Paper: arXiv 2403.18831
Authors: Armand Mihai Cismaru
Abstract
In this paper, we introduce DeepTraderX (DTX), a simple Deep Learning-based trader, and present results that demonstrate its performance in a multi-threaded market simulation. In a total of about 500 simulated market days, DTX has learned solely by watching the prices that other strategies produce. By doing this, it has successfully created a mapping from market data to quotes, either bid or ask orders, to place for an asset. Trained on historical Level-2 market data, i.e., the Limit Order Book (LOB) for specific tradable assets, DTX processes the market state \(S\) at each timestep \(T\) to determine a price \(P\) for market orders. The market data used in both training and testing was generated from unique market schedules based on real historic stock market data. DTX was tested extensively against the best strategies in the literature, with its results validated by statistical analysis. Our findings underscore DTX’s capability to rival, and in many instances, surpass, the performance of public-domain traders, including those that outclass human traders, emphasising the efficiency of simple models, as this is required to succeed in intricate multi-threaded simulations. This highlights the potential of leveraging “black-box” Deep Learning systems to create more efficient financial markets.
Complexity vs Empirical Score
- Math Complexity: 4.0/10
- Empirical Rigor: 6.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper uses a deep learning model but does not present heavy mathematical derivations, focusing more on simulation setup and performance results. Empirical rigor is moderately high due to extensive backtesting in a simulated market with statistical validation against benchmarks.
Research Flowchart
flowchart TD A["<b>Research Goal</b><br>Can a simple DL model rival<br>conventional trading strategies?"] --> B["<b>Methodology: DeepTraderX (DTX)</b><br>Deep Learning Trader trained via RL"] B --> C["<b>Input Data</b><br>Level-2 LOB Data<br>Historic Market Schedules"] C --> D["<b>Market Simulation</b><br>500+ Trading Days<br>Multi-threaded Environment"] D --> E["<b>Core Process</b><br>State S<sub>t</sub> → DTX → Quote P<sub>t</sub>"] E --> F["<b>Key Findings</b><br>DTX rivals/surpasses<br>public-domain traders<br>Simple models succeed in<br>complex simulations"]
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 DeepTraderX (DTX), a simple deep learning-based trader that maps market data to bid/ask quotes using a DLNN trained on Level‑2 order‑book snapshots.
Source sentence
In this paper, we introduce DeepTraderX (DTX), a simple Deep Learning-based trader, and present results that demonstrate its performance in a multi-threaded market simulation. Trained on historical Level-2 market data, i.e., the Limit Order Book (LOB) for specific tradable assets, DTX processes the market state $S$ at each timestep $T$ to determine a price $P$ for market orders.
Assumes 6
- Profit per trader (PPT) is the sole metric for evaluating trader performance, as the simulator does not model losses. statistical
Source sentence
The one met-ric we are interested in is the profit that each trader type achieves at the end of the session, namely profit per trader (PPT). We cannot assess these algorithmic traders the same way we would with real traders, as TBSE doesn’t simulate loss, so we are judging based on profits only.
- All available data are used for training the DLNN, without a separate validation or test split. model
Source sentence
3.0.2 Model Architecture and Training Contrary to the usual practices for training and vali-dating a DLNN, which consist of splitting the dataset into training, validation, and test subsets, we used all the dataset for training.
- The market simulation uses stochastic, unique supply‑demand schedules to avoid repetition and bias. market
Source sentence
TBSE was designed to use real-world historical data to introduce variability in its supply and demand schedules, generating more realistic data. The sim-ulator uses a stochastically-altered schedule at each session, avoiding repetition and bias.
- Input features are min‑max normalised to preserve scale of financial variables. data
Source sentence
The choice was to use min-max normalisation, given that we are working with multivariate features derived from fi-nancial data, so it is important to preserve their scale while using an easy-to-understand model.
Show 2 more
- The simulator generates synthetic market data based on real historic stock schedules, but the results are evaluated only on simulated profit outcomes. other
Source sentence
The market data used in both training and testing was generated from unique market schedules based on real historic stock market data. TBSE was designed to use real-world historical data to introduce variability in its supply and demand schedules, generating more realistic data.
- Trading costs (commissions, slippage, market impact) are not incorporated into the reported profit results. execution
Source sentence
4 RESULTS The following section presents the results of our ex-periments, obtained through 4,000 individual market simulations.
Data
- IBM stock price data (August 31, 2017 NYSE trading day) market-prices public
Source sentence
For our train-ing and experiment sessions, we used IBM stock price data from the August 31, 2017 NYSE trading day.
- Synthetic market data generated by TBSE synthetic proprietary
Source sentence
The market data used in both training and testing was generated from unique market schedules based on real historic stock market data. TBSE was designed to use real-world historical data to introduce variability in its supply and demand schedules, generating more realistic data.
Costs
Not modeled
Finds 5
- DTX dominates in 6 out of 8 experiments, showing very significant differences in profit per trader (PPT). performance
Source sentence
The outcome largely supports our re-search hypothesis, with DTX dominating in 6 out of 8 experiments, with very significant differences in PPT for a number of them.
