Financial Assets Dependency Prediction Utilizing Spatiotemporal Patterns

TL;DR

This work tackles predicting dynamic asset dependencies by modeling time-evolving Asset Dependency Matrices (ADMs) and forecasting them with a ConvLSTM-based framework. The authors introduce the Asset Dependency Neural Network (ADNN), which integrates two transformation streams—Positional Rearrangement and Mixture of Experts (MoE)—to produce a representation of the ADM optimized for spatiotemporal prediction, followed by a ConvLSTM predictor and a PSD enforcement step. Empirical results show that ADNN variants outperform traditional statistical models and plain deep-learning baselines in ADM prediction, with downstream gains in portfolio risk reduction and enhanced pair-trading performance. The approach provides a scalable, end-to-end pipeline from ADM construction to practical financial tasks, offering a flexible framework for modeling nonlinear, time-varying asset dependencies in real markets.

Abstract

Financial assets exhibit complex dependency structures, which are crucial for investors to create diversified portfolios to mitigate risk in volatile financial markets. To explore the financial asset dependencies dynamics, we propose a novel approach that models the dependencies of assets as an Asset Dependency Matrix (ADM) and treats the ADM sequences as image sequences. This allows us to leverage deep learning-based video prediction methods to capture the spatiotemporal dependencies among assets. However, unlike images where neighboring pixels exhibit explicit spatiotemporal dependencies due to the natural continuity of object movements, assets in ADM do not have a natural order. This poses challenges to organizing the relational assets to reveal better the spatiotemporal dependencies among neighboring assets for ADM forecasting. To tackle the challenges, we propose the Asset Dependency Neural Network (ADNN), which employs the Convolutional Long Short-Term Memory (ConvLSTM) network, a highly successful method for video prediction. ADNN can employ static and dynamic transformation functions to optimize the representations of the ADM. Through extensive experiments, we demonstrate that our proposed framework consistently outperforms the baselines in the ADM prediction and downstream application tasks. This research contributes to understanding and predicting asset dependencies, offering valuable insights for financial market participants.

Figures (8)

• Figure 1: Correlation matrix of multiple assets: (a) assets are randomly posited and (b) assets from the same industry are posited together.
• Figure 2: ADMs Construction. $r_q^{a_j}$ denotes the return of asset $j$ at time $q$, and $\mathcal{M}_t$ denotes the ADM at time $t$. We first generate ADMs sequence from the asset returns sequence. In this figure, $k=10, u=2$ and $q=t-(k-1)u=t-18$. With fixed $u$, $k$, $h$, and shift of current time $t$, we can obtain multiple input sequences and their corresponding targets.
• Figure 3: The overall architecture of the ADNN framework. The input $\mathcal{M}_t^{k,u}$ is processed through the positional rearrangement block and the transformation block using the quadratic transformation learned by the MoE network. The resulting converted matrix is then passed to the spatiotemporal block to generate the forecast matrix, which is further adjusted by the PSD enforcement block to ensure positive semi-definiteness.
• Figure 4: Illustration of ADM positional rearrangement process. Before the positional rearrangement, assets inside the ADM are randomly positioned, resulting in poor spatial patterns. In contrast, the ADM after the positional rearrangement is structured based on the cluster information, thereby enhancing the locality.
• Figure 5: MoE block structure. Assuming there are $n$ experts in the MoE block, the input is fed to the top 2 experts chosen by the gating network.