Comparing time series and neural network models of long memory for electricity price forecasting

• Long memory in day-ahead electricity prices is identified and quantified. • Time-series and neural network models capturing long memory are evaluated. • Fractionally integrated time series models outperform other models in all forecasting scenarios. • Results support explicit modeling of long memory for electricity price forecasting. In this study, we compare time-series and neural networks models capturing long memory for electricity price forecasting in the Russian day-ahead market. We identify the presence of long memory in hourly wholesale electricity price series across six regions of the Russian power grid. Using dedicated statistical tests, we confirm strong long-range dependence and estimate the corresponding long memory parameters. To investigate the potential enhancement in forecasting accuracy offered by long-memory models, we implement a set of fractionally integrated time series models, alongside Long Short-Term Memory (LSTM) and Deep Neural Network (DNN) machine learning models. We evaluate forecasting performance using both in-sample and out-of-sample tests: the in-sample evaluation corresponds to one-hour-ahead predictions, while the out-of-sample evaluation simulates actual day-ahead market conditions. In most cases, the Seasonal Autoregressive Fractionally Integrated Moving Average model with calendar regressors (SARFIMAX) outperforms other time series models and neural networks. The best-performing model, SARFIMAX(1,0,0)(0, D ,0) 24 , achieves an average forecasting improvement of approximately 0.5 % compared to the DNN.