Performance of Electricity Price Forecasting Models: Evidence from Turkey

ABSTRACT

In this article, hourly prices of the Turkish Day Ahead Electricity Market are forecasted by using various univariate electricity price models, then the out-of-sample forecasts are compared with each other and the benchmarks. This article has two main contributions to the literature: Firstly, it provides a factorial Analysis of Variance (ANOVA) as a pre-whitening method of the price series and allows one to work with the stationary residuals series. Secondly, it is the first work, which compares the performances of all important statistical univariate forecast models in the Turkish electricity market. Results indicate the importance of the factorial ANOVA application and the SARIMA model’s success under the given conditions.

KEY WORDS:

• electricity price forecasting
• factorial ANOVA
• pre-whitening
• Turkish day-ahead electricity market
• univariate time-series modeling

JEL classification:

• C22
• C52
• C53
• G17

Acknowledgments

We would like to thank Michael Coulon, Burc Ulengin, Mehmet Fuat Beyazit, Tarryn Catherine Williams and all the participants of the 4th Young Finance Scholars’ Conference, and The Commodity and Energy Markets Annual Meeting 2017. We also would like to express our gratitude to the editor, Ali M. Kutan, and four anonymous referees for their invaluable comments and suggestions on the draft of this article.

Supplemtary Material

Supplementary data for this article can be accessed here.

Notes

^1. It is surely possible that balancing market prices could be lower than the day-ahead market prices, however it is the last chance to buy/sell the electricity; and if prices occur in unexpectedly high/low levels, it might cause significant losses to the companies. Furthermore, in the Turkish Balancing Market, there is a 3% penalty for trading in the balancing market (EMRA 2017).

^2. Electricity prices in the day-ahead markets are bounded in some markets for preventing big losses. See Negahdary and Ware (2016) for Alberta power prices.

^3. Negative prices is an important topic in the electricity prices, especially in the high renewables share markets, such as Germany. Bublitz, Keles, and Fichtner (2017), Fanone et al. (2013) and Keles et al. (2012) can give more information about the negative prices.

^4. Eleven of them are European markets and the other one is the GEFCom2014 competition data.

^5. Interestingly, this is the NATAF transformation of Díaz and Planas (2016). Uniejewski, Weron, and Ziel (2017) mention that Díaz and Planas (2016) call NATAF transformation to the N-PIT transformation, misleadingly.

^6. It is impossible to remove the non-stationarity by taking log returns due to the zeros.

^7. https://seffaflik.epias.com.tr/transparency/.

^8. 8760 hours for 2012–2015 and 8784 for 2016.

^9. Numbers given under the hours represent the following hour. For example; 1 represents, 01:00:00–01:59:59.

^10. “Weekdays” is used for Monday, Tuesday, Wednesday, Thursday, and Friday in this case.

^11. In Turkey, most of the companies work half-day on Saturday and factories work all-day.

^12. P-value of the Augmented Dickey-Fuller test is 0.1175.

^13. ANOVA has the assumptions of the independence of the cases, normal distribution of the residuals, homoscedasticity and no multicollinearity. It is very difficult to fulfill them, which was the case for us even after attempts with various transformations from Uniejewski, Weron, and Ziel (2017). It must be kept in mind that ANOVA assumptions couldn’t be fulfilled.

^14. Interaction effects are the effects between the independent variables. For example, if day of the week have an additional effect via month of the year, it is named as interaction effect. It is a kind of combined effect of two or more independent variables.

^15. P-value of the Augmented Dickey-Fuller test is 0.0000.

^16. Although autocorrelation - partial autocorrelation functions of the residuals show that the effect of the autocorrelations and partial autocorrelations are decreased in the residuals, according to TBATS model (De Livera, Hyndman, and Snyder 2011) seasonality is still in the residuals (stochastic part) for the 24th and 168th lags.

^17. It is actually a SETAR model, which will be discussed in this section.

^18. For detailed information about the SARIMA model, see Box and Jenkins (1976).

^19. See Hamilton (1994).

^20. Expanding window scheme has been used. The estimation period was from 01.01.2013 to the previous day of the forecast day.

^21. Lag selections are done according to autocorrelation – partial autocorrelation functions and all the estimations are available upon request.

^22. Turkish Day-Ahead Market doesn’t allow prices to exceed 0 and 2000 TL/MWh, downside and upside, respectively. In other markets, such as German EEX, market-makers also allow negative prices, which cause forecasting problems as well.

^23. 2.88% of the prices are exactly same with the previous day’s prices and 11.01% have less than 1 TL/MWh difference in 2016, which are quite high levels compared to the other methods’ forecasts. For further information, bid structure must be examined.

^24. Due to high levels of air-conditioning usage.

^25. As an example, prices of 15 July, which is analyzed in this article, is given in the Table S6 (available online) with the previous day’s prices and AE.

Additional information

Funding

In the last period of research, Umut Ugurlu is supported by The Scientific and Technological Research Council of Turkey, 2214/A Programme. This work is supported by Research Fund of the Istanbul Technical University; project number: SDK-2018-41160.