Today, the performance of pre-trained machine learning models is comparable to some of the fastest numerical methods. Moreover, neural networks easily outperform classical regression techniques such as principal component analysis (PCA) in representing historical interest rate curve shapes and their evolution. They also have the potential to surpass classical curve basis methods such as the Nelson– Siegel (NS) basis and its extension, the Nelson–Siegel–Svensson (NSS) basis.
In this presentation, we focus on the architecture of VAEs and AEMMs, explain how they work, and provide hands-on examples.
- The roles of an encoder and a decoder
- Deliberately introducing uncertainty in a reconstruction
- The loss function and optimization loop
- Reconstruction and generation with a VAE
- Curve representation
- Training on historical data
- One-hot encoding of currency
- VAE with a dimensional latent space
- VAE with a separable two-dimensional latent space
- VAE with a non-separable two-dimensional latent space
- Comparison to the NS and NSS bases
- VAE for handwritten digits from the MNIST dataset
- VAE for the yield curve