{"paper_id":"2307.04863","title":"Interpretable ML for High-Frequency Execution","authors":["Timothée Fabre","Vincent Ragel"],"paper_date":"2023-07-10","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2307.04863v2","page_url":"https://thequant.space/flowcharts/interpretable-ml-for-high-frequency-execution/","scores":{"empirical_rigor":8,"math_complexity":6.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes a simple neural network model to infer the fill probability function and a new backtest methodology for high‑frequency order placement that accounts for market impact.","quote":"We set an optimal execution problem where the agent needs to choose between posting a limit order and sending a marketable order. We propose a new backtest methodology that, by essence, takes into account the market impact of limit order insertion for better performance assessment.","summary":"Order placement tactics play a crucial role in high-frequency trading algorithms and their design is based on understanding the dynamics of the order book. Using high quality high-frequency data and a set of microstructural features, we exhibit strong state dependence properties of the fill…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"Latency is negligible compared to the execution horizon.","type":"execution","stated":"explicit","quote":"We suppose in the rest of the discussion that ℓ is negligible compared to the time horizon T, for example ℓ/T < 10⁻³."},{"assumption":"Order lifetimes are independent of censoring variables.","type":"statistical","stated":"explicit","quote":"we assume that the order can be subject to independent right censoring modeled by a random variable Cn and the censoring variable is assumed independent from Ln."},{"assumption":"Fill probability is a smooth non‑linear function of market features.","type":"model","stated":"explicit","quote":"Section 4 discusses two survival analysis methods to compute fill probability on level 3 data and provides empirical evidence of smooth non‑linearity of the fill probability as a function of market features."},{"assumption":"Weighting of censored observations improves fill probability estimation.","type":"model","stated":"explicit","quote":"The significant presence of right censoring in the data set leads us to consider a loss weighting methodology"},{"assumption":"Tick size remains very small on Coinbase during the sample period.","type":"market","stated":"explicit","quote":"We chose to use Coinbase data for two reasons. First, the tick size remains very small during the covered period of time which is not the case for Bitstamp…"},{"assumption":"Feature importance can be assessed with Shapley values on out‑of‑sample observations.","type":"model","stated":"explicit","quote":"We analyze the importance of features with Shapley values computed over out‑of‑sample observations using the SHAP library (Lundberg and Lee, 2017)."}]},"datasets":{"status":"available","items":[{"name":"Coinbase order‑book data","provider":"SUN ZU Lab’s proprietary feed handlers (Coinbase)","kind":"order-book","access":"proprietary","period":"2022‑11‑05 to 2022‑12‑05","universe":"BTC‑USD, ETH‑USD","quote":"We have 1 month of data at our disposal, from 2022‑11‑05 to 2022‑12‑05 on BTC‑USD and ETH‑USD.","known_id":"coinbase"},{"name":"BEDOFIH equity data","provider":"Base Européenne de Données Financières à Haute‑fréquence (BE","kind":"order-book","access":"proprietary","period":"2017‑01‑01 to 2017‑08‑31 (training) and ","universe":"BNPP (Euronext equity)","quote":"August 31st, 2017— and validate it on 1.5 months —from September 1st, 2017 to October 15th, 2017."}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"l2-book"},"transaction_costs":{"status":"available","modeled":"yes","headline_net_of_costs":null,"detail":"Coinbase fee policy ranging from 0.006 to 0.0005 (taker) and 0.004 to 0 (maker) depending on monthly turnover, used in the execution algorithm.","quote":"Table 3: Trading fees — Coinbase fee policy of spot trading on April 2023. Level","source":"front_matter","rule":"R3","candidates":{"front_matter":"yes","evidence_card":"yes","evidence_card_quoted":true},"card_detail":"Coinbase fee policy ranging from 0.006 to 0.0005 (taker) and 0.004 to 0 (maker) depending on monthly turnover, used in the execution algorithm.","card_quote":"Table 3: Trading fees — Coinbase fee policy of spot trading on April 2023. Level","meaning":"costs are part of the reported headline results"},"results":{"status":"available","evaluation_type":"unclear","findings":[{"finding":"Using handcrafted features and a state‑dependent market risk V improves the F‑score from 0.34 to 0.51 for BTC‑USD.","kind":"performance","quote":"It demonstrates that a simple and interpretable neural network architecture"},{"finding":"The optimal placement distance scales linearly with the expected adverse price move V and inversely with the fill‑probability decay rate.","kind":"theoretical","quote":"Equation 21 indicates that in this simple framework, the optimal distance should scale linearly with respect to the expected adverse price move V, and inversely with respect to the decay rate of the fill probability."},{"finding":"Higher taker‑maker fee gaps push the algorithm to post more aggressive (inside‑spread) orders.","kind":"performance","quote":"The algorithm tends to post aggressive orders when the taker‑maker fee gap ε⁻‑ε⁺ is large, pushing the optimal distance towards −ψ as this gap increases."},{"finding":"Applying IPCW weighting deforms the fill probability similarly to the Kaplan‑Meier estimator, giving extra weight to highly censored configurations.","kind":"statistical","quote":"The IPCW procedure deforms the fill probability in a similar way that the Kaplan‑Meier function does by giving extra weights to orders that were not executed under the horizon or posted in highly censored configurations."},{"finding":"Feature importance differs between small‑tick crypto pairs and Euronext equities, especially regarding placement inside the spread versus deep in the book.","kind":"predictive-accuracy","quote":"our work is the first to analyze the feature importance of the fill probability depending on the placement of the order, i.e. if the order is placed inside the spread, at the current best queue, or deep in the book."}],"reported_performance":null,"out_of_sample":{"method":"unclear","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":22,"citations":{"cited_by_in_corpus":0,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":1,"extracted_at":"2026-11-14T18:24:58Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}