Paper: arXiv 2405.08101

Authors: G. Ibikunle, B. Moews, D. Muravyev, K. Rzayev

Abstract

High-frequency trading (HFT) accounts for almost half of equity trading volume, yet it is not identified in public data. We develop novel data-driven measures of HFT activity that separate strategies that supply and demand liquidity. We train machine learning models to predict HFT activity observed in a proprietary dataset using concurrent public intraday data. Once trained on the dataset, these models generate HFT measures for the entire U.S. stock universe from 2010 to 2023. Our measures outperform conventional proxies, which struggle to capture HFT’s time dynamics. We further validate them using shocks to HFT activity, including latency arbitrage, exchange speed bumps, and data feed upgrades. Finally, our measures reveal how HFT affects fundamental information acquisition. Liquidity-supplying HFTs improve price discovery around earnings announcements while liquidity-demanding strategies impede it.

Complexity vs Empirical Score

  • Math Complexity: 3.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper employs advanced machine learning (ensemble models) for prediction but focuses more on application and validation than deep mathematical theory. It demonstrates high empirical rigor through extensive backtesting on multiple datasets, validation via natural experiments (speed bumps, latency arbitrage, feed upgrades), and detailed out-of-sample performance metrics across a long time period and broad stock universe.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Identify & Measure HFT in Public Data"] --> B["Methodology:<br>Machine Learning Model Training"]
  
  B --> C{"Data/Inputs"}
  C --> C1["Proprietary HFT Dataset<br>Ground Truth Labels"]
  C --> C2["Public Intraday Data<br>Features for Prediction"]
  
  C1 & C2 --> D["Computational Process:<br>Train ML Models"]
  D --> E["Generate HFT Measures<br>for Full US Stock Universe<br>(2010-2023)"]
  
  E --> F{"Validation & Outcomes"}
  F --> F1["Superior Performance<br>vs. Conventional Proxies"]
  F --> F2["Validated via Shocks<br>Latency, Speed Bumps, Feeds"]
  F --> F3["Key Finding:<br>Liquidity Supplying HFT<br>Improves Price Discovery"]
  F --> F4["Key Finding:<br>Liquidity Demanding HFT<br>Impedes Price Discovery"]

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 novel data‑driven measures of high‑frequency trading that separate liquidity‑supplying and liquidity‑demanding strategies, using machine‑learning models trained on proprietary data and applied to public intraday data.

Source sentence
We develop novel data-driven measures of HFT activity that separate strategies that supply and demand liquidity.

Assumes 6

  • The mapping learned from the 2009 Nasdaq sample remains informative for other U.S. stocks and later dates. model
    Source sentence
    This step assumes that the Nasdaq mapping remains informative for other U.S. stocks and later dates.
  • HFT activity is endogenous to liquidity and trading characteristics, requiring instrumental variables for identification. statistical
    Source sentence
    HFT activity is endogenous to the liquidity and trading characteristics that may also shape information acquisition.
  • The proprietary Nasdaq HFT dataset accurately identifies aggressive (HH, HN) and passive (HH, NH) trades. data
    Source sentence
    Trades are classified as HH when both parties are HFTs, HN when an HFT demands liquidity from a non-HFT, NH when a non-HFT demands liquidity from an HFT, and NN when neither party is an HFT.
  • Public intraday variables (24 TAQ indicators) contain enough information to predict observed HFT shares. model
    Source sentence
    inputs to predict the observed HFT shares in the Nasdaq sample.
Show 2 more
  • Tree‑based ensemble methods (extremely randomized trees) provide accurate out‑of‑sample predictions of HFT activity. model
    Source sentence
    Extremely randomized trees provide the highest out-of-sample accuracy with low dispersion; therefore, we use them for the main measures.
  • Spread variables can be excluded from the feature set without materially altering the HFT measure predictions. statistical
    Source sentence
    To account for spread variables appearing in the training feature set, we also reconstruct the HFT measures by excluding them; the spread results remain similar.

