Paper: arXiv 2601.07942
Authors: Brandon Luo, Jim Skufca
Abstract
Our work focuses on deep learning (DL) portfolio optimization, tackling challenges in long-only, multi-asset strategies across market cycles. We propose training models with limited regime data using pre-training techniques and leveraging transformer architectures for state variable inclusion. Evaluating our approach against traditional methods shows promising results, demonstrating our models’ resilience in volatile markets. These findings emphasize the evolving landscape of DL-driven portfolio optimization, stressing the need for adaptive strategies to navigate dynamic market conditions and improve predictive accuracy.
Complexity vs Empirical Score
- Math Complexity: 7.0/10
- Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs only partially modelled; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)
Why this score: The paper employs advanced deep learning architectures (LSTMs, transformers) and optimization techniques (gradient ascent, Sharpe ratio loss), representing high mathematical complexity. It demonstrates strong empirical rigor with specific backtesting protocols, statistical tests (Mann-Whitney U), and robustness experiments across extended timeframes and asset compositions.
Research Flowchart
flowchart TD
R["Research Goal: Enhancing Portfolio Optimization with DL"] --> I["Data & Inputs"]
subgraph I ["Inputs"]
I1["Multi-Asset Market Data"]
I2["Limited Regime Data"]
I3["State Variables"]
end
I --> M["Methodology"]
subgraph M ["Key Methodology"]
M1["Pre-training on Limited Data"]
M2["Transformer Architecture Integration"]
M3["Regime Adaptation Training"]
end
M --> C["Computational Process"]
subgraph C ["Process"]
C1["Model Training"]
C2["Long-Only Strategy Generation"]
C3["Multi-Asset Portfolio Allocation"]
end
C --> F["Key Findings & Outcomes"]
subgraph F ["Outcomes"]
F1["Promising Results vs. Traditional Methods"]
F2["Resilience in Volatile Markets"]
F3["Adaptive Strategies for Dynamic Conditions"]
end
F --> End["Conclusions: Evolving DL-Driven Optimization Landscape"]
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 training deep‑learning portfolio models with limited regime data using pre‑training techniques and transformer architectures that incorporate state variables.
Source sentence
We propose training models with limited regime data using pre‑training techniques and leveraging transformer architectures for state variable inclusion.
Assumes 5
- Daily trading adjustments are possible with negligible slippage. execution
Source sentence
Considering the active trading nature of our LSTM model, like the approach used in the Oxford study, we accounted for a 0.01% transaction fee on daily adjustments in allocation weights to determine net returns.
- Missing daily data points can be filled by forward‑filling and back‑filling. data
Source sentence
To ensure consistency in daily (trading days) data points, we employed forward filling and backfilling techniques.
- Economic indicators reported in units can be transformed to year‑over‑year percent changes for comparability. data
Source sentence
we transformed twelve economic indicators reported in units into year‑over‑year percent changes, ensuring a standardized comparison across the dataset.
- Pre‑training on a broader historical dataset improves transfer learning for under‑represented regimes. model
Source sentence
The Oxford study's data limitations led us to hypothesize that pretraining the model on broader historical data from 1986 could improve transfer learning, especially for less represented scenarios in the 2006‑2023 period.
Show 1 more
- L2 regularization and dropout mitigate overfitting in the transformer model. model
Source sentence
To enhance model robustness, we added L2 Regularization to the Adam optimizer and included a dropout layer for normalization and to counter overfitting.
Data
US equities, bonds, commodities and volatility index · multi-asset · daily
- Oxford dataset macro unknown
Source sentence
The Oxford study's data limitations led us to hypothesize that pretraining the model on broader historical data from 1986 could improve transfer learning, especially for less represented scenarios in the 2006‑2023 period.
- Pre‑training dataset market-prices unknown
Source sentence
Our pre‑training dataset included stock and bond price data from our asset experiments (Wilshire and BofA) dating back to 1986, for initializing weights in VTI and AGG funds.
- Economic indicators (21 new features) macro public
Source sentence
These features comprised two financial market indicators and 19 economic indicators (Figure 1).
Cboe Kaggle Yahoo Finance
Costs
Unclear - sources disagree
Source sentence (the authors' words)
Considering the active trading nature of our LSTM model, like the approach used in the Oxford study, we accounted for a 0.01% transaction fee on daily adjustments in allocation weights to determine net returns.
Finds 3
Sharpe 2.70 (benchmark 0.80)
- The replicated LSTM model achieves a Sharpe ratio of 1.828 (net of transaction costs), closely matching the Oxford study's 1.858 and outperforming MVO and Balanced benchmarks in annual return, profitability, and max drawdown. performance
Source sentence
Our replicated portfolio's Sharpe ratio closely aligns with the Oxford study (1.828 vs 1.858), demonstrating superior returns in comparison to the MVO and Balanced portfolios in terms of annual return, profitability (percent of positive returns), and risk metrics (max drawdown).
- Training and validation curves show most validation Sharpe peaks between 20 and 40 epochs, with some models improving up to epoch 100, suggesting early stopping to avoid overfitting. statistical
Source sentence
These curves revealed five distinct training patterns, with most validation curves peaking in Sharpe ratios between 20 and 40 epochs. One curve improved until epoch 100, while another peaked around epoch 90, suggesting the Oxford team might have stopped training at epoch 100 for generalization and to avoid overfitting.
- Dynamic asset allocations vary widely: VTI weights from 0.1‑0.6, AGG from 0.2‑0.9, DBC from 0.0‑0.07, and ^VIX from 0.0‑0.25 across time. performance
Source sentence
Figure 6 shows the dynamic nature of our LSTM model in reallocating capital across asset classes, with significant variations in allocations to VTI (0.1 - 0.6) and AGG (0.2 - 0.9), while allocations to DBC and ^VIX ranged from 0.0 to 0.07 and 0.0 to 0.25, respectively.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 30 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.