Modeling electricity price formation during Poland’s energy transition using market data and sentiment analysis

1 Introduction

The transformation of contemporary energy systems is shaped by a combination of technological, regulatory, and market-related factors. One of the structural challenges accompanying this process is the management of greenhouse gas emissions in the energy sector, including COx and SOx, which are widely discussed in the context of long-term climate dynamics (Anwar et al., 2019; Lashof and Ahuja, 1990). Electricity generation historically based on fossil fuels continues to play a significant role in many power systems, particularly under conditions of growing population, urbanization, and industrial development, which collectively contribute to a sustained increase in electricity demand (West and Marland, 2002). In this context, thermal energy resources such as coal and natural gas remain important elements of electricity supply and price formation, while simultaneously being subject to increasing regulatory and technological pressures associated with emission reduction objectives (Wrigley, 2013; Zou et al., 2021). As a response to these challenges, international climate and energy initiatives have increasingly emphasized the expansion of renewable energy resources and improvements in energy efficiency. At the Climate Change Conference of the Parties (COP28 UAE), held in Dubai in December 2023, commitments to accelerate renewable energy deployment and enhance energy efficiency were endorsed by 133 countries, reflecting a global shift toward structurally reconfigured energy systems. Within the European Union, these commitments have translated into reinforced policy frameworks aimed at increasing the integration of renewable energy resources into power systems while maintaining security of supply and economic stability (COP28 UAE). In parallel, United Nations initiatives highlight the importance of balancing emission reduction efforts with energy affordability and system reliability. In this framework, wind and solar technologies are increasingly treated as integral components of low-emission energy systems, contributing to changes in generation structures, cost dynamics, and operational conditions of electricity markets (United Nations Framework Convention on Climate Change, 2026). Figure 1 illustrates the global fossil fuel-based electricity generation market and projected trends under current resource availability and regulatory constraints.

Energy costs represent an important component of overall economic activity, and their variability has direct implications for electricity markets and system stability. At the same time, rising electricity demand and structural changes in generation portfolios require an increasing share of renewable energy sources, both to support long-term system sustainability and to mitigate cost-related pressures. In countries such as Poland, where electricity generation has historically relied on conventional energy sources, the transition toward renewable-based systems introduces additional economic, infrastructural, and regulatory challenges (Menegaki and Tiwari, 2017). International experience indicates that large-scale deployment of renewable energy technologies, particularly wind power, has already led to substantial structural changes in electricity generation in several major economies. In China and selected European Union member states, renewables account for a significant and growing share of the electricity mix, influencing market operation and price formation mechanisms (Ochoa et al., 2019; Valodka et al., 2015). These developments provide an important comparative background for analyzing national energy transitions within the EU. Within this context, Poland’s energy transition is guided by national and European strategic frameworks (Bluszcz and Manowska, 2021; Manowska, 2021). The Energy Policy of Poland until 2040 (PEP 2040) remains the formally binding long-term strategy for the energy sector, while an updated National Energy and Climate Plan (Krajowy Plan w dziedzinie Energii i Klimatu, KPEiK) is being procedurally advanced in 2025 in response to revised EU climate and energy targets, although its final adoption has not yet been completed (Gawlik and Mokrzycki, 2019; NoMonday.pl Council, 2040). Figure 2 presents the pictorial analysis of the various resources used for the electricity generation in Poland.

Fluctuations in fossil-fuel and carbon prices directly affect electricity-price dynamics, making their behavior an important subject of analysis in energy-transitioning systems. Beyond observable market fundamentals, energy-related news may contain information concerning regulatory developments, supply disruptions, geopolitical risks, and changing market expectations. Textual analysis therefore provides a complementary method for quantifying the tone and thematic intensity of the information environment surrounding electricity markets. In this study, news articles identified through a GDELT-assisted collection procedure are analyzed using two separate natural-language-processing branches. TF–IDF is employed to examine the thematic prominence of selected energy and policy terms, whereas VADER is applied directly to article text to measure news tone.

In this study, sentiment refers to the tone of selected energy- and climate-related news articles rather than to a representative measure of public opinion. The corpus does not directly observe consumer attitudes, trader expectations, social-media reactions, or policymaker behavior. Accordingly, the CCNI is interpreted as a news-based information proxy that may contain incremental predictive information, but not as a direct measure of societal sentiment. Throughout the manuscript, the terms “sentiment,” “news tone,” and “CCNI” therefore refer to the polarity of the selected news corpus and not to representative public attitudes or directly observed market-participant expectations.