Data

  • Nasdaq HFT dataset · Nasdaq order-book proprietary
    2009 · 120 stocks
    Source sentence
    The first is the Nasdaq HFT dataset used by Brogaard et al. (2014), which identifies the parties to each trade for 120 stocks in 2009 and reports time to the millisecond, price, volume, trade direction, and counterparty type.
  • WRDS TAQ Intraday Indicators · WRDS market-prices public
    2010‑2023 · U.S. stock universe
    Source sentence
    The second source is WRDS TAQ Intraday Indicators for the same period. We select 24 public variables covering prices, trading volume, trading costs, liquidity, and volatility.
  • LSEG Datascope Tick History · LSEG order-book commercial
    Source sentence
    Conventional HFT proxies are constructed from quote-level Millisecond TAQ data. Intraday transactions and corresponding bid and ask quotes come from LSEG Datascope Tick History.
  • I/B/E/S announcements · I/B/E/S fundamentals commercial
    Source sentence
    Earnings and merger-and-acquisition announcement dates come from I/B/E/S and the Thomson Reuters Securities Data Company database, respectively.
  • CRSP stock returns and volume · CRSP market-prices commercial
    Source sentence
    Stock returns and trading volume come from CRSP.
  • Eurofidai‑Bedofih data · Eurofidai‑Bedofih order-book proprietary
    Euronext Paris stocks
    Source sentence
    Finally, proprietary Eurofidai‑Bedofih data identify HFT activity in Euronext Paris stocks; Section 5 describes that sample.

Recognised sources: NYSE TAQ Refinitiv / LSEG (Datastream, Eikon, Tick History) WRDS CRSP

Costs

Not applicable The paper reports no trading or execution result; the cost the extractor found is a model parameter.

The paper-text extraction found a cost parameter inside a model (Spreads are used as an outcome variable; a one‑standard‑deviation increase in liquidity‑supplying HFT reduces spreads by 28%, while the same increase in liquidity‑demanding HFT raises spreads by 38%. The analysis also re); the classifier found no reported trading result, so costs are not applicable.

Finds 5

  • The ML‑generated HFT measures outperform five conventional public‑data proxies in capturing both cross‑sectional differences and time‑series variation of liquidity‑demanding and liquidity‑supplying HFT. predictive-accuracy
    Source sentence
    the ML predictions better capture both cross-sectional differences and time-series variation in liquidity-demanding and liquidity-supplying HFT.
  • Liquidity‑supplying HFT improves price discovery around earnings announcements, while liquidity‑demanding HFT impedes it. statistical
    Source sentence
    Liquidity-supplying HFTs improve price discovery around earnings announcements while liquidity-demanding strategies impede it.
  • A one‑standard‑deviation increase in HFT demand reduces information acquisition by 18%, whereas the same increase in HFT supply raises information acquisition by 9%. statistical
    Source sentence
    a one-standard-deviation increase in HFT demand corresponds to 18% less information acquisition, while the same increase in HFT supply corresponds to 9% more.
  • During the COVID‑19 crisis, liquidity‑supplying HFT remained resilient and its negative association with spreads strengthened. performance
    Source sentence
    During COVID-19, liquidity-supplying HFT increased and its negative association with spreads strengthened, evidence consistent with resilient passive HFT during prolonged market stress.
  • Public proxies used in prior information‑acquisition literature mainly capture liquidity‑demanding HFT, reconciling earlier findings that aggregate algorithmic trading reduces information acquisition. other
    Source sentence
    We show that Weller’s (2018) public proxies primarily capture HFT demand, helping reconcile his aggregate result with our findings.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 24 related papers). Coverage is partial, so this is neither replication nor refutation.

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 (72)

