Paper: arXiv 2408.11740

Authors: CJ Finnegan, James F. McCann, Salissou Moutari

Abstract

In this paper we introduce a multi-agent deep-learning method which trades in the Futures markets based on the US S&P 500 index. The method (referred to as Model A) is an innovation founded on existing well-established machine-learning models which sample market prices and associated derivatives in order to decide whether the investment should be long/short or closed (zero exposure), on a day-to-day decision. We compare the predictions with some conventional machine-learning methods namely, Long Short-Term Memory, Random Forest and Gradient-Boosted-Trees. Results are benchmarked against a passive model in which the Futures contracts are held (long) continuously with the same exposure (level of investment). Historical tests are based on daily daytime trading carried out over a period of 6 calendar years (2018-23). We find that Model A outperforms the passive investment in key performance metrics, placing it within the top quartile performance of US Large Cap active fund managers. Model A also outperforms the three machine-learning classification comparators over this period. We observe that Model A is extremely efficient (doing less and getting more) with an exposure to the market of only 41.95% compared to the 100% market exposure of the passive investment, and thus provides increased profitability with reduced risk.

Complexity vs Empirical Score

  • Math Complexity: 3.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.0)

Why this score: The paper relies on standard machine learning architectures (LSTM, Random Forest, GBT) with minimal novel mathematical derivations, but demonstrates strong empirical rigor through a 6-year backtest on real futures data with detailed performance metrics and risk adjustments.

Research Flowchart

  flowchart TD
  A["Research Goal: Develop an efficient<br>AI model for S&P 500 Futures trading"] --> B["Methodology: Model A<br>Multi-Agent Deep Learning"]
  B --> C["Data Input: Historical Prices & Derivatives<br>(2018-2023)"]
  C --> D["Computational Process: Day-to-day<br>Trading Signals (Long/Short/Closed)"]
  D --> E["Comparison: vs. LSTM, RF, GBT<br>and Passive Buy-and-Hold"]
  E --> F["Outcome 1: 41.95% Market Exposure<br>(vs 100% Passive)"]
  E --> G["Outcome 2: Outperformed Passive &<br>ML Comparators in Key Metrics"]
  F & G --> H["Conclusion: High Efficiency<br>(Less Exposure, More Profit)"]

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 a multi‑agent deep‑learning method called Model A for day‑to‑day futures trading on the S&P 500 index.

Source sentence
In this paper we introduce a multi-agent deep-learning method which trades in the Futures markets based on the US S&P 500 index. The method (referred to as Model A) is an innovation founded on existing well-established machine-learning models...

Assumes 3

  • Trading decisions are made at the daily open and closed at the end of the same day. execution
    Source sentence
    The model needs to instruct us on the correct open position in the S&P futures market, beginning at the daily opening time and closing the position at the end of the day.
  • Only standard price data (open, high, low, close) for ES futures and VIX are used as inputs. data
    Source sentence
    the data environment was limited to: (a) CME E-mini S&P 500 futures contract (ticker symbol ES) price data (open, high, low, close) and (b) the CBOE volatility index (ticker symbol VIX) price data (open, high, low, close).
  • Reduced market exposure (41.95%) lowers risk and trading costs compared with full exposure. market
    Source sentence
    Model A is extremely efficient (doing less and getting more) with an exposure to the market of only 41.95% compared to the 100% market exposure of the passive investment, and thus provides increased profitability with reduced risk.

Data

Backtest sample: 2018 – 2023 · US S&P 500 index futures · futures · daily

  • CME E‑mini S&P 500 futures (ES) price data · CME market-prices commercial
    2018-2023 · ES futures contract
    Source sentence
    the data environment was limited to: (a) CME E-mini S&P 500 futures contract (ticker symbol ES) price data (open, high, low, close)...
  • CBOE Volatility Index (VIX) price data · CBOE market-prices commercial
    2018-2023 · VIX index
    Source sentence
    ...(b) the CBOE volatility index (ticker symbol VIX) price data (open, high, low, close).

Recognised sources: Cboe

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Not modeled

Paper text extraction: Partial

Source sentence (the authors' words)
This reduces trading costs as well as reducing exposure (see Table 5) yet has better profits than the fully exposed methods.

Finds 4

Reported performance (Model A): Sharpe 1.16 (benchmark 0.15) · annual return 14.9% · max drawdown 7.9%

  • Model A’s market exposure was only 41.95% compared with 100% for the passive strategy, delivering higher profitability with reduced risk. performance
    Source sentence
    Model A is extremely efficient (doing less and getting more) with an exposure to the market of only 41.95% compared to the 100% market exposure of the passive investment...
  • Model A outperformed the three machine‑learning classification comparators (LSTM, Random Forest, Gradient‑Boosted Trees) over the six‑year test period. performance
    Source sentence
    Model A also outperforms the three machine-learning classification comparators over this period.
  • Model A chose not to trade on many days, leading to a distribution of daily returns concentrated at 0%, reducing exposure. performance
    Source sentence
    Its large frequency concentration at 0% reflects how frequently the agent action makes a decision not to open a position, and thus reduces exposure.
  • The Dynamic Deep Neural Network approach with limited data inputs can significantly outperform a passive S&P 500 benchmark. performance
    Source sentence
    The results show that a Dynamic Deep Neural Network approach such as Model A, with very limited data inputs can significantly outperform its passive S&P 500 benchmark.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 17 related papers). Coverage is partial, so this is neither replication nor refutation.

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.