Forecasting REIT volatility with high-frequency data: a comparison of alternative methods

ABSTRACT

Volatility is a crucial input for many financial applications, including asset allocation, risk management and option pricing. Over the last two decades the use of high-frequency data has greatly advanced the research on volatility modelling. This article makes the first attempt in the real estate literature to employ intraday data for volatility forecasting. We examine a wide range of commonly used methods and apply them to several major global REIT markets. Our findings suggest that the group of reduced form methods deliver the most accurate one-step-ahead forecast for daily REIT volatility. They outperform their GARCH-model-based counterparts and two methods using low-frequency data. We also show that exploiting intraday information through GARCH does not necessarily yield incremental precision for forecasting REIT volatility. Our results are relatively robust to the choice of realized measure of volatility and the length of evaluation period.

KEYWORDS:

• Volatility forecasting
• high-frequency data
• REIT market
• out-of-sample analysis

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

¹ In calculating RV, we use m – the number of intraday time intervals as shown below. For some trading days, the actual m is less than the number shown, due to delayed openings and/or early closings of the exchange.

Table

² Equations (4) and (5) are used only in the case of a multi-step-ahead forecast.

³ For HAR_CJ of Equation (8), this means that the right-hand side uses $log(C_t)$ and $log(1+J_t)$ (since $J_t$ can be 0 for some days).

⁴ For HEAVY, we do not estimate Equations (4) or (5). See the reason in Note 2.

⁵ We use the R package – ‘forecast’ and ‘fracdiff’ to carry out the algorithms. The main function used is ‘arfima’. Once the model is estimated, we then use the function ‘forecast’ to generate volatility forecasts.

⁶ For the reduced form methods, the forecast is actually for log volatility. We need to exponentiate the forecast to obtain a prediction of future volatility.

⁷ GARCH and ARFIMA_SR are fully based on daily data while GARCHX_RV and HEAVY are partially based since RV is used in their variance equation.

⁸ Market microstructure effects would prevent us from using sampling intervals that are too short.

⁹ Plots for other markets are not presented but they are available upon request.

¹⁰ To preserve space, we do not present the technical details of how to carry out the SPA test. Interested readers may refer to Hansen (2005).