Paper: arXiv 2306.09862

Authors: Lifan Zhao, Shuming Kong, Yanyan Shen

Abstract

Stock trend forecasting is a fundamental task of quantitative investment where precise predictions of price trends are indispensable. As an online service, stock data continuously arrive over time. It is practical and efficient to incrementally update the forecast model with the latest data which may reveal some new patterns recurring in the future stock market. However, incremental learning for stock trend forecasting still remains under-explored due to the challenge of distribution shifts (a.k.a. concept drifts). With the stock market dynamically evolving, the distribution of future data can slightly or significantly differ from incremental data, hindering the effectiveness of incremental updates. To address this challenge, we propose DoubleAdapt, an end-to-end framework with two adapters, which can effectively adapt the data and the model to mitigate the effects of distribution shifts. Our key insight is to automatically learn how to adapt stock data into a locally stationary distribution in favor of profitable updates. Complemented by data adaptation, we can confidently adapt the model parameters under mitigated distribution shifts. We cast each incremental learning task as a meta-learning task and automatically optimize the adapters for desirable data adaptation and parameter initialization. Experiments on real-world stock datasets demonstrate that DoubleAdapt achieves state-of-the-art predictive performance and shows considerable efficiency.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)

Why this score: The paper introduces a meta-learning framework with a bi-level optimization formulation, involving data and model adapters, which requires understanding of advanced optimization and distribution shift theory. It is validated on real-world stock datasets with code availability, demonstrating practical implementation and performance metrics, making it empirically rigorous.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Incremental Learning for<br>Stock Trend Forecasting"] --> B{"Core Problem"}
  B --> C["Distribution Shifts<br>(Concept Drifts) Prevent<br>Effective Updates"]
  B --> D["Data: Real-World<br>Stock Market Data"]
  C --> E["DoubleAdapt Framework<br>Meta-Learning for<br>Incremental Adaptation"]
  D --> E
  E --> F["Process: Double Adapters"]
  F --> G["Data Adapter<br>Transforms Data to<br>Locally Stationary Distribution"]
  F --> H["Model Adapter<br>Meta-Learned Initialization<br>for Efficient Updates"]
  G & H --> I["Findings: SOTA Predictive<br>Performance & Efficiency"]

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 DoubleAdapt, an end-to-end incremental learning framework that adapts both data and model for stock trend forecasting.

Source sentence
We propose DoubleAdapt, an end-to-end incremental learning framework for stock trend forecasting, which adapts both the data and the model to cope with distribution shifts in the online environment.

Assumes 3

  • The original incremental data is reliable, allowing test data of a different distribution to be treated as biased. data
    Source sentence
    assuming the original incremental data is reliable, test data of a different distribution can be deemed as a biased dataset.
  • Only features of test data can be adapted because test labels are unknown during online inference. data
    Source sentence
    because test labels are unknown during online inference time.
  • The evaluation focuses on improving the test performance of the current task, as past predictions cannot be withdrawn and future datasets are unseen. execution
    Source sentence
    we only focus on improving the test performance of the current task since past predictions cannot be withdrawn and future datasets are unseen.

Data

  • CSI 300 · Qlib (Alpha360 features) market-prices public
    300 largest China A-share stocks
    Source sentence
    CSI 300 consists of the 300 largest stocks, reflecting the overall performance of the market.
  • CSI 500 · Qlib (Alpha360 features) market-prices public
    500 largest remaining China A-share stocks after excluding CSI 300
    Source sentence
    CSI 500 comprises the largest remaining 500 stocks after excluding the CSI 300 constituents, reflecting the small-mid cap stocks.

Recognised sources: Qlib datasets

Costs

Not modeled A trading or execution result is reported without costs.

Finds 5

  • DoubleAdapt achieves state-of-the-art predictive performance on real-world stock datasets. performance
    Source sentence
    Experiments on real-world datasets demonstrate that DoubleAdapt is generic and efficient, achieving state-of-the-art predictive performance in stock trend forecasting.
  • Applying both model and data adapters, DoubleAdapt attains the best results across different distribution shift types. performance
    Source sentence
    into IL, DoubleAdapt achieves the best results against different kinds of distribution shifts.
  • Data adaptation alone (IL+DA) improves IC by 5.6% over DDG-DA. predictive-accuracy
    Source sentence
    IL+DA outperforms DDG-DA by 5.6% improvement on IC
  • Model adaptation alone (IL+MA) improves IC by 3.3% over C-MAML. predictive-accuracy
    Source sentence
    IL+MA outperforms C-MAML by 3.3% improvement on IC
  • The multi-head version of DoubleAdapt performs best under gradual shifts but may overfit under abrupt shifts. performance
    Source sentence
    In the cases of gradual shifts, the multi-head version with better expressiveness significantly outperforms the single-head one. Nevertheless, the multi-head version performs worse in the cases of abrupt shifts because its high complexity is a double-edged sword and may incur overfitting issues.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 25 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 9 papers in the archive (8 by independent authors)

Code

https://github.com/SJTU-DMTai/qlib — the authors' own repository

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