Using Hourly Data to Estimate Multi-Day Mean Reversion

Summary

The document asks whether hourly observations can be used to estimate dynamics that matter over a horizon of several days, particularly mean reversion or cointegrating relationships. It refers to a proposed analogy between estimating monthly relationships from daily data and estimating daily relationships from hourly data, with time-scale conversions offered as the initial rationale.

The key issue is that simple time scaling may not preserve the relevant dynamics. Sampling frequency, serial dependence, market trading hours, microstructure effects, and changes in the relationship across horizons can all affect estimates. A model fitted to hourly data may therefore describe short-interval behavior without reliably capturing multi-day adjustment. The post reports that the cited method’s author doubted the proposed use but gives no explanation or empirical test. It asks whether higher-frequency data can support estimation at a longer horizon; no answer, estimation procedure, or results are included, so the question remains open.

Key ideas

• The post asks whether hourly data can estimate mean reversion over a multi-day horizon.
• It proposes translating between sampling intervals using calendar time as a starting point.
• Simple rescaling may fail when serial dependence, market hours, or microstructure affect the data.
• The cited author expressed doubt, but the document provides no rationale or empirical evidence.
• No method or conclusion for estimating longer-horizon cointegration is supplied.

Tags

# Estimating Daily Dynamics using Hourly Data


# Estimating Daily Dynamics using Hourly Data












This article gives a nice outline of how daily data can be used to estimate cointegration on a monthly horizon.

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1404905

I'd like to use the same method to use hourly data to estimate mean reversion on a horizon of a few days.

The scaling seems to be the same one month=approx 22 working days and, one day =24 hours.

I spoke to the author and he seemed to think that using hourly data to estimate the daily dynamics would not work but didn't explain why.

Q1.) I respect the author so I assume he's correct, but I'd still like to understand why you can't use the hourly data to estimate daily dynamics?

Q2.) Assuming he is is correct is there any way I can estimate cointegrating relationships that occurs over the period of days from higher frequency data?

Thanks

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.