[Help] Real-world Pair Trading Issue: Spread Keeps Rising Despite Cointegration

General
1

After conducting a full scan of S&P 500 stocks for pair trading opportunities, I identified NFLX and PLTR as a viable pair. The cointegration test was performed using a 126-trading-day rolling window, and the spread was defined as:

Spread = PLTR – A × NFLX + B

The chart shows a clear cointegration relationship between February and late July 2025.

Backtesting suggested that a Z-score entry range of [-1.5, 1] provided consistent performance. I executed several trades using this strategy, generating actual profits.

Most Recent Trade

  • Entry Date: Around July 15, 2025, when Z-score ≈ 1
  • Subsequent Events:
    • On July 17, NFLX released its earnings. Since then, its stock has steadily declined with falling volume.
    • At the same time, PLTR surged, rising from ~$130 to $150+, causing the spread and Z-score to rise sharply.

As of July 28, 2025 (U.S. Eastern Time)

  • Z-score: 2.9
  • Half-life: 12.34 trading days
  • Cointegration (126-day window): Still valid
  • Bayesian probability of mean reversion: Estimated at 70–80%

:question: Key Question:

1. Should I now execute a simple stop-loss (e.g., exit at –X% loss)?

If not, what would be a more appropriate decision framework?

:repeat: Broader Concern – Model Fragility

Single-pair trading is inherently unstable.

Yes, one could stop out when divergence persists and switch to another pair. But this approach naturally limits returns and may even incur a series of losses. In other words, this method is not self-sustaining over time.

:white_check_mark: Looking for Better Solutions

Are there stronger theoretical foundations to handle this issue?
In this case, a direction is needed instead of a solution with details.

I’m considering:

  • PCA-based spread construction for more stable latent factor portfolios
  • Multi-pair selection with scoring models (e.g., cointegration strength, half-life, Sharpe ratio)
  • Or any robust statistical arbitrage frameworks that adapt to regime changes

If you there are relevant papers, books, or implementation ideas, I’d really appreciate your recommendations.

Thanks in advance!

Spread Visualization Chart

2

Hi, we are looking into your query, we will shortly give an update.

3

Firstly Rong, thank you for sharing. Such an occurrence is a very real challenge in live pair trading. From what you’ve described, the divergence post-NFLX earnings looks more like a regime shift than normal spread noise, even though the cointegration still holds statistically. In such cases, sticking to a simple Z-score model without accounting for fundamental catalysts like earnings can be risky. I’d lean towards closing the trade, especially since the post-event price action is clearly driven by asymmetric news flows.

You’re also absolutely right about the broader issue, single-pair strategies are inherently fragile. Cointegration may hold in the long run, but short-term dislocations due to events or changing regimes are common. Relying solely on one pair increases the risk of extended drawdowns and limits diversification.

To address this structurally, I’d suggest looking into factor-based or PCA-driven spread construction to reduce idiosyncratic risk.

You could also build a multi-pair scoring model that selects and rotates among several candidates based on statistical strength, volatility, and event sensitivity. For further robustness, consider regime-switching models like Hidden Markov Models or Bayesian filters that can adapt when the dynamics shift.

Hope this helps.