Proceedings of The 13th International Conference on Modern Research in Management, Economics and Accounting
Analyzing the Dynamics of the Swaption Market Using Neural Networks
Sándor Kunsági-Máté, Gábor Fáth and István Csabai
The SABR stochastic volatility model is a widely used option pricing tool proposed by Hagan et al. (2002). It was the first that could successfully capture the static pattern of the volatility smile while predicting its dynamics much better than previously used local volatility models. This model is often used for hedging where the portfolio is usually hedged to the delta and vega risks representing the uncertainty in the underlying asset price and its volatility. The model assumes that the two non-stochastic parameters – ρ and υ – are stable over time. However, traders regularly recalibrate these model parameters to fit the market, making them stochastic as well and introducing additional parameter risks. Then the question arises whether this behaviour is due to the larger complexity of the market or due to the incorrect model choice? In our study we analyzed the dynamics of the volatility smiles of the GBP swaption market using an autoencoder-like neural network, and created an alternative model, a deformation of SABR, in which we can describe the smiles with two stochastic parameters only as it would be required from a real two factor model. This new model can reproduce well the volatility smiles for several months until it reaches a critical point- the onset of the COVID crisis – where the reproduction error increases suddenly. We found that the sharp discrepancy is caused by a sudden change in the representative low dimensional market manifold, and our model is highly sensitive to pick up this change, making it especially useful for regime change detection.
keywords: Deep Learning; regime change detection; stochastic models; time series; volatility smiles.