In the world of quantitative finance, the "Alpha Factory" usually involves two distinct, often manual steps: first, researchers hand-craft features (like momentum or mean-reversion indicators), and second, they apply a separate rule to decide how much of each asset to buy (position sizing).
A new paper from the University of Oxford and the Oxford-Man Institute introduces Deep Inception Networks (DINs), a framework that collapses this pipeline into a single, end-to-end deep learning model. By automating both feature extraction and portfolio optimization, DINs offer a powerful, data-driven alternative to traditional systematic trading.
TL;DR
Deep Inception Networks (DINs) are an end-to-end deep learning framework that simultaneously learns time-series and cross-sectional market patterns to output optimal portfolio weights. By directly optimizing for the Sharpe ratio while penalizing turnover and market correlation, DINs significantly outperform traditional momentum benchmarks across multiple asset classes.
What Problem Does This Paper Address?
Traditional quantitative strategies often suffer from two main bottlenecks:
- Feature Engineering: Researchers spend years "hand-crafting" indicators. These features might work in specific regimes but often fail to generalize as market dynamics shift.
- Fragmented Optimization: Most Machine Learning models in finance are trained to minimize prediction error (e.g., "will the price go up?"). However, a good prediction doesn't always make a good trade once you account for risk, transaction costs, and portfolio-wide correlations.
The authors tackle these issues by creating a model that takes raw price returns as input and directly outputs risk-adjusted position sizes for an entire portfolio, rather than looking at assets in isolation.
Methods
The DIN architecture consists of two primary components:
- Feature Extractors (FE): The authors introduce "Custom Inception Modules" (OrigCIM and FlexCIM). Inspired by computer vision, these use parallel convolutional filters of varying sizes to capture different types of behavior—such as short-term time-series (TS) trends and cross-sectional (CS) relative performance—simultaneously.
- Position Sizers (PS): The extracted features are fed into either a Long Short-Term Memory (LSTM) network or a Temporal Fusion Transformer (TFT). The TFT is particularly notable because it uses attention mechanisms to weigh historical data points based on their relevance to the current market regime.
The Secret Sauce: The Loss Function Unlike standard models, DINs are trained using a custom loss function that directly optimizes the Sharpe ratio. Crucially, it includes a "turnover regularisation" term to keep transaction costs in check and a "correlation cost" term to prevent the model from simply hugging the market index (systemic risk).
Datasets
The authors evaluated DINs across four diverse asset classes to test generalizability:
- Futures: 50 liquid cross-asset contracts (Pinnacle database, 2000–2022).
- Equities: 50 blue-chip European stocks (EURO STOXX 50, 2001–2022).
- Cryptocurrencies: 8 major coins (CoinMarketCap, 2018–2023).
- Foreign Exchange (FX): 19 spot rates (Federal Reserve Board, 2000–2023).
Key Insights and Ideas
- TS + CS = Alpha: The model proves that learning time-series (how an asset moves relative to its own past) and cross-sectional (how it moves relative to its peers) features at the same time is superior to using them separately.
- Semantic Features: By constraining the "shapes" of the convolutional filters, the features learned by the model retain semantic meaning (e.g., identifying lead-lag effects or momentum) rather than becoming "black box" noise.
- Dynamic Importance: Using Variable Selection Networks (VSNs), the authors show that the model's "logic" changes during crises. For example, during the 2008 Great Financial Crisis, the model automatically shifted its focus from short-term returns to longer-term trend features.
- End-to-End Efficiency: DINs perform better without manual volatility scaling at the asset level. This suggests the model learns a more sophisticated internal weighting scheme than the standard "vol-targeting" used by most quants.
Key Results
- Superior Risk-Adjusted Returns: On the futures dataset, the best DIN variant (OrigCIM-TFT) achieved a Sharpe ratio of 2.95, a massive 10x increase over an equal-weight long-only benchmark (0.37).
- Cost Robustness: The strategy remained profitable even after accounting for transaction costs up to 4.82 basis points (bps).
- Benchmark Outperformance: DINs consistently beat established ML benchmarks like LambdaMART and traditional momentum strategies (JT, MOP) across almost all tested metrics.
- Low Correlation: The model maintained a low correlation (approx. ±30%) to the long-only benchmark, proving it provides genuine diversification rather than just leveraged market exposure.
Limitations and Future Work
- Asset Class Specificity: While DINs are general, the "best" feature extractor varies. For example, cross-sectional features were highly effective for futures but less so for equities, where assets are already highly correlated.
- Data Inputs: The current model relies solely on price action. The authors suggest that incorporating fundamental data (like book-to-market ratios) or alternative data (like news sentiment or social media trends) could further improve performance.
- Computational Load: The Transformer-based position sizer is more computationally expensive than simpler models, though the authors argue the interpretability and performance gains are worth the trade-off.
Summary updated