Interpreting Kalman Filter Signals in Mean-Reverting Pairs Trading

Summary

The document raises a question about how to interpret a Kalman filter signal in a mean-reverting spread strategy. In the cited setup, the latent spread follows a mean-reverting process, while observed spread values include measurement noise. The proposed signal compares the current observation with the filter’s estimate based on prior observations.

The author wonders whether an observation above that estimate indicates an overpriced spread that should be shorted. The question highlights a key distinction: the cited rule treats the deviation from the prior estimate as evidence that the spread is unusually large and suggests taking a long spread position, while the author reasons from mean reversion and expects a decline. The document does not resolve this apparent sign or position-convention issue. It gives no empirical evidence or implementation details, so interpreting the rule requires clarifying how the spread portfolio is defined and how the filter’s forecast relates to the strategy’s intended trade.

Key ideas

• The model separates a latent mean-reverting spread from noisy observed spread values.
• The proposed signal compares the observation with the filter’s estimate based on past data.
• The author questions whether a positive deviation calls for a long or short spread position.
• The correct trade direction depends on the spread and portfolio sign conventions.

Tags

# Why long a position when the acutal price is higher than the predicted price? (Kalman filter for pairs trading)


# Why long a position when the acutal price is higher than the predicted price? (Kalman filter for pairs trading)












In this paper Elliott, R., van der Hoek, J. and Malcolm, W. (2005) Pairs Trading., the spread (state process) is assumed to follow a mean reverting process $x_{k+1}-x_k=(a-bx_k)\tau+\sigma\sqrt{\tau}\varepsilon_{k+1}$, and the observed spread is assumed to follow a (observation) process $y_k=x_k+Dw_k$.

In section 2.3, the authors have mensioned that "if $y_k>\hat{x}_{k|k-1}=E[x_k|\sigma(y_0,\cdots,y_{k-1})]$ then the spread is regarded as too large, and the trader could take a long position in the spread portfolio and profit when a correction occurs."

I want to ask why this trading strategy is reasonable. If $y_k>\hat{x}_{k|k-1}=E[x_k|\sigma(y_0,\cdots,y_{k-1})]$, I think that the predicted spread at time $k$ is smaller than oberserved spead, so the price is high right now and it will go down eventually since it is assumed to be mean reverting, so maybe I will short the position.

Appreciate any suggestion and comments.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.