Paper: arXiv 2310.10500

Authors: Kieran Wood, Samuel Kessler, Stephen J. Roberts, Stefan Zohren

Abstract

Forecasting models for systematic trading strategies do not adapt quickly when financial market conditions rapidly change, as was seen in the advent of the COVID-19 pandemic in 2020, causing many forecasting models to take loss-making positions. To deal with such situations, we propose a novel time-series trend-following forecaster that can quickly adapt to new market conditions, referred to as regimes. We leverage recent developments from the deep learning community and use few-shot learning. We propose the Cross Attentive Time-Series Trend Network – X-Trend – which takes positions attending over a context set of financial time-series regimes. X-Trend transfers trends from similar patterns in the context set to make forecasts, then subsequently takes positions for a new distinct target regime. By quickly adapting to new financial regimes, X-Trend increases Sharpe ratio by 18.9% over a neural forecaster and 10-fold over a conventional Time-series Momentum strategy during the turbulent market period from 2018 to 2023. Our strategy recovers twice as quickly from the COVID-19 drawdown compared to the neural-forecaster. X-Trend can also take zero-shot positions on novel unseen financial assets obtaining a 5-fold Sharpe ratio increase versus a neural time-series trend forecaster over the same period. Furthermore, the cross-attention mechanism allows us to interpret the relationship between forecasts and patterns in the context set.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10 (capped from 9.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 9.0)

Why this score: The paper employs advanced deep learning architectures (cross-attention) and few-shot learning frameworks with formal loss functions, indicating high mathematical complexity. It also provides extensive empirical validation with backtests across multiple asset classes, specific performance metrics (Sharpe ratios), and a reference to public code, demonstrating high empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal: Develop a fast-adapting financial forecaster for changing market regimes."] --> B["Methodology: Few-Shot Learning with Cross-Attention.<br>Proposed Model: X-Trend Network"]
  B --> C{"Data & Inputs: Financial Time-Series"}
  C --> D["Training: Context Set<br>(Source Regimes)"]
  C --> E["Inference: Target Regime<br>(New Market Conditions)"]
  D --> F["Computation: X-Trend Network<br>1. Attend to Context Patterns<br>2. Transfer Trends<br>3. Generate Forecast"]
  E --> F
  F --> G["Outcome: Trend-Following Trading Strategy"]
  G --> H["Key Findings:<br>• 18.9% ↑ Sharpe Ratio vs. Neural Forecaster<br>• 10x ↑ Sharpe vs. Conventional Momentum<br>• 2x Faster Recovery from COVID Drawdown<br>• Zero-Shot capability on unseen assets<br>• Interpretability via Cross-Attention"]

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 the Cross Attentive Time-Series Trend Network (X‑Trend), a few‑shot learning model that attends over a context set of financial regimes to make trend‑following forecasts.

Source sentence
We propose the Cross Attentive Time-Series Trend Network – X-Trend – which takes positions attending over a context set of financial time-series regimes.

Assumes 5

  • Transaction costs are set to zero for each asset. execution
    Source sentence
    For this paper we set C(i) = 0 rather than assuming a cost for each asset, focusing…
  • Volatility targeting is used with a leverage factor σtgt/σ(i). model
    Source sentence
    We use volatility targeting [5, 26, 27], which introduces a leverage factor σtgt/σ(i)
  • An expanding window training approach is employed for out‑of‑sample testing. data
    Source sentence
    we use an expanding window approach, where we initially train on 1990 to 1995, test out‑of‑sample on the period from 1995 to 2000, expand the training window to from 1990 to 2000, then test out‑of‑sample on the subsequent 5 years and so on.
  • Gaussian return distribution is assumed in low‑resource (zero‑shot) settings. statistical
    Source sentence
    the simpler assumption of Gaussian returns is favourable in a low‑resource setting.
Show 1 more
  • Change‑point detection is used to segment the context set into regimes. model
    Source sentence
    by segmenting the context set with CPD, we improve Sharpe by a further 11.3%, showing the benefit of constructing context sets with regimes.

Data

  • Pinnacle Data Corp CLC Database futures contracts · Pinnacle Data Corp market-prices commercial
    1990‑2023 · 50 most liquid continuous futures contracts
    Source sentence
    We backtest our X‑Trend variants on a portfolio of 50 of the most liquid, continuous futures contracts

Costs

Not modeled A trading or execution result is reported without costs.

Source sentence
For this paper we set C(i) = 0 rather than assuming a cost for each asset, focusing…

Finds 5

  • X‑Trend improves Sharpe by 18.9% over the baseline (neural forecaster) and about 10‑fold over conventional time‑series momentum during 2018‑2023. performance
    Source sentence
    During this 5‑year period X‑Trend improves upon the Sharpe of the baseline strategy by 16.9% and X‑Trend‑Q improves upon the baseline by 18.9%.
  • X‑Trend recovers from the COVID‑19 drawdown almost twice as quickly as a neural forecaster (162 vs 254 trading days). performance
    Source sentence
    the drawdown ends after 162 trading days for X‑Trend‑Q compared to 254 days for the Baseline which is an entire year and almost twice as long as X‑Trend‑Q.
  • Zero‑shot X‑Trend achieves a Sharpe of 0.47 on unseen futures, outperforming loss‑making time‑series momentum baselines. performance
    Source sentence
    we achieve a Sharpe of 0.47 compared to loss‑making time‑series momentum baselines, both deep‑learning based and conventional TSMOM.
  • Segmenting the context set with change‑point detection adds an 11.3% Sharpe improvement. performance
    Source sentence
    by segmenting the context set with CPD, we improve Sharpe by a further 11.3%, showing the benefit of constructing context sets with regimes.
  • X‑Trend‑Q outperforms X‑Trend‑G by 7.1% in Sharpe in the few‑shot setting, indicating non‑Gaussian modeling benefits. performance
    Source sentence
    Sharpe for X‑Trend‑Q outperforms X‑Trend‑G by 7.1% in the few‑shot setting over the turbulent period of 2018 to 2023.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 21 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 4 papers in the archive (1 by independent authors)

Code

https://github.com/kieranjwood/x-trend — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.