Paper: arXiv 2312.05827
Authors: Álvaro Cartea, Gerardo Duran-Martin, Leandro Sánchez-Betancourt
Abstract
This paper develops a framework to predict toxic trades that a broker receives from her clients. Toxic trades are predicted with a novel online learning Bayesian method which we call the projection-based unification of last-layer and subspace estimation (PULSE). PULSE is a fast and statistically-efficient Bayesian procedure for online training of neural networks. We employ a proprietary dataset of foreign exchange transactions to test our methodology. Neural networks trained with PULSE outperform standard machine learning and statistical methods when predicting if a trade will be toxic; the benchmark methods are logistic regression, random forests, and a recursively-updated maximum-likelihood estimator. We devise a strategy for the broker who uses toxicity predictions to internalise or to externalise each trade received from her clients. Our methodology can be implemented in real-time because it takes less than one millisecond to update parameters and make a prediction. Compared with the benchmarks, online learning of a neural network with PULSE attains the highest PnL and avoids the most losses by externalising toxic trades.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 9.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper introduces a complex novel Bayesian online learning method (PULSE) with significant mathematical foundations in variational approximations and Kalman filtering, yet is grounded in a proprietary high-frequency FX dataset with real-time execution speed metrics and reported PnL.
Research Flowchart
flowchart TD
A["Research Goal: Predict Toxic Trades in FX"] --> B{"PULSE Methodology"}
B --> C["Online Bayesian NN Training"]
C --> D["Foreign Exchange Proprietary Data"]
D --> E["Real-time Prediction < 1ms"]
E --> F{"Benchmark Comparison"}
F --> G["Superior PnL & Loss Avoidance"]
F --> H["Outperforms LogReg, RF, MLE"]
G & H --> I["Broker Strategy: Internalize vs Externalize"]
style A fill:#f9f,stroke:#333,stroke-width:2px
style I fill:#bbf,stroke:#333,stroke-width:2px
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 novel online Bayesian learning method called PULSE for training neural networks and a broker internalise‑externalise strategy based on toxicity predictions.
Source sentence
a novel online learning Bayesian method which we call the projection‑based unification of last‑layer and subspace estimation (PULSE).
Assumes 5
- The broker quotes her clients the best available rates in LMAX Exchange net of fees. execution
Source sentence
we assume that the broker quotes her clients the best available rates in LMAX Exchange net of fees.
- Transaction costs in FX are around $3 per million euros traded, providing clients a discount of $3 per million euros. market
Source sentence
Transaction costs in FX are around $3 per million euros traded (see e.g., Cartea and Sánchez‑Betancourt (2023)), so this assumption provides clients with a discount of $3 per million euros traded when trading with the broker.
- Permanent price impact of externalising trades is ignored in the strategy evaluation. model
Source sentence
we ignore the permanent price impact of externalising trades, as this would require the formulation of a stochastic control problem.
- A trade is toxic if the client can unwind it at a profit within a given time window. other
Source sentence
In this paper, a trade is toxic over a given time window if the client can unwind the trade at a profit within the time window.
Show 1 more
- The broker’s decision rule uses a cutoff probability p to internalise or externalise a trade. model
Source sentence
For a given toxicity horizon and a cutoff probability, the strategy internalises the trade if the probability that the trade is toxic is less than or equal to the cutoff probability, otherwise it externalises the trade.
Data
- Proprietary FX dataset (EUR/USD) market-prices proprietary
Source sentence
We employ data for the currency pair EUR/USD from LMAX Broker and from LMAX Exchange for the period 28 June 2022 to 21 October 2022.
Costs
Partial
Assumes a flat $3 per million euros traded cost, providing a discount to clients; costs are not incorporated into the reported PnL results.
Source sentence
Transaction costs in FX are around $3 per million euros traded (see e.g., Cartea and Sánchez‑Betancourt (2023)), so this assumption provides clients with a discount of $3 per million euros traded when trading with the broker.
Finds 3
- PULSE achieves the highest average AUC among all models for every toxicity horizon and outperforms benchmarks in profitability and loss avoidance. performance
Source sentence
... the mean PnL of the internalise‑externalise strategy we obtain when training the MLP with PULSE is the highest (when compared with the benchmarks).
- A universal model trained on all traders yields higher toxicity prediction accuracy than models trained per individual trader. predictive-accuracy
Source sentence
We obtain higher accuracies (when predicting toxic trades) when we train one model for all traders than when we train one model per trader; higher accuracies result from having more data.
- Including client‑specific variables (identity and recent activity) improves PULSE performance substantially compared to a model without them. performance
Source sentence
the performance of PULSE deteriorates substantially when compared to the performance of a model that includes the identity and unique features of the trader.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 12 related papers). Coverage is partial, so this is neither replication nor refutation.
7 papers 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.
Cited References (45)
- Best practice in statistics: The use of log transformation (2022)
- When do neural nets outperform boosted trees on tabular data? (2023)
- Market-making by a foreign exchange dealer (2022)
- A mean field game between informed traders and a broker — → 2401.05257
- Algorithmic market making in dealer markets with hedging and market impact (2023)
- Strategic learning and trading in broker-mediated markets — → 2412.20847
- Broker-Trader Partial Information Nash-Equilibria — → 2412.17712
- Liquidity competition between brokers and an informed trader — → 2503.08287
and 37 more
- Nash equilibrium between brokers and traders — → 2407.10561
- An introduction to ROC analysis (2006) — link
- Optimal execution with stochastic delay (2023)
- Pattern Recognition and Machine Learning (Information Science and Statistics) (2006)
- The Only Game in Town (1971) — link
- Liquidity, information, and infrequently traded stocks (1996)
- The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (2019) — link
- Measuring the Intrinsic Dimension of Objective Landscapes (2018) — link
- Information Effects on the Bid-Ask Spread (1983)
- Execution in an aggregator (2017) — link
- Bid, ask and transaction prices in a specialist market with heterogeneously informed traders (1985)
- Dealership market: Market-making with inventory (1980)
- The recursive variational G aussian approximation ( R-VGA ) (2021) — link
- ASML asset description (2023)
- Bayesian Filtering and Smoothing (2013) — link
- Training Multilayer Perceptrons with the Extended Kalman Algorithm (1988) — link
- Online Natural Gradient as a K alman Filter (2017) — link
- Efficient Online Bayesian Inference for Neural Bandits (2022)
- Probabilistic Machine Learning: An Introduction (2022) — link
- Probabilistic Machine Learning: Advanced Topics (2023) — link
- Kalman Filters (2001)
- How many degrees of freedom do we need to train deep networks: a loss landscape perspective (2021) — link
- Adam: A Method for Stochastic Optimization (2015)
- The Elements of Statistical Learning (2001)
- On the limited memory BFGS method for large scale optimization (1989) — link
- Stein’s Lemma for the Reparameterization Trick with Exponential Family Mixtures (2019) — link
- How and When are High-Frequency Stock Returns Predictable? (2022)
- Brokers and informed traders: dealing with toxic flow and extracting trading signals (2025)
- On the impossibility of informationally efficient markets (1980)
- Continuous auctions and insider trading (1985)
- Internalisation by electronic FX spot dealers (2019)
- Insider trading with temporary price impact (2021)
- A mean-field game of market-making against strategic traders (2022) — link
- Informed speculation with imperfect competition (1989)
- Insider trading in an equilibrium model with default: a passage from reduced-form to structural modelling (2007)
- Deep attentive survival analysis in limit order books: Estimating fill probabilities with convolutional-transformers — → 2306.05479
- Bandits for Algorithmic Trading with Signals (2023)
6 of 45 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.