💰 quantitative finance

Vector-Quantized Discrete Latent Factors Meet Financial Priors: Dynamic Cross-Sectional Stock Ranking Prediction for Portfolio Construction

The paper introduces PRISM-VQ, a dynamic factor framework that integrates expert priors with vector-quantized discrete latent factors and a structure-conditioned Mixture-of-Experts to enhance cross-sectional stock return prediction and portfolio performance while maintaining interpretability.

Original authors: Namhyoung Kim, Jae Wook Song

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Namhyoung Kim, Jae Wook Song

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict which stocks in a massive market will go up or down tomorrow. It's like trying to hear a single whisper in a hurricane. The signal (the real reason a stock moves) is very weak, and the noise (random market chatter) is overwhelming.

The paper introduces a new AI system called PRISM-VQ to solve this problem. Think of it as a high-tech "prism" that takes the chaotic, messy light of the stock market and splits it into a clean, organized rainbow of patterns.

Here is how it works, broken down into simple steps:

1. The Problem: The "Noisy Room"

Traditional methods are like trying to guess the weather by looking at a single thermometer; they are too rigid. Newer AI methods are like having a super-fast computer that guesses based on patterns, but they often get confused by the noise and forget the basic rules of economics (like "low prices usually mean value").

2. The Solution: The "PRISM" Approach

The authors built a two-stage system that acts like a filter and a smart team of experts.

Stage 1: The "Discrete Filter" (Vector Quantization)

Imagine you have a giant bucket of mixed-up Lego bricks (stock data). Some are red, some are blue, some are broken, and some are just dust.

What PRISM does: It forces these bricks into specific, pre-defined "buckets" (called a codebook). Instead of saying a brick is "50% red and 51% blue," it forces a decision: "This brick is definitely Red Bucket #42."
Why this helps: By snapping the data into these fixed buckets, the system ignores the tiny, random dust (noise) and keeps only the strong, repeating patterns. It's like sorting a messy pile of mail into 500 specific mailboxes; once sorted, you can see the patterns clearly.
The "Contrastive" Trick: The system also learns to make sure that bricks that should be together are in the same bucket, and bricks that are different are pushed far apart. This creates a clean map of the market's structure.

Stage 2: The "Smart Team" (Mixture-of-Experts)

Now that the data is sorted into clean buckets, the system needs to predict the future.

The Setup: Imagine a team of 8 different financial experts (like a Value Expert, a Growth Expert, a Momentum Expert).
The Switch: In the past, one AI model tried to do everything. In PRISM, the "bucket" the stock was sorted into in Stage 1 acts as a switch.
- If a stock falls into "Bucket #42" (which represents a specific market pattern), the system automatically routes that stock to Expert #6.
- If it falls into "Bucket #10," it goes to Expert #2.
The Result: Each expert specializes in a specific type of market condition. They don't all try to guess at once; the right expert is called upon for the right situation. This makes the prediction much more accurate and adaptable.

3. The "Safety Net" (Financial Priors)

The system doesn't just learn from scratch; it respects the "old rules" of finance.

Imagine the AI is a student. The "Financial Priors" are the textbooks the student must read.
The system is given established economic factors (like "Value," "Size," and "Momentum") as a starting point. It uses these as anchors to ensure its wild guesses don't drift too far from reality. This keeps the model stable even when the market gets crazy.

4. The Results: A Better Portfolio

The authors tested this system on two major markets: the CSI 300 (China) and the S&P 500 (USA).

The Score: Compared to other top AI models and traditional methods, PRISM-VQ consistently predicted stock rankings better.
The Money: When they built a portfolio (a basket of stocks) based on these predictions, they made more money with less risk (a higher "Sharpe Ratio") and fewer huge losses (lower "drawdown").
The "Why": The system worked because it successfully filtered out the noise (Stage 1), used the right expert for the right situation (Stage 2), and didn't forget the basic rules of economics (Priors).

Summary Analogy

Think of the stock market as a crowded, noisy dance floor.

Old AI tries to predict the next move by listening to everyone shouting at once.
PRISM-VQ first puts everyone into specific, quiet rooms based on how they dance (Stage 1: Vector Quantization). Then, it sends a specialized dance coach to each room to teach the next move (Stage 2: Mixture-of-Experts). Finally, it reminds the coaches of the basic steps of the dance (Financial Priors).

The result is a much clearer, more accurate prediction of where the dancers (stocks) will move next.

Technical Summary: PRISM-VQ

Problem Statement

Predicting cross-sectional stock returns is a fundamental challenge in quantitative finance due to three primary obstacles: persistently low signal-to-noise ratios (SNRs), complex and evolving cross-asset dependencies, and frequent market regime shifts. Classical linear factor models offer interpretability but lack the flexibility to capture nonlinear interactions and time-varying dynamics. Conversely, deep learning approaches, while powerful, often suffer from insufficient regularization in noisy environments, fail to condition on evolving cross-sectional structures, and underutilize established financial priors. Existing methods typically struggle to balance the expressiveness of neural networks with the interpretability and robustness required for portfolio construction.

Methodology: PRISM-VQ

The authors propose PRISM-VQ (PRior-Informed Stock Model with Vector Quantization), a unified framework designed to address these limitations through a two-stage learning paradigm that integrates discrete representation learning, structure-conditioned temporal modeling, and domain-informed priors.

