Realiseret Volatilitets Forecasting ved HAR-Inspirerede Neurale Netværk
Translated title
Realized Volatility Forecasting Using HAR-Style Neural Networks
Author
Linde, Sebastian Tved
Term
4. semester
Publication year
2022
Submitted on
2022-06-28
Abstract
Volatilitet er central i finans, men kan ikke observeres direkte og må derfor estimeres. Med adgang til højfrekvente data kan realiseret volatilitet – beregnet ud fra intradag-afkast – give en stærk ikke‑parametrisk målestok, særligt for kryptovalutaer, der handles kontinuerligt uden ‘overnight’-effekter. Dette speciale undersøger, hvordan neurale netværk kan forbedre HAR‑inspirerede modeller til at forudsige realiseret volatilitet i fire kryptovalutaer. To dybe læringsmodeller udvikles og sammenlignes med en klassisk HAR‑model: et enkelt feed‑forward netværk og et netværk med dilaterede, kausale konvolutioner inspireret af WaveNet. Modellerne evalueres bl.a. med MSE og QLIKE. Resultaterne viser, at HAR‑modellerne præsterer bedst målt med MSE, mens det simple feed‑forward netværk er bedst på QLIKE. Dette peger på, at HAR‑modeller bedre fanger ekstreme volatilitetstoppe og dag‑til‑dag udsving, mens de neurale netværk bedre modellerer grundniveauet og længerevarende bevægelser i volatiliteten. Fundene kan bruges som vejledning i modelvalg eller kombineres i ensembler for mere balancerede prognoser, og de motiverer yderligere forskning i forskellene mellem de to modeltyper.
Volatility is central to finance but cannot be observed directly and must be estimated. With high‑frequency data, realized volatility—computed from intraday returns—offers a strong non‑parametric measure, especially for cryptocurrencies that trade continuously without overnight effects. This thesis examines how neural networks can enhance HAR‑style models for forecasting realized volatility in four cryptocurrencies. Two deep learning models are developed and compared with a classical HAR model: a simple feed‑forward network and a network using dilated causal convolutions inspired by WaveNet. Models are evaluated using metrics including MSE and QLIKE. The results show that HAR models perform best on MSE, while the simple feed‑forward neural network performs best on QLIKE. This suggests HAR models better capture extreme volatility spikes and day‑to‑day fluctuations, whereas neural networks better model the baseline level and longer‑term movements in volatility. The findings can guide model selection or be combined in ensemble forecasts for more balanced predictions, and they motivate further research into the differing strengths of the two model families.
[This summary has been generated with the help of AI directly from the project (PDF)]
Keywords
Documents