Liquid Neural Networks for Quant Models

We have been early but cautious adopters of #AI for #quantitative modeling of #equity and credit markets. Our philosophy was to adopt a #Deep #Learning model only if it offered significant improvement over statistical models. Most DL models were augmentative in nature, i.e. sentiment analysis from structured text, pattern recognition using capsule networks, and risk modeling using GAN. For the core models, we preferred interpretable statistical models.

The main reasons for this cautious approach were

1. Financial markets are adaptive by nature. Hence, constant retraining and subjectivity regarding observation weights become imperative to keep the models relevant.
2. The improvement offered vs. benchmark models is not significant to waste compute resources and coding effort.

In 2024, we started working with #Liquid #Networks after reading the seminal paper by #Ramin #Hasani, and are very near to conversion. The initial results are promising. We used Liquid Neural Networks for factor selection, thinking about the problem as analogous to self driving cars. Thinking of factor maps as guide posts to steer markets, we know from experience that style / factor drifts are slow and persistent for short durations. Hence, we predict factor maps and thus expected return to factor, which can be used for smart index construction, risk model construction or ensemble alpha prediction.

We find that Liquid Networks with considerably less neurons and connections are able to predict better than CNN (the smallest model had about 16000 nodes), and can identify changes in the sub structures of factor maps.

Happy to collaborate with others working in this area.