← Back to research← 返回研究列表

Mixed-Frequency Deep Learning for Cross-Sectional Stock Prediction混频深度学习与横截面股票预测

Do financial statements still add alpha beyond price and volume? An empirical study on CSI 300 constituents.财务信息是否仍能在量价数据之外贡献增量 alpha？基于沪深 300 成分股的实证研究。

Deep LearningEquitiesPyTorch627k observations1,225-day backtest深度学习股票PyTorch62.7 万观测1,225 日回测

Built a mixed-frequency deep learning pipeline combining quarterly financial statement data (MLP branch) with daily price/volume data (CNN-LSTM branch), fused by a self-attention layer.
Trained with a pairwise ranking loss so the model optimizes the cross-sectional ranking of stocks — the objective that actually matters for selection — instead of pointwise return regression.
Adding financial features improved cross-sectional Mean IC by ~26% (0.0073 → 0.0092) and RankIC by ~33% (0.0050 → 0.0067); ICIR rose from 0.11 to 0.16.
A top-10 portfolio backtest returned +69.3% annualized (Sharpe 3.28) with +43.1% annualized excess return versus the equal-weighted CSI 300, in simulation under the simplified assumptions described below.

构建了混频深度学习流程：将季度财务数据（MLP 分支）与日频量价数据（CNN-LSTM 分支）通过自注意力层融合。
采用成对排序损失训练，直接优化股票的横截面排序——这才是选股真正需要的目标，而非逐点收益率回归。
引入财务特征后，横截面 Mean IC 提升约 26%（0.0073 → 0.0092），RankIC 提升约 33%（0.0050 → 0.0067），ICIR 从 0.11 升至 0.16。
Top-10 组合回测年化 +69.3%（夏普 3.28），较等权沪深 300 年化超额 +43.1%——基于下文所述简化假设的模拟结果。

1. Motivation1. 研究动机

Most daily-frequency stock-selection models are trained only on price/volume data. Accounting fundamentals move slowly — they are updated quarterly — but they remain the primary channel through which firm value is communicated to the market.

多数日频选股模型仅使用量价数据训练。会计基本面变化缓慢——按季度更新——但仍然是市场理解公司价值的主要渠道。

This study asks a narrow, practical question: when market and fundamental data have different sampling frequencies, does a deep model with an explicit fusion mechanism extract incremental predictive power from financial statements?

本研究提出一个具体的问题：当市场数据与基本面数据具有不同采样频率时，带有显式融合机制的深度模型能否从财务报表中提取增量预测能力？

Design choice	Why
Separate MLP branch for quarterly financials	Fundamentals are tabular, low-frequency and smooth — a different inductive bias from price series
CNN-LSTM branch for daily OHLCV	CNN captures local price-volume structure; LSTM captures longer temporal dependence
Self-attention at the fusion layer	Lets the model dynamically re-weight fundamental vs. technical signals as the regime changes
Pairwise ranking loss	Portfolio construction only needs the order of predicted returns, not their level

设计选择	原因
季度财务走独立 MLP 分支	基本面是表格型、低频、平滑的数据，与价格序列的归纳偏置不同
日频 OHLCV 走 CNN-LSTM 分支	CNN 捕捉局部量价结构，LSTM 捕捉更长的时间依赖
融合层使用自注意力	让模型随市场状态动态调整基本面与技术面信号的权重
成对排序损失	组合构建只需要预测收益的排序，而非绝对水平

2. Data2. 数据

Block	Frequency	Content	Source
Price / volume	Daily	OHLCV and derived technical inputs for CSI 300 constituents	Wind terminal
Financial statements	Quarterly (aligned to daily)	Balance-sheet, income-statement and cash-flow items, standardized	Wind terminal

数据块	频率	内容	来源
价格 / 成交量	日频	沪深 300 成分股 OHLCV 及衍生技术指标	Wind 终端
财务报表	季度（对齐至日频）	资产负债表、利润表、现金流量表科目，已标准化	Wind 终端

Sample: CSI 300 constituents, 2016 – 2026, 627,075 stock-day rows. Financial features are aligned to the trading calendar from each item's disclosure date, so the model can only use information available at prediction time.

样本：沪深 300 成分股，2016 – 2026，共 627,075 条股票-日观测。财务特征按披露日期对齐至交易日历，确保模型仅使用预测时点可得的信息。

3. Methodology3. 方法

quarterly financials ──► MLP encoder ──────────┐
                                               ├─► Self-Attention fusion ─► ranking head
daily OHLCV ──► CNN (local) ──► LSTM (temporal)┘        (pairwise ranking loss)

季度财务报表 ──► MLP 编码器 ────────────┐
                                       ├─► 自注意力融合 ─► 排序输出头
日频量价数据 ──► CNN（局部）─► LSTM（时序）┘      （成对排序损失）

Two asymmetric encoders are combined by a single-head self-attention block that learns per-sample fusion weights, followed by a linear ranking head. The pairwise ranking objective penalizes the model whenever a lower-ranked stock is predicted above a higher-ranked one, which reduced the tendency of plain regression to chase the magnitude of noisy daily returns.

