Paper: arXiv 2406.08742
Authors: Joel Ong, Dorien Herremans
Abstract
This paper introduces DeepUnifiedMom, a deep learning framework that enhances portfolio management through a multi-task learning approach and a multi-gate mixture of experts. The essence of DeepUnifiedMom lies in its ability to create unified momentum portfolios that incorporate the dynamics of time series momentum across a spectrum of time frames, a feature often missing in traditional momentum strategies. Our comprehensive backtesting, encompassing diverse asset classes such as equity indexes, fixed income, foreign exchange, and commodities, demonstrates that DeepUnifiedMom consistently outperforms benchmark models, even after factoring in transaction costs. This superior performance underscores DeepUnifiedMom’s capability to capture the full spectrum of momentum opportunities within financial markets. The findings highlight DeepUnifiedMom as an effective tool for practitioners looking to exploit the entire range of momentum opportunities. It offers a compelling solution for improving risk-adjusted returns and is a valuable strategy for navigating the complexities of portfolio management.
Complexity vs Empirical Score
- Math Complexity: 8.0/10
- Empirical Rigor: 8.5/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced deep learning architectures (multi-task learning with multi-gate mixture of experts) involving complex formulations, while demonstrating rigorous empirical validation with comprehensive backtests across multiple asset classes including transaction costs.
Research Flowchart
flowchart TD A["Research Goal:<br>Create unified momentum portfolio<br>via Multi-task Learning & MoE"] --> B["Data & Inputs<br>Asset Classes: Equity, FX, Bonds, Commodities"] B --> C["Core Methodology<br>DeepUnifiedMom Framework"] C --> D["Multi-Task Learning<br>Long/Short signals across time horizons"] C --> E["Multi-Gate Mixture of Experts<br>Specialized expert networks"] D & E --> F["Computational Process<br>Unified Portfolio Construction"] F --> G["Backtesting & Evaluation<br>Transaction costs applied"] G --> H["Key Outcomes<br>Consistent outperformance<br>Superior risk-adjusted returns"]
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 DeepUnifiedMom, a deep multi‑task learning framework with a multi‑gate mixture‑of‑experts to construct unified time‑series momentum portfolios across multiple speeds.
Source sentence
We propose an innovative approach to bridge the research gaps identified earlier inspired by recent advancements in deep learning for portfolio construction
Assumes 5
- The model is trained to maximize the Sharpe Ratio as the objective function. model
Source sentence
Since the objective of the CAN differs from those of the task-specific networks, we use the Sharpe Ratio (Sharpe, 1994) as the objective function for training.
- A soft‑capped Sharpe Ratio is used to prevent over‑fitting to noisy batches. model
Source sentence
We introduce a modified objective function called the Sharpe Ratio with a Soft Capping Mechanism to mitigate this risk.
- Transaction costs of 3 basis points are applied to all backtest results. execution
Source sentence
Table 2: Backtest results (net) for the period from January 2000 to December 2023, with transaction costs set at 3 basis points.
- Continuous futures contracts are constructed using the backward‑ratio method to create a continuous price history. data
Source sentence
We leverage the continuous contract history of each asset, constructed through end‑to‑end concatenation and price adjustment using the backward‑ratio method
Show 1 more
- An expanding‑window cross‑validation with a 10‑year initial training window and 20% validation split is used for model training. other
Source sentence
Throughout the training phase, 20% of the data was allocated for validation purposes.
Data
2000 – 2023 · global equities, fixed income, FX, commodities · multi-asset · daily
- Pinnacle Data Corp CLC database market-prices commercial
Source sentence
The data we used in our experimentation spans from January 1990 to December 2023, providing daily data and encompassing over three decades of historical information.
Costs
Modeled
transaction costs set at 3 basis points
Source sentence
Table 2: Backtest results (net) for the period from January 2000 to December 2023, with transaction costs set at 3 basis points.
