Benchmarking Volatility Forecasts Across Consecutive Trading Days

Summary

The document frames a measurement problem for forecasts of returns from a chosen time on one trading day to the same time on the next. It asks what ex post variance should serve as the benchmark for evaluating those forecasts. A simple sum of squared short-interval returns raises a concern: the market closure and reopening may create a large discontinuity. Squaring only the total cross-day return also appears inadequate to the author, though no alternative estimator is proposed.

The material is a research question rather than a solution. It identifies a mismatch between ordinary intraday realized variance and a target interval that includes an overnight closure, and asks whether existing literature offers a consensus. No papers, data, estimator, or empirical comparisons are provided. The appropriate benchmark may depend on how the forecast target treats overnight price changes and on the market’s trading hours, so readers would need further research before choosing a method.

Key ideas

• The forecast horizon runs between the same clock time on consecutive trading days.
• Intraday squared returns may not represent that horizon cleanly because the market closes overnight.
• Using only the squared total return is also questioned as a benchmark.
• The document poses the issue but supplies no estimator, literature references, or empirical evidence.

Tags

# Cross-day realized volatility


# Cross-day realized volatility












I've been looking for papers on volatility forecast, and most of them focus either on daily volatility (often using daily returns to access predictions for monthly volatility). Others focus on intraday-volatility, where intra-day returns can be used.

However, I'm working on the forecast of volatility of returns between a given time in the a trading day and the same time on the consecutive trading day. What could I use in this case as an ex post variance to use as a benchmark for my models?

Simply using squared returns over say 15 minutes intervals seem wrong because there would be a large squared return due to the closing of the market and its reopening on the next day. It also would seem wrong to simply use the squared return over the whole period.

Is there any reference that has addressed this issue on the literature? Is there a consensus on how to benchmark a cross-day volatility forecast?

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.