Best modelling innovation: CompatibL

CompatibL wins the Best modelling innovation award for the second year running, this time for creating a new type of interest rate model based on autoencoders.

Autoencoders are machine learning algorithms designed to compress data and reveal latent variables. The variational autoencoder (VAE) used by CompatibL is an accurate and stable type used in highly complex applications such as image and speech recognition and generation. CompatibL has used the same VAE methodology to develop an accurate way to compress yield-curve shapes into a small number of model factors. VAEs replace the simple parametric forms of conventional model equations. They are used for the optimal mapping of complex, high-dimensional market data such as yield-curve and volatility surface to low-dimensional latent variables, which become the autoencoder-based model’s factors.

Until now, interest rate models have always been based on stochastic differential equations. These equations use power laws, exponents and other simple mathematical forms to specify the dynamics of interest rates. Attempting to approximate complex market behaviours using such equations is the source of considerable model error. Increasing the number of parameters is the only way to improve conventional models; however, doing so is detrimental to calibration stability and model performance. Recent market turmoil exposed and further amplified the shortcomings of conventional interest rate models. Since these models use a single parametric form for a specific market regime, they are poorly suited to capturing the widely varying set of possible future scenarios and market regimes.

The CompatibL platform offers forward-looking scenarios for risk and portfolio valuation calculations. These scenarios include all possible interest rate regimes covered by the historical time series for all currencies, some with interest rates reaching new heights and others with rates staying at moderate or low levels. The mapping produced by VAE means that model factors are adjusted to the level of the interest rates, producing the best fit across all market regimes. The autoencoder-generated map of market regimes can also be used to identify the possible direction of evolution beginning with the current market regime – a valuable feature in today’s highly uncertain market environment.

The VAEs are trained on a very large and diverse dataset that includes multiple decades of historical time series across many currencies. This allows the autoencoder market models for interest rates to represent, with markedly lower error, the diverse risk scenarios needed in today’s market landscape with its unprecedented levels of risk.
 

Judges said:

• “Nice new algorithm – also used for image and speech recognition.”
• “Genuine modelling innovation with (topical) real-world benefits that promise wide application.”
• “Real innovation – impressed.”

Sponsored content