Paper: arXiv 2511.07434
Authors: Enzo Duflot, Stanislas Robineau
Abstract
We study opportunistic optimal liquidation over fixed deadlines on BTC-USD limit-order books (LOB). We present RL-Exec, a PPO agent trained on historical replays augmented with endogenous transient impact (resilience), partial fills, maker/taker fees, and latency. The policy observes depth-20 LOB features plus microstructure indicators and acts under a sell-only inventory constraint to reach a residual target. Evaluation follows a strict time split (train: Jan-2020; test: Feb-2020) and a per-day protocol: for each test day we run ten independent start times and aggregate to a single daily score, avoiding pseudo-replication. We compare the agent to (i) TWAP and (ii) a VWAP-like baseline allocating using opposite-side order-book liquidity (top-20 levels), both executed on identical timestamps and costs. Statistical inference uses one-sided Wilcoxon signed-rank tests on daily RL-baseline differences with Benjamini-Hochberg FDR correction and bootstrap confidence intervals. On the Feb-2020 test set, RL-Exec significantly outperforms both baselines and the gap increases with the execution horizon (+2-3 bps at 30 min, +7-8 bps at 60 min, +23 bps at 120 min). Code: github.com/Giafferri/RL-Exec
Complexity vs Empirical Score
- Math Complexity: 6.0/10
- Empirical Rigor: 8.5/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced reinforcement learning (PPO) and impact models from optimal execution theory (e.g., Almgren-Chriss, Obizhaeva-Wang), indicating moderate-to-high mathematical density. It is highly empirically rigorous with real BTC-USD LOB data, backtested on historical replays with explicit transaction costs and impact, and uses robust statistical inference (Wilcoxon tests, FDR correction, bootstrapping).
Research Flowchart
flowchart TD
A["Research Goal: Find optimal liquidation strategy<br>for BTC-USD LOBs under fixed deadlines"] --> B["Data & Simulation Engine"]
B["Data & Simulation Engine"] --> C["Historical BTC-USD LOB Replays<br>Jan 2020 (Train) | Feb 2020 (Test)"]
B --> D["Realistic Execution Environment<br>Includes: Partial fills, fees, latency, transient impact"]
C & D --> E["Model Architecture"]
subgraph E ["Computational Process"]
E1["PPO Agent"] --> E2["Observation: Depth-20 LOB + Microstructure"]
E3["Action: Sell-only inventory control"] --> E4["Output: Execution Schedule"]
end
E --> F["Comparison & Inference"]
subgraph F ["Evaluation"]
F1["RL-Exec"]
F2["TWAP Baseline"]
F3["VWAP-like Baseline<br>(Order-book Liquidity Alloc.)"]
end
F --> G["Key Findings / Outcomes"]
subgraph G ["Feb 2020 Test Set"]
G1["RL-Exec significantly outperforms TWAP & VWAP"]
G2["Performance gap widens with horizon"]
G3["Stats: 1-sided Wilcoxon + Benjamini-Hochberg FDR"]
end
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 RL-Exec, a PPO reinforcement‑learning agent that learns impact‑aware opportunistic liquidation policies using depth‑20 LOB features and microstructure indicators.
Source sentence
We present RL-Exec, a PPO agent trained on historical replays augmented with endogenous transient impact (resilience), partial fills, maker/taker fees, and latency.
Assumes 6
- The liquidation task is sell‑only with a fixed residual target inventory. execution
Source sentence
acts under a sell-only inventory constraint to reach a residual target.
- Evaluation is performed on a single asset (BTC‑USD) and a single venue. market
Source sentence
Our study is intentionally narrow to make inference clean, but that comes with caveats. First, the evaluation is single‑asset (BTC‑USD), single‑venue, sell‑only liquidation...
- Impact is modeled as a parametric transient (resilience) process in a replay‑with‑impact simulator. model
Source sentence
augmented with endogenous transient impact (resilience), partial fills, maker/taker fees and latency.
- Maker/taker fees and latency are fixed and known. execution
Source sentence
we assume fixed maker/taker fees and latency, and a constrained action space consistent with the sell‑only task.
Show 2 more
- Training and testing periods are strictly time‑split (Jan‑2020 train, Feb‑2020 test). statistical
Source sentence
Training covers January 2020 (31 days) and testing covers February 2020 (28 days).
- Baseline schedules (TWAP and VWAP‑like) are executed on exactly the same timestamps and cost model as the RL agent. other
Source sentence
proxy are executed on exactly the same timestamps and under the same fee and transient‑impact/resilience settings as the agent, ensuring cost and path parity.
Data
2020-01 – 2020-02 · BTC-USD · crypto · intraday
- BTC‑USD limit‑order‑book data order-book unknown
Source sentence
We use BTC‑USD limit‑order‑book data at 1 s resolution with 20 levels per side. Training covers January 2020 (31 days) and testing covers February 2020 (28 days).
Costs
Modeled
Costs include maker/taker fees, latency, and a parametric transient impact (resilience) model; both RL‑Exec and baselines use the same cost model.
Source sentence
augmented with endogenous transient impact (resilience), partial fills, maker/taker fees and latency.
Finds 3
- RL‑Exec outperforms TWAP and the VWAP‑like proxy on all horizons, with mean daily gaps increasing from +2.25 bps at 1,800 s to +22.96 bps at 7,200 s. performance
Source sentence
RL‑Exec beats both baselines on every horizon, with one‑sided Wilcoxon tests (paired by day) significant after Benjamini–Hochberg correction and 95% bootstrap CIs strictly positive:
- The advantage of RL‑Exec grows monotonically with execution horizon (≈+2–3 bps at 30 min, ≈+7–8 bps at 60 min, ≈+23 bps at 120 min). performance
Source sentence
The advantage monotonically increases with the horizon (≈+2–3 bps at 30 min →≈+8 bps at 60 min →≈+23 bps at 120 min).
- A +2–3 bps improvement corresponds to roughly $200–$300 reduction in implementation shortfall per $1 M sell program, after accounting for spreads, fees, latency, and transient impact. other
Source sentence
A +2–3 bps improvement over a $1,000,000 sell program corresponds to roughly $200–$300 of implementation‑shortfall reduction per program, with spreads, fees, latency, and transient impact already included in our cost model.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 20 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.