- Wilcoxon signed‑rank tests (95% significance) reject the null hypothesis of no profit difference for most opponent pairings, confirming DTX’s superiority. statistical
Source sentence
Thus, for each time DTX was poised against the other traders, we performed a Wilcoxon-signed rank-test with a significance level of 95%. A p-value lower than 0.05 indicates that we can reject the hypothesis, concluding that one strategy outperforms the other in a given experiment.
- In ZIC vs. DTX experiments, DTX is identified as the dominant strategy with a statistically significant higher mean profit. performance
Source sentence
4.0.1 ZIC vs. DTX Figure 2a shows a narrow difference in means be-tween ZIC and DTX in the BGTs, a fact visible by the dispersion of profits in Figure 3. The statistical test for 95% significance level has confirmed DTX as the dominant strategy of this experiment.
- DTX shows clear dominance over GDX and other traders in box‑plot and scatter‑plot analyses, with most points above the diagonal reference line. performance
Source sentence
-tion, the bar plot shows a significant difference be-tween the means of DTX and its competitor, with the scatter plot placing most of its points above the diag-onal, indicating the clear dominance of DTX in this experiment, with the same outcome confirmed by the outcome of our statistical test. Figures 8b and 10 show PPT score comparisons in the OTM experiment.
- Against AA, DTX’s profit distribution is similar and statistical tests do not show a significant difference, indicating comparable performance. performance
Source sentence
4.0.4 AA vs. DTX Figures 11a and 12 present a visual representation of the profits obtained by AA and DTX in the BGTs, in-dicating similar results for both traders. The statistical test applied failed to prove that there is a significant difference in terms of mean profits between AA and Figure 9: Scatter plot of PPT in BGTs for GDX vs. DTX.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 28 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (46)
- Which Trading Agent is Best ? Using a Threaded Parallel Simulation of a Financial Market Changes the Pecking-Order (2020) — link
- Minimal-Intelligence Agents for Bargaining Behaviors in Market-Based Environments (1997)
- Strategic Sequential Bidding in Auctions Using Dynamic Programming (2002)
- The Structure and Behaviour of the Continuous Double Auction (2006)
- Agent-Based Modeling in Economics and Finance : Past , Present , and Future (2018)
- Artificial Intelligence : Financial Trading and Neurology of Decision (2018)
- A Comparative Study of Bitcoin Price Prediction Using Deep Learning (2019)
- Automated Trading System for Stock Index Using LSTM Neural Networks and Risk Management (2020)
and 38 more
- Language models are few-shot learners (2020)
- An open-source limit-order-book exchange for teaching and research (2018)
- High-Performance Bidding Agents for the Continuous Double Auction (2001)
- Human-agent auction interactions: Adaptive-aggressive agents dominate (2011)
- Adaptive switching circuits (1960)
- Untersuchungen zu dynamischen neuronalen Netzen (1991)
- Long short-term memory (1997)
- Exploring segment representations for neural semi-Markov conditional random fields (2020)
- Supervised Learning (2008)
- The perceptron: A probabilistic model for information storage and organization in the brain. (1958)
- The information content of an open limit-order book (2009)
- High-frequency trading in a limit order book (2008)
- DeepLOB: Deep Convolutional Neural Networks for Limit Order Books (2019)
- Deep Learning in Drug Discovery (2015)
- Algorithmic Trading Market Size Expanding at 12.9% CAGR , Set to Reach USD 41.9 Billion By 2030 (2023)
- A State -of-the- Art Survey on Deep Learning Theory and Architectures (2019)
- Diving Deep into Deep Learning : History, Evolution , Types and Applications (2020)
- Universal Features of Price Formation in Financial Markets : Perspectives from Deep Learning (2019)
- Market Making Strategy Optimization via Deep Reinforcement Learning (2022)
- Exploring Assignment-Adaptive (ASAD) Trading Agents in Financial Market Experiments . (2013)
- Chat GPT (2023)
- Adaptive- Aggressive Traders Don \textquoteright t Dominate (2019)
- Price Formation in Double Auctions (1998)
- Allocative Efficiency of Markets with Zero - Intelligence Traders : Market as a Partial Substitute for Individual Rationality (1993)
- An Experimental Study of Competitive Market Behavior (1962)
- GitHub - BSE is a Simple Minimal Simulation of a Limit Order Book Financial Exchange (2022)
- Automated Creation of a High - Performing Algorithmic Trader via Deep Learning on Level -2 Limit Order Book Data (2020)
- Deep learning for sleep analysis on children with sleep-disordered breathing: Automatic detection of mouth breathing events (2023)
- Deep Learning Techniques For Improving Breast Cancer Detection And Diagnosis (2022)
- AN Approach To Deep Learning (2022)
- Deep Learning can Replicate Adaptive Traders in a Limit - Order - Book Financial Market (2018)
- What Is a Trader , and What Do Traders Do ? (2023)
- Does Fairness Prevent Market Clearing ? An Experimental Investigation (1993)
- Comparisons of various types of normality tests (2011)
- The Crucial Thing a Computer Can \textquoteright t Do (2022)
- A double-blind trial of the analgesic properties of physostigmine in postoperative patients (1986)
- Using field experiments to test equivalence between auction formats: Magic on the Internet (1999)
- Can Intelligence and Speed Make Profit ? A Deep Learning Trader Challenging Existing Strategies in a Multi-Threaded Financial Market Simulation (2023)
0 of 46 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.