Detecting Changes in Asset Co-Movement Using the Autoencoder Reconstruction Ratio

Bryan Lim, Stefan Zohren, Stephen RobertsarXiv 2020 · arXiv

Abstract

Identifying shifts in asset co-movement is critical for risk management, as high correlation often precedes market instability. This paper introduces the Autoencoder Reconstruction Ratio (ARR), a non-linear indicator derived from a Deep Sparse Denoising Autoencoder. By measuring the model's ability to compress and reconstruct market returns, the ARR identifies periods of "tight coupling" that traditional linear methods like Principal Component Analysis (PCA) fail to capture. Empirical tests on high-frequency US equity data demonstrate that the ARR significantly improves short-term volatility forecasting and crash prediction.

Introduction

In stable markets, asset returns are often idiosyncratic, allowing for effective portfolio diversification. However, during systemic stress, these assets frequently move in lockstep—a phenomenon known as co-movement. Traditional metrics like the Absorption Ratio (AR) utilize PCA to quantify this, but they are limited by assumptions of linearity and Gaussian return distributions. Furthermore, PCA-based methods often lack the responsiveness needed for high-frequency market shifts. This study addresses these gaps by replacing the linear framework with a deep learning-based metric: the ARR.

Methodology: The Autoencoder Reconstruction Ratio (ARR)

The researchers utilize a Deep Sparse Denoising Autoencoder, a neural network designed to map input sector returns to a lower-dimensional latent space and then reconstruct the original inputs. To ensure the model learns robust features rather than noise, the authors employ masking noise (dropout) and an L1 penalty for sparsity.

The ARR is defined as the normalized mean squared error of this reconstruction.

  • Low ARR: The model easily reconstructs returns from a few latent variables, signifying high co-movement and market fragility.
  • High ARR: The model struggles to reconstruct returns, meaning assets are moving independently (low co-movement).

Mathematically, the authors demonstrate that the ARR serves as a non-linear generalization of the inverse traditional Absorption Ratio ($ARR \approx 1 - AR$).

Data and Empirical Analysis

The study utilizes second-by-second intraday data from the CRSP US Total Market Index and its 11 constituent sectors from 2012 to 2019. The autoencoder demonstrated a 35% improvement in out-of-sample reconstruction $R^2$ compared to linear PCA, proving that asset relationships are far more complex than simple linear correlations.

Key Results

Volatility Forecasting

The ARR was integrated into various machine learning models (Linear Regression, Gradient Boosted Decision Trees, and Multi-layer Perceptrons) to predict realized volatility. The inclusion of ARR consistently improved predictive performance at 5-minute and 1-hour horizons. For example, in the 5-minute window, the $R^2$ for the MLP model improved from 0.631 to 0.641.

Joint distribution of Log Realized Volatility and ARR across multiple time horizons, showing a clear negative correlation where lower ARR (higher co-movement) aligns with higher volatility.

Crash Prediction

The ARR also proved to be a potent signal for predicting market crashes. Using AUROC (Area Under the Receiver Operating Characteristic) as a metric, models enhanced with ARR showed statistically significant improvements in identifying sharp declines at the 99% confidence level. Historical analysis showed that "dips" in the ARR often preceded major sell-offs, such as the market turbulence in August 2015 and early 2018.

Joint distribution of Drawdown (DD) and ARR across multiple time horizons, illustrating how lower ARR values coincide with higher drawdown risks.

Discussion and Limitations

While the ARR is a powerful tool, its predictive power is strongest for very short horizons (minutes to hours). For longer horizons (1-day to 1-week), the signal's effectiveness decays, suggesting that high-frequency co-movement patterns are transient. Additionally, unlike the mathematically transparent PCA, deep autoencoders require careful hyperparameter tuning, which introduces operational complexity.

Conclusion

The ARR provides a robust, non-linear framework for monitoring market fragility. By capturing complex dependencies that linear models miss, it offers a significant advancement in high-frequency risk monitoring. Future research may explore temporal architectures like LSTMs to further refine these signals across different asset classes like commodities or foreign exchange.

Summary updated

Abstract

Detecting changes in asset co-movements is of much importance to financial practitioners, with numerous risk management benefits arising from the timely detection of breakdowns in historical correlations. In this article, we propose a real-time indicator to detect temporary increases in asset co-movements, the Autoencoder Reconstruction Ratio, which measures how well a basket of asset returns can be modelled using a lower-dimensional set of latent variables. The ARR uses a deep sparse denoising autoencoder to perform the dimensionality reduction on the returns vector, which replaces the PCA approach of the standard Absorption Ratio, and provides a better model for non-Gaussian returns. Through a systemic risk application on forecasting on the CRSP US Total Market Index, we show that lower ARR values coincide with higher volatility and larger drawdowns, indicating that increased asset co-movement does correspond with periods of market weakness. We also demonstrate that short-term (i.e. 5-min and 1-hour) predictors for realised volatility and market crashes can be improved by including additional ARR inputs.
BibTeX
@misc{lim2020detecting,
  author = {Bryan Lim and Stefan Zohren and Stephen Roberts},
  title = {{Detecting Changes in Asset Co-Movement Using the Autoencoder Reconstruction Ratio}},
  year = {2020},
  eprint = {2002.02008},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2002.02008},
}
Figure