Predicting energy futures high-frequency volatility using technical indicators: The role of interaction

Highlights

• Technical indicator has significant effects on crude oil and natural gas volatility.

• Interaction variables can significantly improve volatility forecasting performance.

• The improved CSIS model outperforms extensive rival methods.

• The improved model has robust out-of-sample performance.

Abstract

In this paper, we investigate the predictability of technical indicators on energy futures volatility from the high-frequency and high-dimensional perspectives. We show that the technical indicators have significant impacts on crude oil and natural gas futures volatility based on in- and out-of-sample analysis. Further, we analyze the impacts of interactions among predictor variables on future volatility. Based on an improved conditional sure independence screening model, we find that the interactions contribute to the out-of-sample predictive power significantly. The improved model has robust and better forecasting performance relative to extant popular dimension reduction methods, forecast combination methods, and regularization methods. Moreover, we show that the out-of-sample predictability is robust during various periods. Finally, we show that technical indicators improve economic value in the crude oil market but the economic increment is not significant in the natural gas market.

Introduction

Researchers pay much attention to energy commodities since it plays an irreplaceable role in the real economy (Ergen and Rizvanoghlu, 2016). As two important non-renewable and strategic resources in the energy market, crude oil and natural gas have attracted wide attention from investors and other finance professionals. Besides, they are the first two largest and most liquid commodity markets (Zhang and Wei, 2011, Borovkova and Mahakena, 2015). A growing number of scholars try to understand the dynamics of energy commodities prices and volatility through investigating the determinants of swings (Borovkova and Mahakena, 2015, Ergen and Rizvanoghlu, 2016, Yin and Yang, 2016, Qadan and Nama, 2018).

In this paper, we investigate the impacts of technical trading activities, measured by technical indicators, on crude oil and natural gas futures volatility from the high-frequency and high-dimensional perspectives. This study is associated with recent literature showing that technical indicators have significant impacts on asset prices (Neely et al., 2014, Yin and Yang, 2016, He et al., 2021) and stock volatility (Liu and Pan, 2020). Technical indicators rely on past information in a period to identify price or volatility trends believed to persist into the future (Neely et al., 2014). Thus, the technical indicator can reflect investors’ beliefs via capturing the changes in market information.

Theoretically, that technical indicators predict market prices and volatility is connected with an informationally inefficient market (Neely et al., 2014, Liu and Pan, 2020). The remarkable phenomenon is under-reaction and overreaction to information (Hong and Stein, 1999). The news-watchers predict the market based on their observation and do not rely on past information. Whereas, momentum traders make a decision depending on historical information. Investors under-react to the news at the beginning owing to behavioral biases. When the market rises, momentum traders overreact and thus increase prices. Thus, the market anomalies, such as momentum effect and reversal effect, appear, which leads to market inefficiency.

Besides, Liu and Pan (2020) attribute the predictability of technical indicators to the business cycle in the stock market. Using equal-weighted mean forecast combination framework, they document that the out-of-sample $R^2$ (a measure for evaluating out-of-sample performance) is larger in the economic recessions than economic expansions. However, we find that there is no consistent result in the energy futures markets, especially in the case of using high-frequency data. More specifically, our results indicate that almost all technical indicators have predictive power on future volatility only during economic expansions. Furthermore, the multi-factor models that consider all technical indicators also perform significant out-of-sample predictive ability only during economic expansions. Thus, our result shows that there is no significant relationship between the predictability of technical indicators and business cycle in the energy futures markets.

Using technical analysis to understand asset prices/returns behavior has received much attention. Related literature pays attention on explaining the predictability of technical indicators on asset returns in different markets, e.g., equity market (Neely et al., 2014), commodity market (Wang et al., 2020, He et al., 2021), cryptocurrency market (Hudson and Urquhart, 2021), currency market (Panopoulou and Souropanis, 2019), etc.. These technical indicators are constructed based on typical buy or sell signals caused by changes in prices, such as moving averages (Brock et al., 1992) and momentum effect (Conrad and Kaul, 1998). Thus, these proposed prevailing technical indicators link to the asset returns. However, investigation of how to construct technical indicator from the risk (volatility, secondary moment of returns) perspective is less common,¹ which is what we are interested in this study because return and risk are equally important and gain equally attention. Moreover, using technical analysis to understand the asset volatility behavior is also less common in the literature, especially in energy commodity markets. Although Liu and Pan (2020) discuss the impacts of technical indicators on stock market volatility, the empirical evidence is not enough since several questions are still unknown.

First, except for the technical indicators in Liu and Pan (2020)’s work, whether other technical indicators have significant impacts on asset volatility remains interest to scholars. Second, whether the technical indicator is valid for forecasting volatility in the energy futures markets? Furthermore, whether the technical indicator with higher frequency, say daily, can still predict volatility well? Third, so many extant studies focus on analyzing the individual influence of explanatory variables on the target variable. They usually use a dimension reduction method to improve prediction accuracy owing to the in-sample over-fitting problem. Very few scholars investigate the interaction among predictor variables, which could affect volatility prediction. Finally, if considering the interactions among a large set of predictors, the covariate will be hyper-dimensional, which brings many troubles into the prediction task. Whether there is an appropriate method to solve that is a challenge in this study.

