Paper: arXiv 2511.17462
Authors: Ryan Engel, Yu Chen, Pawel Polak, Ioana Boier
Abstract
Conditional Autoencoders (CAEs) offer a flexible, interpretable approach for estimating latent asset-pricing factors from firm characteristics. However, existing studies usually limit the latent factor dimension to around K=5 due to concerns that larger K can degrade performance. To overcome this challenge, we propose a scalable framework that couples a high-dimensional CAE with an uncertainty-aware factor selection procedure. We employ three models for quantile prediction: zero-shot Chronos, a pretrained time-series foundation model (ZS-Chronos), gradient-boosted quantile regression trees using XGBoost and RAPIDS (Q-Boost), and an I.I.D bootstrap-based sample mean model (IID-BS). For each model, we rank factors by forecast uncertainty and retain the top-k most predictable factors for portfolio construction, where k denotes the selected subset of factors. This pruning strategy delivers substantial gains in risk-adjusted performance across all forecasting models. Furthermore, due to each model’s uncorrelated predictions, a performance-weighted ensemble consistently outperforms individual models with higher Sharpe, Sortino, and Omega ratios.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 9.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper introduces advanced statistical learning concepts, including a high-dimensional conditional autoencoder (CAE) and uncertainty-aware factor selection via quantile forecasts from foundation models, resulting in high mathematical density. Furthermore, it is highly empirical, explicitly backtesting with specific models (Chronos, XGBoost, bootstrap), robust metrics (Sharpe, Sortino, Omega ratios), and implementation details (RAPIDS, ensemble methods) on financial time-series data.
Research Flowchart
flowchart TD
A["Research Goal: Scale Conditional Autoencoders<br>for Portfolio Optimization via<br>Uncertainty-Aware Factor Selection"] --> B["Input: High-Dimensional Firm Characteristics<br>and Asset Returns"]
B --> C["Step 1: Generate Latent Factors<br>using Conditional Autoencoder (CAE)"]
C --> D["Step 2: Quantile Prediction & Uncertainty Estimation"]
subgraph D ["Forecasting Models"]
D1["Zero-shot Chronos (ZS-Chronos)<br>Pretrained Time-Series Foundation"]
D2["Quantile Gradient Boosting (Q-Boost)<br>XGBoost & RAPIDS"]
D3["IID Bootstrap Sample Mean (IID-BS)<br>Baseline Model"]
end
D --> E["Step 3: Uncertainty-Aware Factor Selection<br>Rank factors by forecast uncertainty<br>Retain top-k most predictable factors"]
E --> F["Step 4: Portfolio Construction & Evaluation"]
F --> G["Key Outcomes & Findings"]
G --> G1["Substantial gains in risk-adjusted performance<br>across all models"]
G --> G2["Performance-weighted ensemble<br>outperforms individuals"]
G --> G3["Higher Sharpe, Sortino, and Omega ratios<br>achieved via pruning strategy"]
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 a scalable framework that couples a high-dimensional Conditional Autoencoder with an uncertainty‑aware factor selection procedure for portfolio optimization.
Source sentence
we propose a scalable framework that couples a high-dimensional CAE with an uncertainty-aware factor selection procedure.
Assumes 5
- Latent factor returns are i.i.d. from a stationary distribution. model
Source sentence
s=1 are drawn i.i.d. from a stationary distribution.
- Transaction costs are linear in turnover with a proportionality constant κ (e.g., κ=0.001 for 10 bps). execution
Source sentence
κ>0 is the transaction cost proportionality constant, e.g., κ=0.001 corresponds to 10 bps costs.
- The investment universe consists of large‑cap and mid‑cap stocks within the top 2,000 by market equity, updated monthly. data
Source sentence
The investment universe consists of large-cap and mid-cap stocks within the top 2,000 by market equity, updated each month on a rolling basis.
- Factor selection is based on minimizing expected out‑of‑sample utility loss. statistical
Source sentence
By ranking factors based on predictive confidence, the procedure explicitly minimizes expected out-of-sample utility loss.
