Paper: arXiv 2307.13832

Authors: Tom Liu, Stefan Zohren

Abstract

Cryptocurrency trading represents a nascent field of research, with growing adoption in industry. Aided by its decentralised nature, many metrics describing cryptocurrencies are accessible with a simple Google search and update frequently, usually at least on a daily basis. This presents a promising opportunity for data-driven systematic trading research, where limited historical data can be augmented with additional features, such as hashrate or Google Trends. However, one question naturally arises: how to effectively select and process these features? In this paper, we introduce Multi-Factor Inception Networks (MFIN), an end-to-end framework for systematic trading with multiple assets and factors. MFINs extend Deep Inception Networks (DIN) to operate in a multi-factor context. Similar to DINs, MFIN models automatically learn features from returns data and output position sizes that optimise portfolio Sharpe ratio. Compared to a range of rule-based momentum and reversion strategies, MFINs learn an uncorrelated, higher-Sharpe strategy that is not captured by traditional, hand-crafted factors. In particular, MFIN models continue to achieve consistent returns over the most recent years (2022-2023), where traditional strategies and the wider cryptocurrency market have underperformed.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper uses advanced machine learning concepts like Deep Inception Networks and multi-factor models, which involve non-trivial mathematics and neural network architectures, scoring high on math complexity. It demonstrates strong empirical rigor through a detailed backtesting framework on real cryptocurrency data, employing robust validation methods like expanding-window splits, transaction cost adjustments, and volatility scaling to ensure realistic, implementable results.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Effective Feature Utilization<br>in Crypto Trading"] --> B["Data Inputs<br>Price Returns + Additional Features<br>e.g., Hashrate, Google Trends"]
  B --> C["Methodology<br>Multi-Factor Inception Network MFIN<br>End-to-End Deep Learning"]
  C --> D["Computational Process<br>Deep Inception Architecture<br>Auto-feature learning & Sharpe Optimization"]
  D --> E["Key Findings<br>Higher Sharpe Ratio & Uncorrelated Strategy"]
  E --> F["Outcome<br>Consistent Returns 2022-2023<br>Outperforms Traditional Benchmarks"]

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 Multi-Factor Inception Networks (MFIN), an end‑to‑end framework that learns features from multiple cryptocurrency factors and outputs portfolio positions to optimise Sharpe ratio.

Source sentence
In this paper, we introduce Multi‑Factor Inception Networks (MFIN), an end‑to‑end framework for systematic trading with multiple assets and factors.

Assumes 5

  • The spread between price and alternative feature is stationary. statistical
    Source sentence
    We verify that the spread 𝛿𝑖,𝑗,𝑡 is stationary by applying an Augmented Dicky Fuller (ADF) test, requiring a p-value ≤1%.
  • Model training uses an expanding window with a 90%/10% train/validation split. model
    Source sentence
    Following [20, 27, 32, 33], we train our model with an expanding window approach with a 90%/10% train/validation split for hyper‑parameter optimisation.
  • Transaction costs are incorporated into the loss function for realistic performance evaluation. execution
    Source sentence
    This ensures that the transaction costs in the loss function of Eq. 5 are correct.
  • Data are standardised daily returns for each factor, reducing dimensionality. data
    Source sentence
    We apply MFIN to standardised daily returns in each factor. This reduces the dimensionality of the input and features are instead learnt automatically from the data.
Show 1 more
  • The cryptocurrency dataset uses 1‑year increments for training. data
    Source sentence
    The cryptocurrency dataset uses 1‑year increments, as shown in Figure 1.

Data

  • Cryptocurrency price and alternative data set · Binance and public web sources alternative public
    1‑year increments (e.g., 2020‑2023) · multiple cryptocurrencies
    Source sentence
    The cryptocurrency dataset uses 1‑year increments, as shown in Figure 1.
  • Google Trends data · Google alternative public
    Source sentence
    many metrics describing cryptocurrencies are accessible with a simple Google search and update frequently, usually at least on a daily basis.
  • Hashrate data other public
    Source sentence
    additional features, such as hashrate or Google Trends.

Recognised sources: Binance CoinMarketCap

Costs

Modeled Costs are part of the reported headline results.

Breakeven transaction cost (BRK) is reported; MFIN delivers positive cost‑adjusted Sharpe up to BRK = 17.5 bps, above Binance fee of 10.0 bps.

Source sentence
Considering fees, MFIN is also able to deliver positive cost‑adjusted Sharpe ratios up to 𝐵𝑅𝐾= 17.5 bps, above the standard Binance trading fee of 10.0 bps.

Finds 5

  • MFIN outperforms traditional momentum and reversion strategies, achieving a higher Sharpe ratio (1.89) and lower maximum drawdown (0.64 σ). performance
    Source sentence
    MFIN has higher 𝑆𝑕𝑎𝑟𝑝𝑒= 1.89 and 𝐵𝑅𝐾= 20.9 bps, whilst experiencing lower 𝑀𝐷𝐷= 0.64 𝜎 than either underlying strategy.
  • MFIN’s Calmar ratio is 2.05, a threefold increase over Long‑only and double that of the combined benchmark. performance
    Source sentence
    Accounting for returns, 𝐶𝑎𝑙𝑚𝑎𝑟= 2.05 is a threefold increase compared to Long‑only, and double that of CMB.
  • MFIN remains profitable after transaction costs up to 5 bps, with breakeven cost of 17.5 bps. performance
    Source sentence
    Considering fees, MFIN is also able to deliver positive cost‑adjusted Sharpe ratios up to 𝐵𝑅𝐾= 17.5 bps, above the standard Binance trading fee of 10.0 bps.
  • Traditional strategies achieve Sharpe ratios between 1.4 and 1.7, which can be combined to a Sharpe of 3.07 using a volatility‑scaled CMB portfolio. performance
    Source sentence
    A simple volatility‑scaled CMB portfolio increases 𝑆𝑕𝑎𝑟𝑝𝑒 to 3.07, with a high breakeven transaction cost of 54.0 bps.
  • BAZ strategies with MACD indicators retain performance out‑of‑sample with only an 18 % decay in Sharpe after removing look‑ahead bias. performance
    Source sentence
    Removing lookahead bias, BAZ strategies with MACD indicators continue to perform out‑of‑sample, with only an 18% decay in Sharpe ratio.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 33 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive

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.