Deep Learning Architectures to Perform Parameter Estimation & State Inference in Econometric Models
Artificial Intelligence & Machine Learning
Paul Bilokon of the Thalesians and Imperial College London and 2023 Quant of The Year introduces a powerful new concept in quantitative modeling: deep econometrics. This emerging field blends deep learning with classical econometric theory to improve parameter estimation and latent state inference in dynamic, nonlinear systems. (See Paul’s other work: Bayesian Techniques for Estimation and Forecasting in High-Frequency Trading)
The study demonstrates how convolutional neural networks (CNNs), trained on synthetic data, outperform traditional fully connected networks (FCNs) in recovering parameters from nonlinear functions and autoregressive (AR) processes. Notably, the CNNs excelled at generalizing from a single example—something FCNs failed to do even with 60,000 examples.
Bilokon and his collaborators go further by tackling a longstanding challenge in time-series modeling: filtering latent volatility in stochastic volatility models. Traditionally, this task relies on Kalman or particle filters. The researchers introduced a custom deep learning filter capable of approximating these hidden processes directly from observed data. By incorporating both scaled and unscaled data into their architecture, they achieved robust real-time state estimation in nonlinear, non-Gaussian settings.
This work marks a shift in econometric modeling—showing that deep learning architectures, especially CNNs with up to 10 layers, can effectively replace or augment statistical methods in complex inference tasks. The authors emphasize that all code is publicly available, encouraging reproducibility and further exploration.
As Bilokon concludes, “deep econometrics” could redefine the econometric toolkit. Its scalability and flexibility offer a promising path forward for modeling high-dimensional macroeconomic and financial systems, where conventional inference techniques often fall short.
Future research will likely explore deeper architectures, hybrid probabilistic-neural models, and real-world financial applications, signaling a new era in the convergence of AI and economics.
Download the full paper: Deep Econometrics by Paul Bilokon :: SSRN
Join us at: Cornell Financial Engineering Manhattan 2024 Future of Finance & Ai Conference
