Statistical Arbitrage (StatArb)

统计套利：基于历史/预测的统计关系（相关性、协整、距离）构造多空对冲头寸，押注均值回归，追求与市场无关的绝对收益。

Definition

Statistical arbitrage (StatArb) is the class of equity-market strategies that:

Group stocks into pairs, baskets, or clusters using a dependence metric (correlation, cointegration, Euclidean distance, or a learned similarity).
Form market-neutral long/short positions that bet on mean reversion of relative prices within those groups.
Rebalance on a fixed cadence (3-day, 10-day, monthly) to harvest the reversion.

The dependence metric is the soul of the strategy: it defines the trade. The standard implementation uses rolling-window correlations or cointegration, which is precisely the bottleneck Fanshawe Masih Cameron 2025 THGNN Correlation Forecasting removes by predicting the dependence matrix forward in time.

Historical evolution

Era	Dependence metric	Limitation
Early (Gatev et al. 2006)	Euclidean distance on prices	Pure price distance; sensitive to drift
Distance / cointegration (Rad et al. 2016)	Cointegration, copula	Cointegration tests have low power on short samples
Correlation clustering (Chen et al. 2019)	Rolling correlation	Backward-looking; lag in regime change
Sector classifications	Fama-French 48 industries	Static; useless in crises
Graph clustering (Mantegna 1999; Tumminello 2005; Cartea 2023)	MST / PMFG / SPONGEsym	Better structure; still backward-looking
ML-filtered SPONGEsym (Korniejczuk & Slepaczuk 2024)	Rolling correlation + 5-classifier soft-voting filter	Reduces turnover; correlations still lag
Forward-looking SPONGEsym (Fanshawe et al. 2025)	THGNN-predicted 10-day correlations	Forward-looking; adapts to regimes

Standard architecture of a modern StatArb strategy

A modern (post-2023) implementation has four layers:

flowchart TB
    A[Universe: S&P 500 or similar large-cap] --> B[Dependence metric: correlation / cointegration / GNN-predicted]
    B --> C[Clustering: SPONGEsym / spectral / hierarchical]
    C --> D[Cluster-specific long-short positions]
    D --> E[Signal filter: ML ensemble / vol target / risk parity]
    E --> F[Execution with TCA, impact model]

The dependence-metric layer is what Fanshawe Masih Cameron 2025 THGNN Correlation Forecasting upgrades. Everything downstream (clustering, position construction, signal filtering) is reusable.

Position construction — equal-weight contrarian

Inside each cluster c with n_c stocks:

Compute the cluster-mean cumulative return over the last H days (5 in the THGNN paper).
For each stock i \in c:
- Long weight = +1/n_c if i's H-day return is below cluster mean.
- Short weight = -1/n_c if i's H-day return is above cluster mean.

This is contrarian within cluster: buy the under-performer, short the over-performer, expecting reversion to the cluster mean. The bet only works if the cluster structure is real and stable enough that reversion actually happens.

The regime-shift problem

StatArb strategies earn their return from stable dependence structure — when correlations are predictable and the relationship between a stock and its cluster peers doesn't shift wildly. The breakdown of this stability during crises is well-documented:

COVID-19 (March 2020): equity correlations spiked sharply within days; previously stable clusters collapsed into one or two highly correlated groups. Pairs and baskets stopped mean-reverting and instead all sold off together. Rolling 30-day correlations took ~30 trading days to "see" what had happened.
Russia–Ukraine (Feb–Jul 2022): commodity-sector correlations spiked, energy-equity dynamics changed. Same lag problem.

The empirical signature: pairs/basket strategies show catastrophic losses in March 2020 that don't recover for weeks. A "good" StatArb strategy is one whose worst drawdown coincides with these known events but recovers within a quarter; a "bad" one has structural drift that survives the regime.

Performance benchmarks (US equity, recent)

Strategy	Sharpe	MDD	Rebalance	Ref
SPONGEsym equal-weight, 3-day	1.1	n/a	3d	Cartea et al. 2023
ML-filtered SPONGEsym, 10-day	(paper claim large)	n/a	10d	Korniejczuk–Slepaczuk 2024
THGNN forward-looking SPONGEsym	1.84	−9.4%	10d	Fanshawe et al. 2025
SPX buy-and-hold (2019–2024)	0.65	−33.9%	—	benchmark

The Korniejczuk–Slepaczuk numbers should be read with caution — see Korniejczuk Slepaczuk 2024 Graph Clustering Stat Arb for the replication discussion.

Why this concept page lives in the equity section

StatArb is overwhelmingly an equity strategy in practice. The same techniques generalise to futures (cross-commodity spreads), FX (carry baskets), and crypto (token-pair mean reversion), but the canonical literature is in US large-cap equities and the S&P 500 is the standard benchmark. Coverage in other asset classes is delegated to cross-references.

Limitations and open questions

Transaction cost modeling is the central unaddressed problem. None of the cited backtests include realistic slippage or market impact. Top-10% filtered basket of ~45 stocks with 10-day rebalance is plausibly execution-friendly, but a square-root-impact model would likely cut Sharpe by 0.2–0.4.
Capacity is unknown. The cited results use S&P 500; capacity at $100M+ AUM is not documented.
Multiple-testing concerns. With hundreds of candidate strategies in the literature, the median published Sharpe is biased upward. Ledoit-Wolf stationary bootstrap (used in THGNN paper) is the right inferential framework but is not universal.
Forward-looking correlations raise the bar. If the dependence metric itself is ML-predicted, the strategy is now a black-box-within-black-box system. Interpretability, monitoring, and decay detection all become harder.

Alternative route: graph on factor exposures (MINGLE)

All the strategies above build the dependence metric on observed co-movement (rolling correlation, ML-predicted correlation). MINGLE 框架 (Chehab et al. 2026) takes a different route: build the graph on factor exposure similarity, jointly learn the loadings and the graph, then allocate Contagion Cut (簇内加权度反比). Result: 2019/01–2026/03 on S&P 500 / Nikkei / STOXX 600, 0 bps CAGR 21.0% / Sharpe 1.07. This is a portfolio diversification strategy rather than a market-neutral StatArb, but shares the graph-clustering back-end and shows the same "induce-covariance-then-allocate" pattern as the SPONGEsym family.