Finds 3
Sharpe 2.33 (benchmark 2.31) · annual return 1.9% · max drawdown 1.0%
- DeepUnifiedMom(CAN) achieves a Sharpe Ratio of 1.92 and maximum drawdown of -1.02%, outperforming benchmark TSMOM strategies. performance
Source sentence
DeepUnifiedMom(CAN) 1.92 0.82 2.33 3.81 -1.02
- The unified portfolio weights assigned by DeepUnifiedMom(CAN) remain remarkably consistent over time, diverging from traditional equal‑weighting approaches. performance
Source sentence
Figure 5 shows that the weights assigned by the DeepUnifiedMom(CAN) to the DeepUnifiedMom portfolios —Fast, Medium, and Slow— maintain remarkable consistency throughout the year.
- Backtest results using the Sharpe Ratio with Soft Capping mechanism show superior risk‑adjusted performance across all asset classes. performance
Source sentence
Our extensive backtesting, spanning various asset classes such as equity indexes, bonds, currencies, and commodities, has consistently shown that DeepUnifiedMom surpasses benchmarks, maintaining its superior performance even after accounting for transaction costs.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 43 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (101)
- A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data (2005)
- Tied Multitask Learning for Neural Speech Translation (2018)
- A multimodal model with Twitter FinBERT embeddings for extreme price movement prediction of Bitcoin (2022) — link
- Forecasting Bitcoin volatility spikes from whale transactions and CryptoQuant data using Synthesizer Transformer models (2022) — link
- Fact, Fiction, and Momentum Investing (2014)
- Auxiliary task reweighting for minimum-data learning (2020)
- Task Clustering and Gating for Bayesian Multitask Learning (2003) — link
- Improving Time-Series Momentum Strategies: The Role of Trading Signals and Volatility Estimators (2012)
and 93 more
- Demystifying Time-Series Momentum Strategies: Volatility Estimators, Trading Rules and Pairwise Correlations (2021)
- Momentum has its moments (2015)
- A Model of Inductive Bias Learning (2000)
- Dissecting Investment Strategies in the Cross Section and Time Series (2015) — link
- Identifying beneficial task relations for multi-task learning in deep neural networks (2017)
- Global Portfolio Optimization (1992) — link
- Multitask Learning (1997)
- Drawdown Measure in Portfolio Optimization (2005) — link
- GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks (2018)
- Portfolio Constraints and the Fundamental Law of Active Management (2002) — link
- Multi-task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics (2008)
- A unified architecture for natural language processing: deep neural networks with multitask learning (2008)
- Momentum crashes (2016)
- Optimal Versus Naive Diversification: How Inefficient is the 1/N Portfolio Strategy? (2007) — link
- Adapting Auxiliary Losses Using Gradient Similarity (2018)
- Low Resource Dependency Parsing: Cross-lingual Parameter Sharing in a Neural Network Parser (2015) — link
- Mean-Semivariance Optimization: A Heuristic Approach (2007) — link
- Efficient Capital Markets: A Review of Theory and Empirical Work (1970)
- Two Centuries of Price-Return Momentum (2016)
- The Trend Is Your Friend: Time-Series Momentum Strategies Across Equity and Commodity Markets (2016)
- Portfolio Selection (1952)
- A Century of Evidence on Trend-Following Investing (2017)
- Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency (1993)
- Profitability of Momentum Strategies: An Evaluation of Alternative Explanations (2001)
- Earnings Announcements are Full of Surprises (2008)
- Which Trend Is Your Friend? (2016)
- Auxiliary Tasks in Multi-task Learning (2018) — link
- Time series momentum (2012)
- A Survey on Transfer Learning (2010) — link
- The efficient market hypothesis and calendar anomalies: a literature review (2015)
- Enhancing the momentum strategy through deep regression (2019) — link
- An Overview of Multi-Task Learning in Deep Neural Networks (2017) — link
- The Sharpe Ratio (1994)
- Learning to Learn: Introduction and Overview (1998)
- Discovering Structure in Multiple Learning Tasks: The TC Algorithm (1996)
- Multi-Task Learning via Generalized Tensor Trace Norm (2021) — link
- TREND: Stevens Analytics. Continuous Futures: Future Daily Open Price | Symbol: – | Exchange: All Exchanges | Symbol: ED, 09/28/1983 - 10/23/2020. Data Planet™ Statistical Datasets: A SAGE Publishing Resource Dataset-ID: 095-001-006 (2020) — link
