Weighted Ensemble for Stock Trading

• Weighted ensemble for stock trading is a strategy that integrates multiple predictive models by assigning adaptive weights based on historical and recent performance.
• It employs dynamic weighting schemes, such as rolling information coefficient measures, to improve forecast accuracy and manage portfolio risk.
• Empirical backtesting demonstrates that dynamic ensembles outperform static approaches, driving enhanced returns and robust factor selection.

A weighted ensemble for stock trading refers to a systematic framework that aggregates predictions or signals from a set of diverse models, where each component is assigned a weight reflecting its recent or historical predictive performance. This methodology aims to achieve enhanced returns, greater robustness, and improved risk management by adaptively leveraging the strengths and compensating for the weaknesses of individual methods. Weighted ensemble strategies are central to recent advances in machine learning-based stock selection, and their efficacy is typically validated via extensive backtesting on realistic trading universes.

1. Architecture of Weighted Ensembles in Stock Trading

Weighted ensemble systems combine predictions from multiple, heterogeneous models, which may include linear regressors, neural networks, tree-based models, or more complex architectures such as deep reinforcement learning agents. A canonical framework trains several base learners—commonly including Ridge regression (linear), Multi-layer Perceptron (MLP), and Random Forest regressors—on rolling time windows of historical features and returns. At each forecast period (e.g., monthly), each model produces a cross-section of return predictions, which are then fused into an ensemble forecast by a convex combination: $$\hat r^{\rm ensemble}_t = \sum_{k=1}^K w_k(t)\,\hat r_{k,t}$$ where $\hat r_{k,t}$ denotes the prediction from model $k$ at time $t$, and $w_k(t)$ are the (potentially time-varying) model weights subject to $\sum_k w_k(t) = 1$ (Cai et al., 26 Aug 2025).

The general workflow includes:

• Rolling-window retraining and prediction with each base learner;
• Dynamic combination of per-stock predicted returns via a weighting scheme (static or dynamic);
• Portfolio construction (rank/quantile selection, transaction cost application), return realization, and time rollout.

2. Weight Determination Strategies

Ensemble weights encode the relative trust in each model and are central in translating improvements in forecast accuracy into trading gains. Several principled weighting schemes are extensively analyzed:

Static weighting:

A vector of weights is assigned to models based on pre-specified evaluation metrics $\phi_k$ (such as RMSE, MAPE, Precision, Recall, or F1-score) computed on a recent history. For regression errors, the reciprocal is used so that larger $\psi_k=1/{\rm RMSE}_k$ indicates better performance. The weights are normalized: $w_k = \frac{\psi_k}{\sum_{j=1}^K \psi_j}$ and remain constant or are updated infrequently (Cai et al., 26 Aug 2025).

Dynamic weighting via predictive performance:

A more robust alternative dynamically adapts weights based on rolling Information Coefficient (IC), the rank correlation between the predicted and realized cross-sectional returns over a specified lag (e.g., monthly). For each model $k$: $\hat r_{k,t}$0 Rolling statistics such as mean IC or an IC-stability ratio (mean divided by standard deviation) over the past $\hat r_{k,t}$1 periods are computed: $\hat r_{k,t}$2

$\hat r_{k,t}$3

Weights are assigned by rescaling nonnegative performance scores: $\hat r_{k,t}$4 where $\hat r_{k,t}$5 may be $\hat r_{k,t}$6 or $\hat r_{k,t}$7 (Cai et al., 26 Aug 2025).

Utility- and return-weighted approaches:

In intraday and short-term contexts, ensemble weights can be based directly on realized trading utility (profits per trade), using exponential moving-average or loss-driven updates (Mukherjee et al., 2024).

Dynamic weighting consistently improves out-of-sample performance, especially under non-stationary market conditions.

3. Factor Engineering and Screening

Empirical improvements in weighted ensemble performance are partially attributable to rigorous factor engineering:

• Construction of a large candidate set of factors, including original and newly constructed descriptors (e.g., 50 total in (Cai et al., 26 Aug 2025));
• Preprocessing steps such as industry-median imputation, median absolute deviation trimming, Z-score standardization, and residual demeaning by industry/size;
• Feature selection procedures, typically LASSO regression with penalization parameter $\hat r_{k,t}$8 tuned to maximize out-of-sample IC or minimize validation error;
• Retention of only predictive factors (e.g., 17 out of 30+ after screening), periodically reselected to adapt to evolving factor efficacy (Cai et al., 26 Aug 2025).

