Korniejczuk & Slepaczuk (2024) — Multi-Pair Stat Arb via Graph Clustering
Authors: Adrian Korniejczuk, Robert Slepaczuk.
Venue: arXiv 2406.10695, dated 15 June 2024.
Contribution
Extended Cartea Cucuringu Jin 2023 SPONGE's SPONGEsym framework with two practical engineering changes:
- Lengthened rebalance from 3 to 10 days to reduce turnover and transaction-cost drag.
- ML ensemble signal filter: 5 classifiers — HistGradientBoosting, AdaBoost, MLP, Logistic Regression, SGD — trained on binary profitability labels derived from a backtest of the baseline clustering strategy. Soft-voting aggregation weighted by Brier score (not accuracy). Trades taken only on the top 10% of signals by predicted profitability → average portfolio size ~45 stocks.
Two profitability definitions are used: (a) threshold-based (cumulative return exceeds 4% at any point during the 10-day holding period), and (b) transaction-cost-adjusted (return exceeds transaction cost at next rebalance, even without hitting the threshold). A trade is profitable if either condition is met.
Replication concern
The headline "large gains from ML filtering" was not reproducible by Fanshawe Masih Cameron 2025 THGNN Correlation Forecasting's authors. Classifier Brier scores and precision match the original paper, suggesting the discrepancy is in portfolio construction or backtest plumbing rather than the ML implementation itself. This is the classic StatArb backtest fragility.
Importance to this wiki
- It is the immediate antecedent of the THGNN paper — they reuse the trading engine and improve only the dependence metric (rolling → forward-predicted correlations).
- The 4% threshold / transaction-cost-adjusted label definition is a reusable labelling recipe for any binary profitability classifier on a holding-period strategy.