Deep Learning Statistical Arbitrage
Jorge Guijarro-Ordonez et al. (3 authors), 2025 · Management Science
What the paper asks for.
Construct daily arbitrage portfolios as IPCA-style residuals against conditional latent factors built from firm characteristics, feed a 30-day cumulative-residual window through a CNN+Transformer to extract a per-stock signal, then map signals to long-short allocation weights with an FFN allocator trained jointly to maximize risk-adjusted returns under a unit L1-leverage constraint; daily rebalanced.
- Type
- ml
- Implementation
- torch
- Rebalance
- daily
- Long-short
- yes
- Data
- One-command fetcher for WRDS-sourced task data
- Operational scope
- Use a reduced liquid-stock universe, executable IPCA-style residualization, and an initial-period fitted model within the benchmark compute envelope.
- 01 Data loading
- 02 Feature engineering
- 03 Model fitting
- 04 Signal generation
- 05 Portfolio construction
- 06 Execution
15 properties · 80 attempts.
Each row is one requirement the strategy must satisfy. The bar reports the share of observed checker verdicts that pass; missing checker output is not converted into a property verdict.
- A2no global normalizestage · feature eng44/76
- A3no future data accessstage · signal gen53/76
- A4causal consistencystage · execution59/76
- A5no survivorship biasstage · data loading70/76
- A6train test temporal orderstage · model fit44/76
- A9signal delay teststage · execution64/76
- A10e2e metric consistency3/76
- B1lookback window l 30§§III.D / §IV “The trading signal is based on a local lookback window of L = 30 days.”provenance · paper-explicitstage · feature eng27/76
- B2weight normalization l1 unit§§III “the absolute values of the individual stock portfolio weights sums up to one, i.e. we use the normalization ||ω_t||_1 = 1.”provenance · paper-explicitstage · portfolio40/76
- B3rebalance frequency daily§§III / §IV “outputs the predicted optimal allocation weight for that residual for the next time ... Our main models estimate arbitrage strategies”provenance · paper-explicitstage · execution44/76
- B4oos period 2002 2016§§IV “out-of-sample trading performance from January 2002 to December 2016 as we use the first four years to estimate the signal and allocation function.”provenance · paper-explicitstage · data loading4/76
- B5universe top liquid us§§IV.A “roughly the 550 largest and most liquid stocks in the U.S. from 1998 to 2016”provenance · benchmark-specifiedstage · data loading69/76
- B6ffn allocation after signal§§III.E “We use a flexible feedforward neural network for the allocation function”provenance · paper-explicitstage · model fit64/76
- B7input is cumulative residuals§§III.D “The input to the signal extraction functions are the last L cumulative residuals.”provenance · paper-explicitstage · feature eng31/76
- B8weights continuous not discretized§§III.C “The allocation function w^ε can be a complex nonlinear function ... outputs the predicted optimal allocation weight”provenance · paper-explicitstage · portfolio50/76
The lowest-passing observed requirement is e2e metric consistency: 3 of 76 observed checker verdicts pass. The full property list above shows whether errors concentrate in domain-level validity or the paper-specific strategy logic.