两个非对称编码器由单头自注意力块融合，学习每个样本的融合权重，后接线性排序头。成对排序目标在市场把低排序股票预测得高于高排序股票时施加惩罚，抑制了普通回归追逐噪声日收益量级的问题。

Two runs are trained with identical pipelines as an ablation: Model-T (price/volume branch only) and Model-F (full fused model). Fusion-layer attention weights are inspected per period to check whether the model shifts between fundamental and technical logic over time.

以完全相同的流程训练两个模型作为消融：Model-T（仅量价分支）与 Model-F（完整融合模型）。我们会按时期检查融合层注意力权重，观察模型是否在市场环境变化时在基本面与技术面逻辑之间切换。

4. Results4. 结果

Model	MSE	Mean IC ↑	Mean RankIC ↑	ICIR ↑	RankICIR ↑
Model-T (price only)	0.000525	0.0073	0.0050	0.110	0.075
Model-F (fused)	0.000530	0.0092	0.0067	0.164	0.114

模型	MSE	Mean IC ↑	Mean RankIC ↑	ICIR ↑	RankICIR ↑
Model-T（仅量价）	0.000525	0.0073	0.0050	0.110	0.075
Model-F（融合）	0.000530	0.0092	0.0067	0.164	0.114

Financial information buys a clear improvement in rank quality (+26% IC, +33% RankIC, +49% ICIR) while leaving point-error metrics essentially unchanged. For stock selection, that is the trade-off that matters. On the study's headline metric — peak validation RankIC — the fused model reached 0.0335 versus 0.0268 for the baseline (+25%).

财务信息带来的是排序质量的明确提升（IC +26%，RankIC +33%，ICIR +49%），而逐点误差指标基本不变。对选股而言，这正是最重要的权衡。在以论文口径为准的指标——验证集峰值 RankIC 上，融合模型达到 0.0335，基线为 0.0268（+25%）。

Model comparison — Validation metrics and NAV comparison between the price-only and fused models.纯量价模型与融合模型的验证指标与净值对比。

Data fusion schematic — Mixed-frequency fusion: quarterly statements and daily price-volume streams.混频融合：季度报表流与日频量价流。

Stock selection behaviour选股行为

Selection frequency — How often individual names are selected by the fused model.各股票被融合模型选中的频率。

Profile of selected names and their scores.被选中股票的画像与得分分布。

5. Backtest5. 回测

Daily selection of the top-10 ranked stocks, in simulation:

每日选取排名前 10 的股票（模拟回测）：

Strategy	Annualized return	Sharpe	Max drawdown	Alpha
Model-F (fused)	69.3%	3.28	−11.1%	+43.1%
Model-T (price only)	50.8%	2.31	−12.6%	+24.5%
CSI 300 (equal-weight benchmark)	26.2%	1.62	−11.1%	—

策略	年化收益	夏普	最大回撤	Alpha
Model-F（融合）	69.3%	3.28	−11.1%	+43.1%
Model-T（仅量价）	50.8%	2.31	−12.6%	+24.5%
沪深 300（等权基准）	26.2%	1.62	−11.1%	—

For reference, the extreme top-1 configuration returned 144.7% annualized (Sharpe 2.98), but concentrating in a single name is not deployable at scale — the top-10 configuration above is the headline result.

作为参考，极端的 top-1 配置年化收益为 144.7%（夏普 2.98），但单票集中不可规模化部署；上表 top-10 配置为主要结果。

6. Backtest caveats6. 回测局限

These are simulation results, not live performance. Known limitations:

以上为模拟结果，并非实盘表现。已知局限：

No full market-impact or capacity modelling; concentrated top-N portfolios are not deployable at scale without significant slippage.
Single market (CSI 300) and a single historical window; the prediction period covers a specific macro regime.
Transaction costs, borrow constraints and limit-up/limit-down frictions are simplified.
Standard overfitting risk for any pipeline selected partly on validation metrics.

未完整建模市场冲击与容量；集中持仓的 top-N 组合在规模上无法在不产生显著滑点的情况下部署。
单一市场（沪深 300）与单一历史区间；预测期覆盖特定的宏观环境。
交易成本、融券约束与涨跌停摩擦均做了简化。
任何部分依据验证指标选择的流程都存在过拟合风险。

The honest reading: the backtest is directional evidence that financial features improve the ranking signal — not a claim of 145% annual returns.

诚实的读法是：回测是方向性证据——财务特征提升了排序信号——而不是年化 145% 的收益承诺。

7. Limitations & next steps7. 局限与后续方向

Industry / size neutralization and incremental IC after neutralization.
Walk-forward retraining instead of a single split.
Realistic costs and capacity; evaluate a top-N portfolio rather than top-1.
Extend the fusion layer to more data frequencies (weekly analyst data, monthly macro).

行业 / 市值中性化，以及中性化后的增量 IC。
以滚动窗口（walk-forward）重训取代单次划分。
真实的成本与容量建模；评估 top-N 组合而非 top-1。
将融合层扩展到更多数据频率（周频分析师数据、月频宏观数据）。