This page is for the open-source project on Autoencoder Market Models (AEMM). The code is provided as a Python library using PyTorch.

An autoencoder market model combines an autoencoder with a stochastic process in a latent space. First, the autoencoder is pretrained to map historical data for the yield curve (and, optionally, also the volatilities) to a well-regularized, low-dimensional latent space with minimal reconstruction loss.

Next, a classical stochastic process or a generative machine learning model is inserted between the encoder and the decoder and calibrated to produce the desired probability distribution in the latent space, which the decoder then converts to future interest rates and volatilities.

Encoding interest rates alone produces a deterministic volatility AEMM, while encoding interest rates and their volatilities together produce a stochastic volatility AEMM.

Under the Q-measure, the decoder is rewritten as an equivalent stochastic process so that classical drift adjustments can be used to satisfy no-arbitrage constraints; this last step is not required for P-measure models.

Autoencoder Market Models (AEMM)

Org chart

Teams

Offices