On the way to answer these questions, this study contributes to the literature in four aspects. Firstly, we expand the categories of volatility technical indicators. Liu and Pan (2020) propose three types of technical indicators in the stock market, based on leverage effect, volatility clustering, and volume effect respectively. We further propose two types of technical indicators based on overnight volatility effect and price-range-based volatility effect, which incorporates overnight information and more prices information. The overnight volatility is viewed as a portion of total daily volatility (Andersen et al., 2011). The consideration of prices information comes from Das and Chen (2007) who use the changes in daily prices to measure volatility. This measure, using high, low, open, and closing prices, is simple but can effectively capture prices volatility on one day. The empirical results show that all types of technical indicators have significant in- and out-of-sample predictability on crude oil and natural gas futures volatility respectively.² Thus, we provide new evidence about technical indicators that influence energy futures volatility significantly.

Secondly, we demonstrate that the technical analysis in higher frequency data is feasible. Most extant literature discusses the impacts of technical indicators on future returns and volatility in monthly frequency or lower frequency, see, e.g., Neely et al., 2014, Yin and Yang, 2016, Liu and Pan, 2020, He et al., 2021. It ignores much trading information since the transaction happens frequently. Besides, using high-frequency data to analyze the behavior of asset prices or volatility can help understand market operational mechanisms from a microscopic point of view. Our results provide convincing evidence that high-frequency technical indicators can affect future high-frequency volatility of crude oil and natural gas significantly. Furthermore, we show that the technical indicators predict crude oil futures and natural gas futures volatility mainly during the economic expansion periods.

Thirdly, we consider the interaction among predictors to investigate the predictive impacts on volatility. Using an improved model, we show this consideration improves prediction accuracy significantly. Many studies investigate the influence of a large set of predictors on the response from two aspects. One is to seek which factor is more important for the target variable, see, e.g., Welch and Goyal, 2008, Christiansen et al., 2012, Gu et al., 2020. Another is to explore how to improve prediction performance using multiple factors based on some dimension reduction techniques, see, e.g., Lin et al., 2018, Zhang et al., 2019. Whereas, a few studies consider the effect of interactions on the target variable. The motivations of considering interaction effects come from three aspects. (i) Interaction terms play a positive role in forecasting tasks. For example, Saunders (1956) provides an early analysis of the product-variable approach to modeling an interaction effect. It is aware that accounting for important interaction effects can improve the prediction of many statistical learning methods. Omitting the product variables leads to the model where covariates have only an additive effect on the response variable. However, in reality, covariates may affect the effects of each other, which incurs cross-level product terms, see a toy example in Hargens (2009). (ii) Interaction effect can refer to nonlinear relation. Cortina (1993) shows that sometimes an interaction term may be statistically significant because of its overlap with nonlinear terms. And nonlinear relation is prevailing in the economic and financial fields, see, e.g., Li et al., 2019, Zhang et al., 2021 (iii) In our model, though we are not sure about the existence of interaction effects at the beginning, they can still be included as candidate predictors. This is because whether they are selected into the final model is fully driven by data. In this way, the model can be more flexible and robust compared with those only considering linear effect.

Under the framework of technical indicators, we find that the interactions among predictor variables have significant power to improve futures volatility prediction. This finding is the first main contribution of this study. What is more, we find that the main contributors for crude oil futures differ from natural gas. More specifically, the interactions among technical indicators are main predictors for crude oil futures volatility while that among volatility historical information and technical indicators contribute more predictability for natural gas futures volatility.

Finally, we develop an improved method that combines the conditional sure independence screening (CSIS) model and model average to improve volatility prediction accuracy, which is another main contribution. After considering the interactions among explanatory variables, the variable dimension is greatly increased, which challenges the prediction task obviously. To address this, based on the conditional sure independence screening model developed by Barut et al. (2016), we exploit the model average technique to improve prediction accuracy and robustness. We show that the improved model has the best prediction performance, in terms of out-of-sample $R^2$, relative to many extant typical methods, including principal components regression, partial least squares regression, Lasso, Elastic Net, and some forecast combination methods. This result is confirmed based on the model confidence set test developed by Hansen et al. (2011). We further explain that the improved model has robust out-of-sample predictive power during high/low attention, high/low sentiment, high/low uncertainty, and high/low volatility periods. As a portfolio application of our findings, we follow Degiannakis and Filis (2022) to construct a trading strategy based on forecasting models. The results show that the improved CSIS model produce significant economic value for crude oil futures. Whereas, the economic increment in the natural gas futures market is not significant, which could be related to the market complexity and investor activity.

The remainder of this paper is organized as follows. Section 2 shows the high-frequency volatility measure and technical indicators. Section 3 presents the improved predictive methods and competing models. Section 4 displays the empirical results, including in- and out-of-sample analyses. Robustness checks are investigated in Section 5. Economic value is discussed in Section 6. Finally, we make the conclusions in Section 7.