- Time Series Prediction Using Support Vector Machines: A Survey (2009)
- Practical Bayesian support vector regression for financial time series prediction and market condition change detection (2017) — link
- Application of support vector machines in financial time series forecasting (2001) — link
- Financial time series forecasting using support vector machines (2003) — link
- Support vector machine with adaptive parameters in financial time series forecasting (2003) — link
- Long Short-Term Memory (1997) — link
- The Extreme Value Method for Estimating the Variance of the Rate of Return (1980)
- On the Estimation of Security Price Volatilities from Historical Data (1980)
- Estimating Variance From High, Low and Closing Prices (1991)
- Drift‐Independent Volatility Estimation Based on High, Low, Open, and Close Prices (2000)
- Enhancing Time Series Momentum Strategies Using Deep Neural Networks (2019) — link
- Trading with the Momentum Transformer: An Intelligent and Interpretable Architecture (2021) — link
- Attention is All you Need (2017)
- Deep learning for portfolio optimisation (2020) — link
- Trading with the Momentum Transformer: An intelligent and interpretable architecture (2022) — link
- COVID-19 and the march 2020 stock market crash. Evidence from S&P1500 (2021) — link
- The 60/40 problem: Examining the traditional 60/40 portfolio in an uncertain rate environment (2017) — link
- We see risk where others may not: 60/40 in 2020 vision (2020) — link
- Multi-Task Learning for Stock Selection (1996) — link
- A comparison of multitask and single task learning with artificial neural networks for yield curve forecasting (2019) — link
- Federated Reinforcement Learning for Portfolio Management (2022) — link
- Strategic risk management: Designing portfolios and managing risk (2021)
- Continuous control with Stacked Deep Dynamic Recurrent Reinforcement Learning for portfolio optimization (2020) — link
- A CWGAN-GP-based multi-task learning model for consumer credit scoring (2022) — link
- COVID19-MLSF: A multi-task learning-based stock market forecasting framework during the COVID-19 pandemic (2023) — link
- Stock Ranking with Multi-Task Learning (2022) — link
- Constructing time-series momentum portfolios with deep multi-task learning — → 2306.13661
- Strategic Risk Management: Out-of-Sample Evidence from the COVID-19 Equity Selloff (2020)
- Network Momentum across Asset Classes — → 2308.11294
- Performance of Time-series Momentum Strategy: US Evidence (2023) — link
- Designing Robust Trend-following System: Behind the Scenes of Trend-following (2018)
- Long-Only Multi-Asset Momentum: Searching for Absolute Returns (2022)
- Uncovering Momentum (2020)
- Momentum Turning Points (2023)
- Trend Following - Expected Returns (2015)
- Adaptive Mixtures of Local Experts (1991) — link
- Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts (2018) — link
- Switch transformers: scaling to trillion parameter models with simple and efficient sparsity (2022)
- Mixture-of-Experts Meets Instruction Tuning:A Winning Combination for Large Language Models (2023) — link
- ST-MoE: Designing Stable and Transferable Sparse Expert Models (2022) — link
- FasterMoE: modeling and optimizing training of large-scale dynamic pre-trained models (2022) — link
- MegaBlocks: Efficient Sparse Training with Mixture-of-Experts (2022) — link
- Estimation Error and Portfolio Optimization: A Resampling Solution (2007)
- Active Risk-Based Investing (2015) — link
- The Sharpe ratio efficient frontier (2012) — link
- Building Diversified Portfolios that Outperform Out of Sample: (2016) — link
- Advances in Financial Machine Learning (2018)
- Can Machines “Learn” Finance? (2020) — link
- The Probability of Backtest Overfitting (2015)
- Evaluating Trading Strategies (2014)
- Estimating the significance of a signal in a multi-dimensional search (2011) — link
- Deep Learning-based Integrated Framework for stock price movement prediction (2023) — link
- Deep learning for financial applications : A survey (2020) — link
- Research on the construction of stock portfolios based on multiobjective water cycle algorithm and KMV algorithm (2022) — link
- Robust mean-risk portfolio optimization using machine learning-based trade-off parameter (2021) — link
- Multiagent-based deep reinforcement learning for risk-shifting portfolio management (2022) — link
2 of 101 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.