:: Economic Performance of Forecasting Methods‎: ‎Evidence from Energy Arbitrage Trading (Alessandro Fiori Maccioni)

Price forecasting is central to arbitrage trading‎, ‎where profits depend on the ability to buy low and sell high‎. ‎Traders must anticipate market movements to execute transactions that exploit temporal price differences‎.

‎For storage operators‎, ‎maximizing arbitrage profits requires both predicting price peaks and timing buy-sell decisions accordingly‎, ‎given the storage constraints‎. ‎Effective arbitrage thus depends on accurately forecasting prices and translating them into optimal trading strategies‎.

‎While many forecasting methods have been proposed and assessed based on statistical accuracy‎, ‎less attention has been paid to their actual impact on the economic performance of storage operations‎.

‎This study investigates whether specific forecasting models consistently lead to higher arbitrage revenues and examines how their economic value can be identified through appropriate evaluation metrics‎. ‎Using data from the Italian day-ahead electricity market between 2020 and 2024‎, ‎we evaluated the economic performance of several forecasting methods of various types‎, ‎including singular spectrum analysis (SSA)‎, ‎SARIMA‎, ‎random forests‎, ‎and simple moving averages‎.

‎Each model was coupled with an optimization algorithm that scheduled battery charge and discharge operations based on forecasted prices‎, ‎computed the resulting arbitrage profits‎, ‎and enabled a direct comparison of financial outcomes under identical market and system constraints‎.

‎Results indicate that univariate SSA consistently outperformed all other models in delivering arbitrage profits‎. ‎Over the five-year horizon‎, ‎SSA achieved average profits equal to 98\% of the theoretical maximum‎, ‎with annual values never falling below 97\%‎. ‎This strong performance stemmed from SSA’s ability to extract dominant price patterns while maintaining low forecast error‎. ‎However‎, ‎this accuracy required substantially greater computational resources than any other method considered‎. ‎Despite this‎, ‎SSA’s advantage remained robust across battery configurations‎, ‎retraining schedules‎, ‎and when accounting for realistic cost factors such as efficiency degradation and transaction fees‎.

‎In contrast‎, ‎simpler models---particularly the 30-day moving average and standard ARIMA‎, ‎SARIMA‎, ‎and random forest---also performed well‎, ‎yielding between 87.8\% and 90.5\% of the perfect-foresight benchmark‎. ‎These models required minimal tuning and offered considerable operational efficiency‎. ‎Notably‎, ‎more granular and computationally intensive variants of random forest did not translate higher forecast accuracy into increased profits‎, ‎likely due to their limited ability to capture key market features such as price oscillations‎, ‎intraday ordering‎, ‎and local extrema‎. ‎This highlights that greater statistical accuracy does not necessarily lead to better economic outcomes‎, ‎and that model complexity must be balanced against practical deployment considerations‎.

‎To further investigate the link between predictive accuracy and realized profit‎, ‎we compared each method’s financial performance with standard forecast accuracy metrics‎. ‎Interestingly‎, ‎methods with the best statistical performance did not consistently yield the highest profits‎. ‎This suggests that conventional accuracy metrics may be insufficient for evaluating the economic value of price forecasts‎. ‎To address this gap‎, ‎we proposed and tested an alternative metric‎, ‎called multistep‎, ‎which proved more effective in estimating profit potential and identifying the most suitable forecasting model for energy storage operations‎.

‎Our study contributes to the literature by linking predictive model performance to realized economic outcomes in an operational setting‎. ‎It demonstrates that univariate SSA‎, ‎despite its computational demands‎, ‎consistently yields the highest arbitrage profits in a real-world application‎. ‎At the same time‎, ‎simpler forecasting models---such as pooled ARIMA configurations and moving averages---proved economically competitive‎, ‎suggesting that model selection for applied forecasting should consider not only statistical accuracy but also the end-use context‎. ‎These findings point to a broader methodological implication‎: ‎standard forecast evaluation metrics‎, ‎such as RMSE‎, ‎may not adequately reflect the financial outcomes of arbitrage trading‎. ‎We show that profit-maximizing decisions often depend more on a model’s ability to capture structural features of the price series---such as local extrema‎, ‎intraday ordering‎, ‎and oscillatory behavior---than on minimizing forecast error alone‎. ‎To address this‎, ‎we propose and validate an alternative evaluation metric that better aligns with economic performance‎. ‎Our results highlight the need for forecast evaluation criteria tailored to decision-oriented contexts‎, ‎providing a way for quantitative researchers to improve model assessment in applied settings‎.