Forecasting electricity spot prices using time-series models with a double temporal segmentation

ABSTRACT

The French wholesale market is set to expand in the next few years under European pressure and national decisions. In this article, we assess the forecasting ability of several classes of time-series models for electricity wholesale spot prices at a day-ahead horizon in France. Electricity spot prices display a strong seasonal pattern, particularly in France, given the high share of electric heating in housing during winter time. To deal with this pattern, we implement a double temporal segmentation of the data. For each trading period and season, we use a large number of specifications based on market fundamentals: linear regressions, Markov-switching (MS) models and threshold models with a smooth transition. An extensive evaluation on French data shows that modelling each season independently leads to better results. Among nonlinear models, MS models designed to capture the sudden and fast-reverting spikes in the price dynamics yield more accurate forecasts. Finally, pooling forecasts give more reliable results.

KEYWORDS:

• Forecasting
• electricity spot prices
• seasonality
• regime-switching
• combinations

JEL CLASSIFICATION:

• C22
• C53
• L94

Acknowledgements

We would like to thank Derek W. Bunn, Gorkem Celik, Michel Cruciani, Cathy Dolignon, Mathieu Bouville, Philippe Vassilopoulos and anonymous referees for their comments and suggestions and Audrey Mahuet for her help in EPEX data. All remaining errors of course are ours.

Notes

¹ The rest of exchanges takes place over-the-counter (OTC), through direct transactions or through intermediaries (brokers and trading platforms). The total volume of OTC transactions is not public.

² The TTF is a virtual trading point for natural gas in the Netherlands, set up in 2003. It became the biggest gas trading platform in Europe.

³ Many authors only include the (forecasted) demand in addition to lagged prices in the forecasting equation. See for instance Nogales et al. (2002), Conejo et al. (2005), Misiorek, Trueck, and Weron (2006), Weron and Misiorek (2008).

⁴ In November 2010, the Central Western Europe Market Coupling (CWE) was launched. CWE is a cooperation of Transmission System Operators and power exchanges covering the Belgian, Dutch, French and German electricity markets.

⁵ Given the delay in the publication of the explanatory variables and the early publication of spot prices, the forecasts performed here are useful over a short window, from 0:00 to noon. Market participants use this window to optimally elaborate their bidding strategies.

⁶ Negative price can occur in situations of overcapacity. We discarded these observations to be able to use the logarithmic transformation.

⁷ To check the stationarity of the variables, we have implemented the Augmented Dickey-Fuller (ADF), Phillips–Perron (PP) and Kwiatkowski–Phillips–Schmidt–Shin tests. The tests were applied to the series in levels and in first differences when necessary. The results are available upon request.

⁸ 293 extreme values out of the 24 792 observed prices have been removed for the readability of the graph.

⁹ Alternative approaches include the use of dummies for the trading periods in a unique equation (Popova (2004)). However, this method is far less flexible or parsimonious. Alternatively, multivariate approaches are employed by Huisman, Huurman, and Mahieu (2007) with a panel framework or by Panagiotelis and Smith (2008) taking into account the hourly heterogeneity in a vector autoregression model.

¹⁰ To remove the weekly seasonality, Crone and Kourentzes (2011) split the hourly data of UK electricity load in seven subclasses of subseries, one for each day of the week. In this article, the weekly seasonality is less of an issue since we focus on working days.

¹¹ We have assessed the existence of a break in the forecasting equations due to the implementation of market coupling in November 2010 (see footnote 4). Rolling regressions do not show any major change in the relationship between spot prices and the price drivers. Therefore, we do not incorporate a structural change in the forecasting equations in November 2010.

¹² The lagged weighted average price is more significant than the price at the same hour the day before. The weighted average price is also smoother and it makes the estimation of nonlinear models easier.

¹³ The linearity tests are not performed at each recursion because it is too time consuming. However, we exclude the 35 forecasted observations to have an out-of-sample experience.

¹⁴ For sake of parsimony, the results are not reported in the paper but they are available upon request.

¹⁵ The number of parameters in the AR and LSTEP models varies at each iteration and it is fixed in the other specifications.

¹⁶ Using Monte Carlo simulations, Busetti and Marcucci (2013) provided a comparison of the finite sample properties of several tests of equal MSE and forecast encompassing.

¹⁷ Like Hansen, Lunde, and Nason (2011), we use a block bootstrap procedure with block length of 12 observations and 1000 resamples for the computation of the p-values.

¹⁸ We thank an anonymous referee for this suggestion.

¹⁹ To compute the model confidence set of Hansen, Lunde, and Nason (2011), we use the Matlab code provided by K. Sheppard.

²⁰ To get a fair comparison, the specification LSTEP (which contains more regressors than the non-linear models) is excluded in the comparison of the linear and nonlinear models.

²¹ This result is probably due to a drop in prices in summer 2009 included in the seasonal dataset. This phenomenon reflects the fall in industrial activity during the 2008–2009 recession in France. This is less visible in other seasons where electricity demand is mainly driven by the residential sector (heating effect).