Paper: arXiv 2311.02088

Authors: Koti S. Jaddu, Paul A. Bilokon

Abstract

High-frequency trading is prevalent, where automated decisions must be made quickly to take advantage of price imbalances and patterns in price action that forecast near-future movements. While many algorithms have been explored and tested, analytical methods fail to harness the whole nature of the market environment by focusing on a limited domain. With the evergrowing machine learning field, many large-scale end-to-end studies on raw data have been successfully employed to increase the domain scope for profitable trading but are very difficult to replicate. Combining deep learning on the order books with reinforcement learning is one way of breaking down large-scale end-to-end learning into more manageable and lightweight components for reproducibility, suitable for retail trading. The following work focuses on forecasting returns across multiple horizons using order flow imbalance and training three temporal-difference learning models for five financial instruments to provide trading signals. The instruments used are two foreign exchange pairs (GBPUSD and EURUSD), two indices (DE40 and FTSE100), and one commodity (XAUUSD). The performances of these 15 agents are evaluated through backtesting simulation, and successful models proceed through to forward testing on a retail trading platform. The results prove potential but require further minimal modifications for consistently profitable trading to fully handle retail trading costs, slippage, and spread fluctuation.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced machine learning architectures (deep learning, reinforcement learning) and complex statistical methods, indicating high mathematical sophistication. Empirically, it involves backtesting and forward testing on multiple assets, though real-world retail trading challenges like slippage and costs are noted as unresolved.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Forecast Returns & Generate<br>Profitable Trading Signals"] --> B["Data Processing & Inputs<br>Order Flow Imbalance<br>5 Instruments: FX, Indices, Commodity"]
  B --> C["Core Methodology<br>Deep Learning on Order Books<br>combined with<br>Reinforcement Learning<br>Temporal-Difference Models"]
  C --> D["Computational Process<br>Train 15 Agents<br>Backtesting Simulation"]
  D --> E{"Performance Evaluation"}
  E -- Successful Models --> F["Forward Testing<br>Retail Trading Platform"]
  F --> G["Key Findings/Outcomes<br>Potential Verified<br>Requires Cost Handling<br>Slippage & Spread Optimization"]
  E -- Unsuccessful Models --> C

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 combining deep learning on order books with reinforcement learning to forecast returns across multiple horizons using order flow imbalance and train three temporal‑difference models for five financial instruments to generate trading signals.

Source sentence
The following work focuses on forecasting returns across multiple horizons using order flow imbalance and training three temporal-difference learning models for five financial instruments to provide trading signals.

Assumes 4

  • Using the first ten order‑flow‑imbalance (OFI) levels is sufficient for the models. data
    Source sentence
    The first ten OFI levels were used by Kolm et al. [5] and should be sufficient to carry out this work.
  • Commissions applied to every trade match the costs at CTrader FXPro. execution
    Source sentence
    The commissions applied to every trade match the costs at CTrader FXPro, which from lowest to highest per lot size is GBPUSD/EURUSD, XAUUSD, FTSE100, and DE40.
  • A train‑validation‑test split of 8:1:1 is appropriate for the ten‑week OFI dataset. data
    Source sentence
    Only ten weeks of order flow imbalance data were collected for each instrument and split into 8:1:1 for training, validation, and testing.
  • Grid‑search will find optimal hyper‑parameters, and literature‑suggested values are likely best. model
    Source sentence
    Grid search was used to find optimal parameters for each model, and more likely than not, values suggested in the literature were found to be best.

Data

  • CTrader FXPro OFI dataset · CTrader FXPro order-book proprietary
    23/5/2023‑3/7/2023 and 25/7/2023‑21/8/20 · GBPUSD, EURUSD, DE40, FTSE100, XAUUSD
    Source sentence
    All data was collected on every working day from 23/5/2023 to 3/7/2023 (6 weeks) and from 25/7/2023 to 21/8/2023 (four weeks).

Recognised sources: LOBSTER

Costs

Modeled Costs are part of the reported headline results.

Commissions applied to every trade match the costs at CTrader FXPro, ranging from lowest (GBPUSD/EURUSD) to highest (DE40).

Source sentence
The commissions applied to every trade match the costs at CTrader FXPro, which from lowest to highest per lot size is GBPUSD/EURUSD, XAUUSD, FTSE100, and DE40.

Finds 4

  • Backtesting with retail costs showed profitability using Q‑Learning on GBPUSD and EURUSD, though forward testing did not replicate the results. performance
    Source sentence
    Overall, backtesting with retail costs produced promising results, with profitability achieved using Q Learning on GBPUSD and EURUSD, but failed to bring similar results during forward testing.
  • Alpha‑extraction performance aligns with Kolm et al.’s MLP results, achieving comparable out‑of‑sample R² values. predictive-accuracy
    Source sentence
    The alpha extraction outcome aligns with Kolm et al.’s [5] results using the MLP architecture.
  • FTSE100 and DE40 agents approached profitability during training, while XAUUSD agents were not profitable despite lower transaction costs. performance
    Source sentence
    It is surprising that FTSE100 and DE40 are close to being profitable during training across all agents, while XAUUSD was not profitable at all when having less transaction costs.
  • DQN agents produced negative daily average profits on all instruments in the test set. performance
    Source sentence
    Daily Average Profit (£) -151,000 ... -11,500 ... -12,000 ... -4,830 ... -6,750

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

https://github.com/KotiJaddu/Masters-Project — 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.