Sjur Westgaard, Florentina Paraschiv,Lina Lassesen Ekern, Ingrid Naustdal, and Malene Roland
1. Introduction
As electricity is a non-storable commodity, a stable power system requires a constant supply and demand balance. This makes electricity a unique commodity with complex price dynamics and relations to fundamentals. Prices are characterized by sudden (positive and negative) spikes, high volatility and volatility clustering, and seasonal patterns over the day, week, and year. Thus, forecasting in electricity markets is arguably more challenging than in traditional financial markets.
Electricity price forecasts are important inputs for decision making by energy companies. For day-to-day market operations, accurate forecasts of short-term prices are crucial. Price forecasts also assist producers, retailers, and speculators who seek to determine their optimal short-term strategies for production, consumption, hedging, and trading. Uncontrolled exposure to market price risk can have devastating consequences for market participants (see the discussion in Deng and Oren (2006)). This has led to an increased focus on risk management in power markets during the last few years.
The recent introduction of smart grids and renewable integration has created more uncertainty in future supply and demand conditions where both very low prices (often negative) and high prices might occur. Over the last 15 years, most research has been concerned with predicting the mean for electricity prices. As stakeholders require explicit control of the risk of both high and low extreme prices, point forecasts are inadequate in many cases (see Nowotarski and Weron (2017)). Academics and practitioners have come to understand that probabilistic electricity price forecasting is now more important for energy systems planning, risk management, and operations than ever before.
Value-at-Risk (VaR) is the market standard for risk measurement and is simply a given quantile that will be found directly from a price distribution forecast. Despite the importance of measuring risk management in power markets, Weron (2014) and Nowotarski and Weron (2017) find that distribution forecasting is “barely touched upon” in the electricity price forecasting literature. This statement is supported by Bunn et al. (2016), who argue that for electricity markets, VaR forecasting remains a highly “under-researched” area. Maciejowska et al. (2016) claim that the lack of such research is probably due to the embedded complexity of the research problem compared with point forecasting. The sparseness in the current literature, combined with the importance of density forecasting, is our motivation for investigating how state-of-the-art econometric models can be applied to forecast VaR. We have chosen to look at the German market for several reasons: (1) it is probably the most important electricity market in Europe; (2) data quality, transparency, and access are excellent; and (3) the input mix of production towards renewables has changed a lot over the years, challenging us to build models that capture this dynamics. We need to identify models that can capture the non-linear sensitivities of electricity prices to fundamentals because of the convex supply curve as well as time-varying sensitivities to fundamentals.
We argue for the use of quantile regression (QR) models when forecasting price distributions and estimating VaR. QR models estimate each quantile with a distinct regression. Moreover, they are simple, insensitive to outliers, avoid distributional assumptions, and help find non-linear sensitivities (e.g., that electricity price sensitivities to gas prices should be higher when electricity prices are higher since gas power plants are used in such a regime “at the end” of the supply or merit order curve). In addition, we also apply QR models with time-varying parameters taking into account that the input mix has changed over time.1 These models are exponentially weighted QR (EWQR) and exponentially weighted double kernel QR (EWDKQR) as proposed by Taylor (2008b). We compare these alternatives with common benchmarks in the VaR prediction literature using GARCH and CAViaR types of models.
By using knowledge of market conditions, we form a set of fundamental factors (supply and demand variables) and perform a variable selection procedure for each trading period and quantile with the aim of (1) proper in-sample fit and (2) optimal out-of-sample fit for a given hour and a given quantile. Hence, in additional to the dimensions of the model choice, we stress the fact that variable selection should be carefully monitored as the various fundamentals influence hours and quantiles differently.
To sum up, the overall goal of our work is threefold. First, we want to identify appropriate fundamental variables for selected hours and quantiles of the price distribution. Second, we assess the gain of using more complex QR models compared to traditional QR models and given GARCH and CAViaR benchmarks. Third, to our knowledge, no such comprehensive VaR prediction study for the German electricity market has been performed.
The paper is structured as follows. In Section 2 we review relevant literature on fundamental electricity price modeling and VaR forecasting. Next, we describe the German power market and price formation process in Section 3. In Section 4 we present and analyze the data set. In Section 5, we give an explanation of the models and evaluation procedures for distributional forecasts. We present and discuss the empirical results in Section 6 and conclude in Section 7.
2. Literature review
We position ourselves between the following groups in the literature: (1) VaR forecasting of asset markets and more specifically energy commodities and (2) fundamental analyses of electricity price formation.
VaR forecasting is complicated by the fact that high-frequency asset prices (including electricity prices) exhibit challenging data features. Time-varying volatility, skewness, and kurtosis induce a lot of complexity to the modeling of VaR (see Hartz et al. (2006)). Kuester et al. (2006) provide a comprehensive review of VaR prediction strategies that can solve some of these issues.
In our context, available models for VaR forecasts can be classified into three main categories:
- Fully parametric modelsassuming a given error distribution (e.g., a GARCH-skew-t model, EVT models).
- Non-parametric approaches such as historical simulation , where one computes empirical quantiles based on past data. Historical simulation can be filtered taking into account time-varying volatility.
- Semi-parametric approaches such as quantile regressionthat directly models specific quantiles with no assumption regarding the error distribution.
Most of these models have been applied without using fundamentals when we look at the energy market literature. Gurrola-Perez and Murphy (2015) evaluate historical and filtered historical simulation models for energy markets. According to paper above these authors the latter approach should be used as it greatly improves the distributional forecast. At present, research on EVT for estimating VaR in energy markets is sparse. However, examples are found in Bystrom (2005), Chan and Gray (2006), and Paraschiv and Hadzi-Mishev (2016), who all report that their results are encouraging. Garcia et al. (2005) use two GARCH models to forecast spot prices for the Spanish and Californian markets: one with price as the only variable, and one including demand. They benchmark these against an ARIMA model. They find that GARCH with price only outperforms ARIMA when time-varying volatility and price spikes are present. Moreover, adding demand as an explanatory variable further improves the forecasting performance. When assuming t or skewed t distribution, GARCH models show promising results when forecasting VaR for commodities, including energy commodities (see Giot and Laurent (2003) and Fuss et al. (2010)). Conditional Autoregressive Value-at-Risk (CAViaR) models by Engle and Manganelli (2004) model quantiles directly as an autoregressive process. The estimation is based on a quantile regression approach. The performance of CAViaR models are promising for electricity markets (see Fuss et al. (2010) and Bunn et al. (2016)). Bunn et al. (2016) also find that classical quantile regression models give excellent VaR forecasts for UK electricity prices. Florentina and Hadzi-Mishev (2016) use a combination of GARCH and EVT to investigate the tails of the German electricity price change distribution. They find that the model delivers relatively precise quantile estimates; however, the quality of the estimates is sensitive to the ...