  1. Ensemble modeling, uncertainty and robust predictions (2013) — link
  2. Classification and Regression Trees. (1984) — link
  3. Pattern Recognition and Machine Learning (2007) — link
  4. Random decision forests (2002) — link
  5. Bringing the Future Forward: The Effect of Disclosure on the Returns‐Earnings Relation (2002) — link
  6. Market microstructure: A survey of microfoundations, empirical results, and policy implications (2005) — link
  7. Inferring investor behavior: Evidence from TORQ data (2000) — link
  8. Extremely randomized trees (2006) — link
and 64 more
  1. High Frequency Trading and the New-Market Makers (2012) — link
  2. Inferring Trade Direction from Intraday Data (1991) — link
  3. Presidential Address: Liquidity and Price Discovery (2003) — link
  4. How Much New Information Is There in Earnings? (2008) — link
  5. High Frequency Quoting, Trading, and the Efficiency of Prices (2013) — link
  6. Information Leakage and Market Efficiency (2005) — link
  7. The High-Frequency Trading Arms Race: Frequent Batch Auctions as a Market Design Response (2013) — link
  8. Does Algorithmic Trading Improve Liquidity? (2011) — link
  9. Top 10 algorithms in data mining (2007) — link
  10. The Impact of SFAS No. 131 Business Segment Data on the Market’s Ability to Anticipate Future Earnings (2005) — link
  11. Induction of Decision Trees (1986) — link
  12. The data complexity index to construct an efficient cross-validation method (2010) — link
  13. Random Forests for Big Data (2017) — link
  14. A random forest guided tour (2016) — link
  15. High-Frequency Trading and Price Discovery (2014) — link
  16. The Economics of High-Frequency Trading: Taking Stock (2016) — link
  17. Where are the risks in high frequency trading (2016)
  18. Who provides liquidity, and when? (2021) — link
  19. To tune or not to tune the number of trees in random forest? (2017) — link
  20. On the Learnability of Fully-Connected Neural Networks (2017)
  21. Tapping the Brakes: Are Less Active Markets Safer and Better for the Economy? (2014) — link
  22. The Elements of Statistical Learning (2009) — link
  23. Need for Speed? Exchange Latency and Liquidity (2017) — link
  24. High Frequency Trading and Co-Movement in Financial Markets (2018) — link
  25. Equilibrium fast trading (2015) — link
  26. High frequency trading and comovement in financial markets (2019) — link
  27. A state-space modeling of the information content of trading volume (2019) — link
  28. High-frequency traders and price informativeness during earnings announcements (2020) — link
  29. High-Frequency Trading around Large Institutional Orders (2015) — link
  30. Does Algorithmic Trading Reduce Information Acquisition? (2017) — link
  31. Trading Fast and Slow: Colocation and Liquidity (2015) — link
  32. Algorithmic Trading and Market Quality: International Evidence (2020) — link
  33. Do speed bumps curb low-latency investment? Evidence from a laboratory market (2020) — link
  34. High frequency trading and extreme price movements (2018) — link
  35. The Diversity of High Frequency Traders (2012) — link
  36. High-Frequency Trading and Market Performance (2015) — link
  37. Clean Sweep: Informed Trading through Intermarket Sweep Orders (2012) — link
  38. Trading aggressiveness and market efficiency (2019) — link
  39. Every Cloud Has a Silver Lining: Fast Trading, Microwave Connectivity and Trading Costs (2016) — link
  40. Market (in)attention and the strategic scheduling and timing of earnings announcements (2015) — link
  41. High Frequency Quoting: Short-Term Volatility in Bids and Offers (2013) — link
  42. Do High-Frequency Traders Anticipate Buying and Selling Pressure? (2020) — link
  43. The Competitive Landscape of High-Frequency Trading Firms (2017) — link
  44. Back-Running: Seeking and Hiding Fundamental Information in Order Flows* (2019) — link
  45. Microstructure in the Machine Age (2020) — link
  46. The market quality implications of speed in cross-platform trading: Evidence from Frankfurt-London microwave networks (2023) — link
  47. Hybrid analytic and machine-learned baryonic property insertion into galactic dark matter haloes (2021) — link
  48. Tracking Retail Investor Activity (2017) — link
  49. Attention: How high-frequency trading improves price efficiency following earnings announcements (2021) — link
  50. Quantifying the High-Frequency Trading “Arms Race” (2021) — link
  51. AI in Finance: Challenges, Techniques and Opportunities (2021) — link
  52. AI in Finance: Challenges, Techniques, and Opportunities (2022) — link
  53. Dark Pool Trading and Information Acquisition (2021) — link
  54. The Conduits of Price Discovery: A Machine Learning Approach (2020) — link
  55. Random Decision Forests (2010) — link
  56. Toxic Arbitrage (2016) — link
  57. The Effect of Tick Size on Managerial Learning from Stock Prices (2020) — link
  58. High-Frequency Trading Strategies (2017) — link
  59. High frequency market making: The role of speed (2023) — link
  60. Algorithmic Trading and Investment-to-Price Sensitivity (2023) — link
  61. Identifying High Frequency Trading Activity without Proprietary Data (2023) — link
  62. Sharks in the dark: Quantifying HFT dark pool latency arbitrage (2023) — link
  63. Informed Trading Intensity (2021) — link
  64. High-frequency trading in the stock market and the costs of options market making (2024) — link

0 of 72 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.