1. Core Architecture

PRISM-VQ decomposes the problem into Spatial Learning (Stage 1) and Temporal Learning (Stage 2):

Stage 1: Spatial Learning (Discrete Factor Discovery)
- Objective: To discover a reusable, discrete cross-sectional structure that acts as an information bottleneck, suppressing noise while preserving market structure.
- Mechanism: Temporal features for each stock are encoded via a GRU and then processed by a cross-asset Transformer to aggregate cross-sectional information. These continuous embeddings are discretized using Vector Quantization (VQ) against a learnable codebook.
- Regularization: To ensure the discrete codes are semantically meaningful and not arbitrary partitions, the model employs:
  - Contrastive Learning: Enlarges the margin between assigned codewords and competing prototypes.
  - Reconstruction: A FiLM-conditioned decoder reconstructs input features from the quantized codes.
  - Multi-horizon Prediction: An auxiliary task predicts future returns to align codes with forward-looking signals.
- Output: Each stock is assigned a discrete code $z_{q,i}$ , which serves as a latent factor value and a structural signal.
Stage 2: Temporal Learning (Dynamic Factor Loadings)
- Objective: To generate time-varying factor loadings conditioned on the learned cross-sectional structure.
- Mechanism: A Structure-Conditioned Mixture-of-Experts (MoE) built on Transformer architectures.
- Routing: The discrete code $z_{q,i}$ from Stage 1 acts as a gating signal. It determines which experts (sub-networks) are activated for a specific stock at a specific time.
- Factor Integration: The model generates dynamic loadings for both expert prior factors (e.g., value, momentum from the JKP Global Factor Library) and the learned latent factors. The final prediction is a linear combination: $y_i = \alpha_i + \beta_{p,i}^T f_p + \beta_{l,i}^T f_{l,i}$ .
- Training: The codebook is frozen after Stage 1. Stage 2 optimizes the temporal Transformer and MoE parameters end-to-end, using a load-balancing regularizer to prevent expert collapse.

2. Key Design Choices

Vector Quantization as Inductive Bias: By mapping continuous embeddings to a finite set of prototypes, the model suppresses idiosyncratic noise and encourages the reuse of stable cross-sectional patterns.
Discrete-Code-Gated MoE: This mechanism allows the model to specialize different experts for distinct market regimes identified by the discrete codes, enabling adaptive factor loadings without sacrificing interpretability.
Two-Stage Decoupling: Separating structure discovery from temporal modeling prevents codebook collapse and enforces a clear information bottleneck, ensuring the temporal model conditions on a stable representation of market structure.

Key Contributions

Unified Framework: PRISM-VQ is the first framework to integrate vector-quantized discrete factors, contrastive representation learning, expert prior factors, and discrete-code-conditioned temporal modeling within a single factor-based model.
Effective Inductive Bias: The paper demonstrates that vector quantization provides a robust inductive bias for financial representation learning, consistently outperforming continuous latent alternatives in low-SNR environments.
Structure-Conditioned Dynamics: The introduction of a discrete-code-gated MoE enables dynamic factor loading generation that adapts to evolving market structures while maintaining a clear factor interpretation.
Empirical Validation: Extensive experiments on the CSI 300 and S&P 500 markets show consistent improvements over strong baselines.

Experimental Results

The model was evaluated on the CSI 300 (China) and S&P 500 (US) markets using data from 2022–2024 (strictly out-of-sample).

Performance Metrics:
- CSI 300: Achieved a RankIC of 0.0646 and a Sharpe Ratio of 1.57, outperforming all baselines including Transformer-based models (e.g., DTML, MASTER) and autoencoder variants.
- S&P 500: Achieved a RankIC of 0.0141 and a Sharpe Ratio of 0.67, representing a 58.4% improvement in RankIC over the best baseline (DTML).
Ablation Studies:
- Removing the discrete codebook caused the most significant performance degradation (RankIC dropped by ~27% on CSI 300 and became negative on S&P 500), highlighting the necessity of the discrete bottleneck.
- Removing expert prior factors reduced performance significantly, particularly in the efficient S&P 500 market, confirming the regularizing role of domain knowledge.
- Replacing the MoE with a single network lowered performance, validating the need for structure-conditioned specialization.
Interpretability:
- t-SNE Visualization: Learned codes formed clusters that did not strictly align with sector taxonomies or market capitalization, suggesting the capture of richer cross-sectional structures.
- Factor Exposure: Codes exhibited systematic but moderate correlations with standard financial factors (e.g., profitability, momentum), indicating they capture multi-dimensional factor combinations rather than replicating single factors.
- Expert Specialization: Different experts in the MoE activated under distinct market regimes, with statistically distinct activation patterns and low temporal overlap.

Significance and Claims

The authors claim that PRISM-VQ addresses the critical gap between the interpretability of classical factor models and the predictive power of deep learning. By treating vector quantization as an information bottleneck and using discrete codes to gate expert networks, the model successfully navigates the low-SNR and non-stationary nature of financial data.

The paper posits that combining a discrete information bottleneck with prior-informed conditional computation is an effective design principle for financial return prediction. The results suggest that this approach not only improves ranking accuracy and portfolio performance (higher Sharpe ratios, lower drawdowns) but also preserves the interpretability required for financial analysis, offering a robust solution for portfolio construction in evolving market regimes. The authors emphasize that their method does not rely on predefined relational topologies (like graphs) but learns implicit cross-sectional structure, enhancing transferability across different markets.