This process ensures high-dimensional signal diversity while controlling overfitting, providing a sound basis for model generalization.

4. Empirical Backtesting and Benchmarking

The impact of weighted ensemble approaches is validated via systematic backtesting, with statistical and economic metrics:

Predictive metrics:

• Rolling RMSE, Recall, F1-score, and IC.

Portfolio-level metrics:

• Cumulative and annualized return;
• Annualized volatility and Sharpe ratio;
• Maximum drawdown (maxDD).

For the CSI 300 universe (2020–2022), the dynamic IC-mean ensemble achieved a cumulative return of 39.09%, annualized Sharpe of 0.6117, and maximum drawdown of 70.2%. This sharply exceeded both the best static-weighted ensemble (cumulative return 16.63%, Sharpe 0.2710) and single-model baselines (best single-model Sharpe 0.1858); see table below (Cai et al., 26 Aug 2025):

Model	Cum. Return	Ann. Sharpe	MaxDD
Ridge	11.26%	0.1858	48.7%
MLP	10.97%	0.1777	53.7%
RF	6.23%	0.0878	74.3%
Static Ensemble	16.63%	0.2710	5.44%
Dynamic IC_Mean	39.09%	0.6117	70.2%

Factor screening almost doubled the Sharpe of the best dynamic ensemble. Similar improvements from weighted ensembles are observed in intraday trading (Mukherjee et al., 2024), cross-sectional stock selection (Filipović et al., 2022), and sector-agnostic backtests (Cai et al., 26 Aug 2025).

5. Practical Guidelines for Implementation

Key operational elements in deploying a weighted ensemble stock trading strategy include:

• Data requirements: High-frequency factor series across a broad universe (≥50 stocks) and sufficiently long rolling history (≥5 years for robust IC estimation);
• Window settings: 12–15 months for rolling model re-training; 20–30 months for IC calculation;
• Transaction cost modeling and turnover control to prevent excessive rebalancing;
• Robust preprocessing and regularization, including minimum weight floors and shrinkage toward equal weighting;
• Risk overlays such as stop-loss, beta neutrality, and sector neutrality (Cai et al., 26 Aug 2025);
• Periodic factor re-screening and parameter retuning to mitigate overfitting and accommodate regime shifts.

Additionally, strategies are portable to new universes (e.g., different equity universes or regions) by redoing factor construction and validating IC-window and rebalancing frequencies under local data conventions.

6. Extensions and Comparative Frameworks

The weighted ensemble paradigm is extensible:

• Other weighting criteria (e.g., validation utility, diversity penalties, mean squared error, economic metrics) can be substituted or combined;
• Complex stacking and dynamic meta-learner selection (TDSE) have been explored for multi-source financial data integration, showing further gains in regime adaptation (Gao et al., 16 Dec 2025);
• Incorporation of reinforcement learning agents and adaptive, regime-aware weighting schemes have achieved strong results in both discrete and continuous action spaces (Holzer et al., 18 Jan 2025);
• Utility-based weighting and smoothing windows are especially effective for intraday ensembles (Mukherjee et al., 2024).

The ensemble approach is robust across factor models, deep learning architectures, cross-sectional and time-series domains, and supports flexible adaptation to evolving market structure (Cai et al., 26 Aug 2025, Filipović et al., 2022, Mukherjee et al., 2024, Gao et al., 16 Dec 2025).

7. Significance and Outlook

Weighted ensemble methods unify the best elements of statistical learning, factor investing, and adaptive signal processing for stock trading. The methodology:

• Outperforms single-model and static rules in both backtested and real-world contexts;
• Adapts to drift and non-stationarity via rolling, information-based dynamic weighting;
• Provides a systematic route for incorporating diverse signal sources, rigorous factor selection, and robust risk management.

Recent developments in machine learning, dynamic stacking, and reinforcement learning are converging on weighted ensemble frameworks as a central tool for systematic trading research and deployment (Cai et al., 26 Aug 2025, Gao et al., 16 Dec 2025, Holzer et al., 18 Jan 2025, Filipović et al., 2022).