Skip to content

Deep Limit Order Book Forecasting

Authors: Antonio Briola, Silvia Bartolucci, Tomaso Aste Institution: Department of Computer Science, University College London; Systemic Risk Centre, LSE Year: 2024 (arXiv March 2024) arXiv: 2403.09267 Categories: q-fin.TR, cs.LG

Plain-language abstract

Deep learning models can forecast short-horizon LOB mid-price direction with high accuracy — but does high accuracy translate to profitable trading? This paper releases LOBFrame, an open-source framework for processing large-scale LOB data and benchmarking deep learning models on 15 NASDAQ stocks (2017–2019). Two headline findings: (i) stocks' microstructural characteristics — specifically whether they are small-, medium-, or large-tick — are the dominant determinant of predictability, not model choice; (ii) high ML accuracy does not reliably correspond to actionable trading signals. The authors propose an operational metric based on the probability of correctly forecasting complete round-trip transactions rather than pure directional classification.

Key contributions

  1. LOBFrame — open-source Python/PyTorch codebase (github.com/FinancialComputingUCL/LOBFrame) for scalable LOB data processing and systematic DL model evaluation on LOBSTER NASDAQ data.
  2. Tick-size-based stock classification — practical thresholds on the ratio \(\langle\sigma\rangle/\theta\) (mean spread / tick size) dividing stocks into small-tick (\(\gtrsim 3\theta\)), medium-tick (\(1.5\theta\)–\(3\theta\)), and large-tick (\(\lesssim 1.5\theta\)) regimes.
  3. Microstructure → predictability mapping — empirically links predictability to spread, depth at bests, volume distribution (CCDF), and "information richness" metrics.
  4. Accuracy vs tradability gap — documents that high directional accuracy frequently fails to survive execution costs and slippage.
  5. Operational metric — evaluates models by probability of correctly predicting a complete transaction in the forecast direction.

Method summary

Data

  • 15 NASDAQ stocks (2017–2019), tick-by-tick L2 data from LOBSTER.
  • 10 LOB levels per snapshot → 40-dim features (4 per level: bid/ask × price/size).
  • Per year: 45 training days / 5 (non-consecutive) validation days / 10 test days.
  • 5-day rolling z-score normalisation (not global) to prevent data leakage under non-stationarity.
  • Trading hours filtered to 09:40–15:50 ET (first and last 10 minutes excluded).

Stock groups

Group Tick-size regime Stocks (6/3/6)
1 — small-tick \(\langle\sigma\rangle \gtrsim 3\theta\) CHTR, GOOG, GS, IBM, MCD, NVDA
2 — medium-tick \(1.5\theta \lesssim \langle\sigma\rangle \lesssim 3\theta\) AAPL, ABBV, PM
3 — large-tick \(\langle\sigma\rangle \lesssim 1.5\theta\) BAC, CSCO, KO, ORCL, PFE, VZ

Model: DeepLOB (Zhang et al. 2019)

The only architecture evaluated. CNN → Inception → LSTM → Dense over a \(100\times40\) input (100 consecutive LOB updates × 40 spatial features): - Convolutional blocks capture inter-level spatial structure and very-short-time dynamics. - Inception module widens the receptive field across time scales. - LSTM@64 units captures residual temporal dependencies. - Dense@3 units outputs the \(\{\text{Down, Stable, Up}\}\) class.

Labels & horizons

Three horizons \(H_{\Delta\tau} \in \{10, 50, 100\}\) measured in LOB updates (not physical time). A sample is labelled Up / Down if mid-price moves by at least the tick size \(\theta\) over the horizon; otherwise Stable.

Training

  • Optimiser: AdamW (weight decay), lr \(6\times10^{-5}\), \(\beta_1=0.9\), \(\beta_2=0.95\).
  • Batch size 32. Max 100 epochs, patience 15.
  • Balanced class sampling in training (up to 5 000 samples per class per day).
  • Sequential sampling in validation/test.
  • 135 experiments run on UCL HPC cluster, ~959 GPU-hours total on P100/V100/Titan-class GPUs.

Main results

  • Predictability clusters by tick-size group. Large-tick stocks are consistently most forecastable: tighter spreads, deeper and broader volume CCDFs, and a characteristic queue-decrease-before-transaction pattern that leaks directional information.
  • Small-tick stocks are the hardest: broader spread distributions, thinner queues at best, and higher volatility dilute the DL signal.
  • Medium-tick stocks are heterogeneous: AAPL leans large-tick-like, while ABBV/PM lean small-tick-like — the "borderline" label is real, not nominal.
  • Class-balance drift across horizons: at \(H_{10}\) the "Stable" class dominates for large-tick stocks and "Up/Down" dominate for small-tick stocks; at \(H_{50}\)/\(H_{100}\) all classes even out. This asymmetry predicts which horizons are usable per stock.
  • Accuracy ≠ tradability: even when directional accuracy is high, the probability of converting a prediction into a completed round-trip transaction is substantially lower. Traditional ML metrics are misleading for execution-aware decisions.

Limitations

  • Single model architecture: only DeepLOB is evaluated in detail. Transformer-based LOB models (e.g., TransLOB) are reviewed in the literature section but not benchmarked, so comparative claims about architecture are out of scope.
  • NASDAQ-only: all 15 stocks from one exchange; generalisation to other venues is untested.
  • Large/mega-cap only: 10B–1T+ capitalisation bucket — mid- and small-cap behaviour may differ.
  • Simulation-to-reality gap partially closed: the operational metric is an improvement, but the paper explicitly does not run full trading simulations (Section 7.2).
  • Mid-price differences, not log-returns: choice made for tick-size-aware control; complicates comparisons to papers that use log-return targets.

Connections