Tim Olson
47471b7700
- Flink update_bars debouncing - update_bars subscription idempotency bugfix - Price decimal correction bugfix of previous commit - Add GLM-5.1 model tag alongside renamed GLM-5 - Use short Anthropic model IDs (sonnet/haiku/opus) instead of full version strings - Allow @tags anywhere in message content, not just at start - Return hasOtherContent flag instead of trimmed rest string - Only trigger greeting stream when tag has no other content - Update workspace knowledge base references to platform/workspace and platform/shapes - Hierarchical knowledge base catalog - 151 Trading Strategies knowledge base articles - Shapes knowledge base article - MutateShapes tool instead of workspace patch
98 lines
4.2 KiB
Markdown
98 lines
4.2 KiB
Markdown
---
|
||
description: "Generalizes pairs trading to N>2 historically correlated stocks within a cluster (e.g., an industry), buying underperformers and shorting overperformers relative to the cluster mean."
|
||
tags: [stocks, mean-reversion, cluster, statistical-arbitrage]
|
||
---
|
||
|
||
# Mean-Reversion — Single Cluster (and Multiple Clusters)
|
||
|
||
**Section**: 3.9 / 3.9.1 | **Asset Class**: Stocks | **Type**: Mean-Reversion / Statistical Arbitrage
|
||
|
||
## Overview
|
||
This strategy generalizes pairs trading to N > 2 stocks that are historically highly correlated — for example, stocks belonging to the same industry or sector. Each stock's return is demeaned relative to the cluster mean, and positions are taken proportional to negative demeaned returns: buy stocks that underperformed the cluster and short stocks that outperformed it.
|
||
|
||
## Construction / Signal
|
||
Let `R_i = ln(P_i(t_2) / P_i(t_1))` be the log return for stock i in the cluster of N stocks.
|
||
|
||
Cluster mean return and demeaned returns:
|
||
|
||
```
|
||
R_bar = (1/N) * sum_{i=1}^{N} R_i (293)
|
||
R_tilde_i = R_i - R_bar (294)
|
||
```
|
||
|
||
Short stocks with positive `R_tilde_i` (outperformers), buy stocks with negative `R_tilde_i` (underperformers).
|
||
|
||
**Dollar-neutrality constraints**:
|
||
|
||
```
|
||
sum_{i=1}^{N} P_i |Q_i| = I (295)
|
||
sum_{i=1}^{N} P_i Q_i = 0 (296)
|
||
```
|
||
|
||
A simple prescription for dollar positions `D_i = P_i Q_i` proportional to demeaned returns:
|
||
|
||
```
|
||
D_i = -gamma * R_tilde_i (297)
|
||
```
|
||
|
||
where `gamma > 0` (short outperformers, buy underperformers). Eq. (296) is automatically satisfied; Eq. (295) fixes gamma:
|
||
|
||
```
|
||
gamma = I / sum_{i=1}^{N} |R_tilde_i| (298)
|
||
```
|
||
|
||
## Entry / Exit Rules
|
||
- **Entry**: Compute demeaned returns over a short measurement window; enter positions D_i = -gamma * R_tilde_i.
|
||
- **Exit**: Close when demeaned returns converge back toward zero, or at a predefined time horizon.
|
||
|
||
## Key Parameters
|
||
- **Cluster definition**: Industry group, sector, or any set of historically correlated stocks
|
||
- **Measurement window**: Short-term (days to weeks) for mean-reversion
|
||
- **Position sizing**: Dollar-neutral via gamma normalization (Eq. 298)
|
||
- **Weights**: Uniform modulus, or non-uniform (e.g., suppressed by volatility)
|
||
|
||
## Variations
|
||
|
||
### 3.9.1 — Mean-Reversion: Multiple Clusters
|
||
|
||
Generalize to K > 1 clusters, where stocks within each cluster are historically highly correlated. Treat clusters independently and combine via linear regression (unified approach).
|
||
|
||
Let `Lambda_{iA}` be the N×K binary loadings matrix: `Lambda_{iA} = 1` if stock i belongs to cluster A, else 0. Cluster sizes: `N_A = sum_{i=1}^{N} Lambda_{iA} > 0`, `N = sum_{A=1}^{K} N_A`.
|
||
|
||
Run a linear regression of stock returns R_i on Lambda_{iA} (no intercept, unit weights):
|
||
|
||
```
|
||
R_i = sum_{A=1}^{K} Lambda_{iA} f_A + epsilon_i (303)
|
||
```
|
||
|
||
Regression coefficients (cluster mean returns):
|
||
|
||
```
|
||
f = Q^{-1} Lambda^T R, Q = Lambda^T Lambda (304, 305)
|
||
Q_{AB} = N_A delta_{AB} (307)
|
||
R_bar_A = (1/N_A) sum_{j in J_A} R_j (308)
|
||
```
|
||
|
||
Demeaned return (residual) for stock i:
|
||
|
||
```
|
||
epsilon_i = R_i - R_bar_{G(i)} = R_tilde_i (309)
|
||
```
|
||
|
||
where G(i) is the cluster to which stock i belongs. These residuals are cluster-neutral:
|
||
|
||
```
|
||
sum_{i=1}^{N} R_tilde_i Lambda_{iA} = 0, A = 1,...,K (310)
|
||
```
|
||
|
||
Also automatically: `sum_{i=1}^{N} R_tilde_i = 0` (dollar-neutral).
|
||
|
||
Investments can be allocated uniformly across the K independent cluster strategies.
|
||
|
||
## Notes
|
||
- The single-cluster strategy (3.9) is the natural generalization of pairs trading to N stocks.
|
||
- The multiple-cluster (3.9.1) formulation uses linear regression to compute all cluster means simultaneously in a unified framework.
|
||
- The binary loadings matrix Lambda_{iA} ensures each stock belongs to exactly one cluster (no overlap assumed).
|
||
- The intercept is automatically included via the constraint that each stock belongs to one cluster (sum_A Lambda_{iA} = 1).
|
||
- Mean-reversion strategies work best when stocks are truly co-integrated or highly correlated; sector/industry groupings provide natural clusters.
|