Work detail

The LSTM approach for Value at Risk prediction

Author: Bc. Nikanor Goreglyad
Year: 2021 - summer
Leaders: Marek Hauzr
Work type: Bachelors
Language: English
Pages: 78
Awards and prizes:
Abstract: This thesis describes a new Value at Risk forecasting method based on a neural
network with Long Short-term Memory architecture trained with Joint Supervision loss function (JS LSTM). By optimizing the number of data points on
both sides of the predicted value, JS LSTM produces VaR prediction for a
given confidence level. The JS LSTM is trained to predict one-day-ahead VaR
for PX, WIG20, BUX, and SAX market indexes. The result was compared
with FIGARCH model, EVT-POT model, and LSTM model trained with realized VaR. The performance evaluation shows that the proposed model has
marginally better performance than benchmark models in periods of normal
volatility but underperform in periods of increased volatility.




Patria Finance