Paper: arXiv 2306.02848

Authors: Zikai Wei, Anyi Rao, Bo Dai, Dahua Lin

Abstract

Factor model is a fundamental investment tool in quantitative investment, which can be empowered by deep learning to become more flexible and efficient in practical complicated investing situations. However, it is still an open question to build a factor model that can conduct stock prediction in an online and adaptive setting, where the model can adapt itself to match the current market regime identified based on only point-in-time market information. To tackle this problem, we propose the first deep learning based online and adaptive factor model, HireVAE, at the core of which is a hierarchical latent space that embeds the underlying relationship between the market situation and stock-wise latent factors, so that HireVAE can effectively estimate useful latent factors given only historical market information and subsequently predict accurate stock returns. Across four commonly used real stock market benchmarks, the proposed HireVAE demonstrate superior performance in terms of active returns over previous methods, verifying the potential of such online and adaptive factor model.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces advanced deep learning architectures (VAEs, hierarchical latent spaces, regime-switching) with complex stabilization algorithms and online learning frameworks, indicating high mathematical density. While it includes real-world benchmarks and backtests over four stock markets, the summary focuses more on model design and theoretical validation than on detailed data preprocessing, hyperparameter tuning, or full implementation details typically found in production-ready strategies.

Research Flowchart

  flowchart TD
  A["Research Goal: Develop an online & adaptive factor model for stock prediction?"] --> B{"Data Inputs<br>Historical Point-in-Time<br>Market & Stock Data"}
  B --> C["Methodology: HireVAE Architecture"]
  C --> D["Hierarchical Latent Space<br>(Market to Factor VAEs)"]
  C --> E["Regime-Switching Mechanism<br>(Adaptive Online Learning)"]
  D & E --> F["Computational Process: Continuous<br>Market Regime Identification &<br>Latent Factor Estimation"]
  F --> G["Output: Accurate<br>Stock Return Predictions"]
  G --> H["Key Findings: Superior<br>Active Returns vs. Benchmarks"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes HireVAE, the first deep learning based online and adaptive factor model with hierarchical latent space and regime-switching capability.

Source sentence
we propose the first deep learning based online and adaptive factor model, HireVAE, at the core of which is a hierarchical latent space that embeds the underlying relationship between the market situation and stock-wise latent factors

Assumes 6

  • The market score s is generated from a Gaussian market regime distribution. statistical
    Source sentence
    We assume that the market score s is a random variable generated from a market regime distribution defined as a Gaussian distribution.
  • All training data must be point‑in‑time with no future information leakage. data
    Source sentence
    Practice requires an online learning setting that all data used for learning should be point-in-time data, meaning that there is no future information leakage at any timestamp in the training data set.
  • The dataset consists of chronologically ordered samples D = {Dt0,…,Dtn}. data
    Source sentence
    In our setting, we have a set of samples labeled with time stamps sorted in chronological order: D = {Dt0, . . . , Dtn}
  • The market latent variable m0 follows a Gaussian distribution with parameters µm0 and σm0. model
    Source sentence
    where m0 ∈RHm follows a Gaussian distribution with parameters µm0 and σm0
Show 2 more
  • The stock latent variable z0 follows a Gaussian distribution with parameters µs0 and σs0. model
    Source sentence
    z0 ∈RHs follows a Gaussian distribution with parameters µs0 and σs0
  • The hierarchical encoder‑decoder structure can capture market‑stock hierarchy and regime‑specific decoding. model
    Source sentence
    we develop a hierarchical encoder with a two-level latent space representing a market‑stock hierarchy, and couple it with a set of regime‑specific decoders that reconstruct or predict stock returns based on different market situations

Data

  • China stock market · WIND database market-prices commercial
    2010 to mid‑2022 · all on‑listing Chinese stocks excluding ST and those listed <3 months
    Source sentence
    We conduct the experiment over the entire China stock market, with price‑volume and fundamental data obtained from the WIND database, and all raw data processed on a point‑in‑time basis.

Recognised sources: Wind

Costs

Modeled Costs are part of the reported headline results.

Finds 4

  • HireVAE outperforms classical linear methods and recent non‑linear data‑driven methods by a large margin on real stock market data. performance
    Source sentence
    Being the first online and adaptive regime‑switch factor model, HireVAE achieves superior performance on real stock market data, outperforming classical linear methods and recent non‑linear data‑driven methods by a large margin.
  • The adaptive regime‑switching algorithm consistently outperforms rule‑based and neural clustering baselines in stock prediction and portfolio construction. performance
    Source sentence
    our adaptive regime switching algorithm can always outperform the rule‑based and neural clustering counterparts.
  • Adding market information does not significantly improve model performance. negative
    Source sentence
    The results show that the additional market information does not help to improve the performance of a model.
  • A hierarchical latent space does not provide significant improvement over a single latent space when regime‑switching is not used. negative
    Source sentence
    encoding market information alone does not provide significant improvement in the non‑regime‑swithcing modeling, since the metrics of VAE and HiVAE are similar.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 19 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.