Show 1 more
- Forecast uncertainty can be estimated via bootstrap resampling (IID‑BS) or quantile regression trees (Q‑Boost) or zero‑shot Chronos. model
Source sentence
bootstrap resamples from the window and compute empirical quantiles
Data
2000 – 2024 · SPY · equities · monthly
- Large‑cap and mid‑cap stock universe (top 2,000 by market equity) fundamentals unknown
Source sentence
The investment universe consists of large-cap and mid-cap stocks within the top 2,000 by market equity, updated each month on a rolling basis.
- Synthetic latent factors generated within the methodological framework synthetic unknown
Source sentence
Thus, the improved performance after 2018 further validates our approach, confirming that the pretrained Chronos model maintains strict data independence and achieves robust generalization.
- M4 dataset (used for pre‑training Chronos) text-news unknown
Source sentence
If pretraining on the M4 dataset introduced any forward‑looking bias, we would expect a measurable decline in performance in the post‑2018 period.
Costs
Modeled
Linear transaction cost model with κ=0.001 (10 bps) per unit turnover; average monthly turnover ≈2.0 (half positions rebalance each month).
Source sentence
The observed average monthly turnover of approximately 2.0 indicates that models rebalance roughly half their positions each month.
Finds 5
Sharpe 2.20 (benchmark 0.56) · annual return 14.4% · total return 3274.2% · max drawdown 9.2%
- Uncertainty‑aware factor selection consistently improves risk‑adjusted performance across all forecasting models, delivering higher Sharpe, Sortino and Omega ratios while keeping max drawdown below 10 %. performance
Source sentence
We show that high-dimensional CAEs, when coupled with forecast‑driven factor selection, significantly outperform conventional low‑dimensional factor models, achieving notably higher Sharpe, Sortino, and Omega ratios while maintaining maximum drawdowns below 10%.
- Ensemble (A) achieves the highest Sharpe ratio (2.20) and Omega ratio (5.95) with lowest volatility (6.52 %) and max drawdown (9.22 %). performance
Source sentence
the highest Sharpe ratio (2.20), and Omega ratio (5.95), as well as the lowest annualized volatility (6.52%) and lowest maximum draw‑down (9.22%).
- Ensemble (B) attains the highest Sortino ratio (4.22) while maintaining a negative market beta. performance
Source sentence
Ensemble (B) achieves the highest Sortino ratio (4.22), while maintaining negative market beta.
- ZS‑Chronos shows superior post‑2018 performance with annual return 11.79 %, Sharpe 1.826 and Sortino 3.898, indicating no forward‑looking bias from pre‑training. performance
Source sentence
ZS‑Chronos attains the highest annual return (11.79%) and demonstrates superior risk‑adjusted metrics, achieving a 1.826 Sharpe ratio and a 3.898 Sortino ratio.
- Average monthly turnover is about 2.0, i.e., roughly half of the portfolio is rebalanced each month, indicating moderate trading intensity. other
Source sentence
The observed average monthly turnover of approximately 2.0 indicates that models rebalance roughly half their positions each month, demonstrating moderate trading intensity relative to the maximum possible.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 55 related papers). Coverage is partial, so this is neither replication nor refutation.
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.
Cited References (17)
- Structural Breaks in Asset Pricing Models (2022)
- Asset Pricing with Omitted Factors (2021)
- Autoencoder Asset Pricing Models (2020)
- Deep Latent Factor Model for Asset Pricing (2025)
- Chronos: Learning the Language of Time Series (2024) — link
- Is There a Replication Crisis in Finance? (2023)
- Greedy Function Approximation: A Gradient Boosting Machine (2001)
- GIFT-Eval : A Benchmark for General Time Series Forecasting Model Evaluation (2024) — link
and 9 more
- RAPIDS : GPU Accelerated Data Science (2018)
- Portfolio Selection and Asset Pricing Models (2000)
- Global Portfolio Optimization (1992)
- Should Investors Assimilate Return Predictability? How to Measure the Incremental Value of Conditioning Information (2010)
- Smooth Minimization of Non-Smooth Functions (2005)
- Ad Click Prediction: A View from the Trenches (2013)
- Introduction to Online Convex Optimization (2016)
- The M4 Competition: 100,000 Time Series and 61 Forecasting Methods (2020)
- The M4 Competition: Results, Findings, Conclusion and Way Forward (2018)
0 of 17 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.