Paper: arXiv 2307.05522

Authors: Tom Liu, Stephen Roberts, Stefan Zohren

Abstract

We introduce Deep Inception Networks (DINs), a family of Deep Learning models that provide a general framework for end-to-end systematic trading strategies. DINs extract time series (TS) and cross sectional (CS) features directly from daily price returns. This removes the need for handcrafted features, and allows the model to learn from TS and CS information simultaneously. DINs benefit from a fully data-driven approach to feature extraction, whilst avoiding overfitting. Extending prior work on Deep Momentum Networks, DIN models directly output position sizes that optimise Sharpe ratio, but for the entire portfolio instead of individual assets. We propose a novel loss term to balance turnover regularisation against increased systemic risk from high correlation to the overall market. Using futures data, we show that DIN models outperform traditional TS and CS benchmarks, are robust to a range of transaction costs and perform consistently across random seeds. To balance the general nature of DIN models, we provide examples of how attention and Variable Selection Networks can aid the interpretability of investment decisions. These model-specific methods are particularly useful when the dimensionality of the input is high and variable importance fluctuates dynamically over time. Finally, we compare the performance of DIN models on other asset classes, and show how the space of potential features can be customised.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces complex deep learning architectures (Inception modules, LSTMs, TFTs) with custom loss functions, requiring advanced math, while providing extensive empirical validation using real-world futures data, transaction cost analysis, multiple asset classes, and robustness checks across random seeds.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>General End-to-End Framework<br>for Multi-Asset Quant"] --> B["Core Methodology:<br>Deep Inception Networks<br>TS + CS Features"]
  B --> C["Data Inputs:<br>Futures Daily Price Returns"]
  C --> D["Computational Process:<br>Optimize Portfolio Sharpe<br>+ Turnover Regularization"]
  D --> E["Key Findings:<br>Outperform Benchmarks<br>Robust to Transaction Costs<br>Interpretable via Attention/VS"]

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 Deep Inception Networks (DINs), a family of deep learning models for end‑to‑end systematic multi‑asset trading that directly output portfolio position sizes.

Source sentence
We introduce Deep Inception Networks (DINs), a family of Deep Learning models that provide a general framework for end-to-end systematic trading strategies.

Assumes 5

  • The model is trained with zero turnover and correlation regularisation in validation (Kvalid = Cvalid = 0). model
    Source sentence
    In this project, we set Kvalid = Cvalid = 0.
  • Higher transaction‑cost coefficient C reduces portfolio turnover. execution
    Source sentence
    Increasing C encourages the model to reduce portfolio turnover.
  • Features are extracted directly from daily price returns. data
    Source sentence
    DINs extract time series (TS) and cross sectional (CS) features directly from daily price returns.
  • A proportional transaction‑cost model is used for evaluating profitability. execution
    Source sentence
    This is profitable until 4.82 basis points (bps) costs under a proportional cost model.
Show 1 more
  • The training loss accounts for transaction costs and correlation to a long‑only benchmark. model
    Source sentence
    where Rp,t are volatility‑scaled portfolio returns adjusted for transaction costs.

Data

  • Futures · Pinnacle market-prices unknown
    2000-2022 · 50 cross‑asset futures contracts
    Source sentence
    over 2000-2022. We select 50 cross-asset contracts
  • Equities · Compustat market-prices unknown
    2001-2022 · EURO STOXX 50 (EUR50) 50 blue‑chip European stocks
    Source sentence
    EURO STOXX 50 (EUR50) comprises 50 blue‑chip European stocks from Compustat
  • Cryptocurrencies · CoinMarketCap crypto-onchain public
    2018-2023 · 8 spot rates, top 25 coins by market cap as of Jan 2019
    Source sentence
    top 25 coins by market capitalisation (MCAP) on CMC as of January 2019

Costs

Modeled Costs are part of the reported headline results.

Performance reported with a proportional cost model; results shown for cost coefficients C from 0 to 5 bps and a breakeven cost of 4.82 bps.

Source sentence
This is profitable until 4.82 basis points (bps) costs under a proportional cost model.

Finds 4

  • The best DIN variant (OrigCIM‑TFT) achieves a Sharpe ratio of 2.95 and a breakeven transaction‑cost of 4.82 bps on the futures dataset. performance
    Source sentence
    The optimal DIN variant for this dataset is OrigCIM‑TFT with Sharpe = 2.95 and BRK = 4.82 bps.
  • DIN models outperform all traditional TS and CS benchmarks before and after transaction costs, delivering about a 10× Sharpe increase over an equal‑weight long‑only portfolio. performance
    Source sentence
    Evaluating on cross‑asset futures data from 2005‑2022, DIN models (OrigCIM x TFT) outperform all benchmarks before and after transaction costs. In particular, there is a 10x increase in Sharpe ratio to 2.95, compared to an equal-weight Long‑only portfolio, prior to transaction costs.
  • DIN performance remains robust across a range of transaction‑cost coefficients up to 4 bps, with Sharpe declining gradually as costs increase. performance
    Source sentence
    Exhibit 7: Strategy performance at different transaction cost coefficients C
  • The TFT position sizer consistently outperforms the LSTM variant and the long‑only benchmark across all time periods, with correlation to long‑only staying low (±30%). performance
    Source sentence
    The TFT‑variant outperforms both the LSTM‑variant and Long‑only benchmark in all time periods. The correlation to Long‑only for both DIN variants also remains stably low between ±30%.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 25 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 7 papers in the archive (5 by independent authors): DSPO: An End-to-End Framework for Direct Sorted Portfolio … (uses its method or data)

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.