditional quantiles of financial returns and volatility |

Key words: | podmíněnéněné kvantily, kvantilová regrese neuronových sítí, neparametrické odhady realizované volatility, VaR |

English key words: | conditional quantiles, quantile regression neural networks, realized measures of volatility, value-at-risk |

Academic year of topic announcement: | 2014/2015 |

Type of assignment: | diploma thesis |

Thesis language: | angličtina |

Department: | Institute of Economic Studies (23-IES) |

Supervisor: | doc. PhDr. Jozef Baruník, Ph.D. |

Author: | hidden - assigned by the advisor |

Date of registration: | 15.06.2015 |

Date of assignment: | 15.06.2015 |

Date and time of defence: | 15.09.2016 00:00 |

Venue of defence: | IES |

Date of electronic submission: | 29.07.2016 |

Date of proceeded defence: | 15.09.2016 |

Reviewers: | prof. Ing. Miloslav Vošvrda, CSc. |

URKUND check: |

References |

Taylor, J. W., 2000. A quantile regression neural network approach to estimating the conditional density of multiperiod returns. Journal of Forecasting. 19(4): 299-311.
Žikeš, F., Baruník, J., 2014. Semi-parametric Conditional Quantile Models for Financial Returns and Realized Volatility. Journal of Financial Econometrics. Clements, M. P., Galväo, A.B., Kim, J. H., 2008. Quantile forecasts of daily exchange rate returns from forecasts of realized volatility. Journal of Empirical Finance. 15(4): 729-750. Berkowitz, J., Christoﬀersen, P.,Pelletier, D., 2011. Evaluating value-at-risk models with desk-level data, Management Science. 52(12): 2213–2227. Hansen, B.E., 1994. Autoregressive Conditional Density Estimation, International Economic Review. 35: 705-730. Donaldson, R.G., Kamstra, M., 1996. Forecast combining with neural networks, Journal of Forecasting. 15(1): 49-61. Atiya, A. F., 2001. Bankruptcy prediction for credit risk using neural networks: A survey and new results. IEEE Transactions on Neural Networks. 12 (4): 929-935. Yeh, I.-C., 2014. Estimating distribution of concrete strength using quantile regression neural networks. Applied Mechanics and Materials 584-586: 1017-1025. Xiong, T., Bao, Y., Hu, Z., 2013. Beyond one-step-ahead forecasting: Evaluation of alternative multi-step-ahead forecasting models for crude oil prices. Energy Economics. 40: 405–415. Panella, M., Barcellona, F., D’Ecclesia, R. L., 2012. Forecasting energy commodity prices using neural networks. Advances in Decision Sciences 2012 Jammazi, R., Aloui, C., 2012. Crude oil price forecasting: Experimental evidence from wavelet decomposition and neural network modeling. Energy Economics 34 (3): 828–841. |

Preliminary scope of work |

Tato teze přispěje do aktuální akademické literatury srovnáním lineárních odhadů kvantilů s potenciálně lepším nelineárním přístupem. QRNN by mělo dát lepší forecastované odhady kvantilů než lineární kvantilová regrese.
Praktické výsledky jsou široké. Například ocenění opcí (Hansen, 1994). Za předpokladu odhadnuté distribuce výnosů a volatility, spočítání očekávaných výnosů a očekávané volatility instrumentů, v této práci se bude jednat o S&P 500 and WTI Crude Oil futures. Dále mohou být odhadnuty výnosy i volatilita v portfoliu instrumentů. Dalším přínosem je použití kvantilové regrese a její schopnost sloužit jako nástroj pro investory, kde mohou predikovat riziko ztrát (Žikeš and Baruník, 2014). |

Preliminary scope of work in English |

Motivation:
Forecasting financial volatility and financial returns is important for making rational decisions. Having the whole distribution of future changes in prices and volatility would be ideal, but having few quantiles may be sufficient. For example modeling 5% VaR, which is a threshold loss value, where there is 5% probability of a loss of given amount over some time period. Linear models estimation of quantiles may be improved by using neural networks. The reason for that is that neural networks are used for estimation of non-linear models where the linear are a subset of more general non-linear models and they still can be estimated by using the neural networks. In forecasting the use of neural networks provides generally better forecasts than linear models, for example Donaldson and Kamstra (1996) found that the use of artificial neural networks on the out of sample basis produces superior forecasts in comparison to linear procedures. Atiya (2001) found that neural networks in addition to classic indicators improve the prediction accuracy of corporate bankruptcies. QRNNs were used for example in civil engineering for estimating the distribution of concrete strength and Yeh (2014) found that QRNN can in this case build accurate quantile models. Xiong et al. (2013) shows that model based on neural networks is better in predicting accuracy in comparison to three commonly used prediction strategies in the case of WTI crude oil. Panella et al. (2012) found that their model based on neural networks may provide improvements in comparison to other well known models. Another paper that used neural networks successfully to predict oil prices is Jammazi and Aloui (2012). Hypotheses: 1. Hypothesis #1: Quantile regression neural networks provides better quantile estimates than standard linear quantile regression. 2. Hypothesis #2: Volatility is important for predicting quantiles of returns. 3. Hypothesis #3: Expected densities of returns are skewed. Methodology: To model non-linearity in quantiles, the quantile regression neural network approach (Taylor, 2000) will be used. The underlying instruments for comparing linear and non-linear models are S&P 500 and WTI Crude Oil futures and they are based on Žikeš and Baruník (2014). To compare relative performance Žikeš and Baruník (2014) suggest to follow Clements et al. (2008) and for absolute performance evaluation to use CAViaR test (Berkowitz, Christoﬀersen & Pelletier 2011). Comparison of model including volatility and returns with a model including only returns will be used for determining the importance of volatility in predicting quantiles of returns. Same ways of relative and absolute performance as mentioned above will be used. Distribution of expected returns may not be Gaussian, for example if the previous return is negative and too big then in some cases we can expect a correction in the market, we can expect that the return will fall, but with a limit to the fall which can result in skewed expected density of return. This will be tested for example by the comparison of mean and median. Expected Contribution: The thesis will contribute to the current academic literature by comparing the linear estimation of quantiles with potentially better approach of non-linear estimation. It should show that the usage of quantile regression neural network approach provides better forecasts for quantiles of return than the simple quantile regression. Practical results are wide. For example option pricing (Hansen, 1994). Under the condition of having an estimated distribution of return and volatility we can calculate the expected return and expected volatility of assets, futures contracts and so on and in the case of this thesis S&P 500 and WTI Crude Oil futures. It can be also used to estimate the expected return and volatility in portfolio investments. Another contribution of applying quantile regression is it's ability to serve as a risk management tool for investors, since they can predict the risk of their losses (Žikeš and Baruník, 2014). Outline: 1. Introduction: Introduction to estimating quantiles of volatility and returns. How different authors use different approaches. 2. Methodology: Description of quantile regression and how it works with neural network. 3. Data description: Description of the dataset that is used. 4. Results: Discussion of results and comparison to different methods. 5. Summary |