Reducing Noisy Signals in Log-Return Pair Trading

Summary

The document raises a practical issue in a cryptocurrency perpetual-swap pair-trading strategy: a spread estimated by regression on log returns can produce opposite signals in consecutive observations. The author calculates the spread from rolling historical candles on a 30-minute timeframe and asks how to reduce or smooth the noise. They consider applying an exponential moving average to returns but are unsure whether that is appropriate.

The document does not provide an answer, test results, or a recommended smoothing method. It is useful as a concise description of a signal-stability problem, but it offers no evidence about whether smoothing, changing the regression window, or another approach would improve performance. Any proposed filter would need to be evaluated for its effect on signal lag, trading costs, and out-of-sample results.

Key ideas

• Rolling regressions on log returns can produce rapidly alternating pair-trading signals.
• The issue is described in a cryptocurrency perpetual-swap strategy using 30-minute data.
• The author asks whether spread noise should be smoothed or canceled.
• An exponential moving average of returns is raised as an untested possibility.

Tags

# Pair Trading - Spread noise from Log returns regression


# Pair Trading - Spread noise from Log returns regression












I have an automated strategy that has been running "ok" for month and generates consistent profit on Perpetual swaps (cryptos). However something annoys me: I calculate the spread based on the regression of log returns and it can often generates two opposite consecutive signals because of the noise. My spread is calculated on 30min timeframes from 200 historical candldes (roling windows). Are you guys familiar with such issue ? Is there a "best practice" to cancel or smoothen the noise ? I had an idea to use the exponential moving average returns instead of the price itself but that sounds really bad.

Thanks in advance for sharing.

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.