Neural Networks and Value at Risk in Asset Management

This study investigates whether machine learning can be utilized to reduce the occurrence of breaching the Value at Risk (VaR) at 5% and 1% thresholds in a portfolio of equity and bond indices. We compare the occurrence and the likelihood of breaches from the Convolutional Neural Network (CNN), the Long Short-Term Memory (LSTM) recurrent neural networks, and the feed forward (FF) using the initiations from the traditional approach (mean/variance or classic) and the Hidden Markov Model (HMM). We perform Monte-Carlo simulations of asset returns to estimate the threshold breaches of VaR using the US, Euro area, UK, and World equity and bond market indices up to 1,343 weeks from January 1987 until mid-June 2020. We also investigate the number and the percentage of breaches under the incentive function that has been trained to take into account the bull and bear markets and the amount of historical data feed (2,000 versus 1,000 days). We find that the LSTM recurrent network initialized with the HMM and balanced incentive function can offer superior protection for asset managers against a downside risk through a reduction in VaR threshold breaches. However, such advancement relies on the availability of long historical data.