Despite extensive research on electricity-price forecasting, comparatively limited attention has been paid to whether news tone and thematic attention provide incremental predictive information beyond observable market fundamentals. The objective of this study is therefore to evaluate, within an observational forecasting framework, whether text-derived indicators improve the prediction of Polish electricity-price dynamics. The study does not seek to identify a causal effect of news sentiment on prices and does not treat the textual corpus as a representative measure of public opinion.

To address this objective, the study constructs separate TF–IDF-based thematic indicators and a VADER-based news-tone index and combines them with fuel, carbon, and financial-market variables. OLS and Elastic Net models are compared using chronological out-of-sample forecasts. The contribution of the text-derived indicators is evaluated by comparing otherwise identical specifications with and without the textual variables.

Although the broader methodological motivation concerns electricity systems with increasing renewable penetration, the empirical application is Poland, which remained predominantly fossil-fuel-based during much of the study period. Poland is therefore interpreted as an energy-transitioning system rather than as an already renewable-dominated system.

2 Literature review

Existing literature addresses electricity price dynamics, energy transition processes, and associated socio-economic, technological, and informational factors from multiple perspectives. A growing body of research highlights the role of public perception, uncertainty, and societal expectations in shaping energy market outcomes, particularly during periods of geopolitical instability and supply disruptions. Several studies focus explicitly on consumer perceptions and expectations regarding energy prices. In (Alberini et al., 2023), a survey-based analysis conducted across four EU member states—Czech Republic, Germany, Italy, and Poland—examined public expectations under conditions of supply disruptions and rising energy prices related to the Russia–Ukraine conflict. The results indicated that respondents anticipated price increases comparable to or exceeding inflation, while simultaneously expressing uncertainty regarding future price trajectories. Complementary insights into public sentiment were provided in (Vrana et al., 2023), which analyzed social media discourse during the 2022–2023 winter energy crisis using multilingual Twitter data. The authors identified energy prices and political events as dominant discussion topics, with prevailing negative emotions such as fear and sadness. In a related context (Nawrocki et al., 2023), examined publicly available data to assess innovation-related sentiment and performance among major Polish energy companies, demonstrating that firms increasing their share of renewable energy sources exhibited more favorable innovation outcomes. A substantial body of literature investigates structural energy transition processes in Poland and other EU member states. The study in (Jorge-Vazquez et al., 2024) analyzed the direction and pace of energy transition in Spain and Poland between 1990 and 2020 using taxonomic methods and descriptive statistics, revealing an acceleration aligned with EU climate and energy policies. The economic implications of Poland’s coal sector transformation were examined in (Bórawski et al., 2023), which documented a significant reduction in coal mines, rising production costs, and declining employment levels using time-series analysis and autoregressive modeling. Broader assessments of energy dependence and resource availability were provided in (Bluszcz et al., 2023), emphasizing Poland’s historical reliance on coal and the gradual shift toward cleaner energy sources. Similar transition-oriented analyses were presented in (Manowska et al., 2024c), (Bluszcz et al., 2024), (Rusin and Wojaczek, 2023), and (Wojtaszek et al., 2024), which collectively examined energy mix evolution, policy frameworks, infrastructure modernization, and long-term decarbonization strategies in Poland and selected EU countries. Interactions between electricity markets, fossil fuel prices, and broader economic conditions have also received considerable attention. In (Liu et al., 2023), a double-layer spillover framework was proposed to analyze interdependencies among European electricity, fossil fuel, and carbon markets, revealing bidirectional spillover effects intensified by geopolitical events. The study in (Bohdan et al., 2023) focused on natural gas price formation in Poland, identifying strong dependencies on international market prices, domestic extraction volumes, and regulatory mechanisms such as tariff setting. Pricing mechanisms and benefit-sharing models in EU electricity markets were reviewed in (Mello et al., 2023), while (Trabelsi et al., 2023) employed copula-based models to quantify dependence structures between carbon futures and energy commodities, highlighting implications for market stability and policy design.

Technological and infrastructural aspects of energy systems have been investigated alongside market and policy considerations. The work in (Manowska et al., 2024a) examined cybersecurity enhancements in power transmission systems, emphasizing the role of advanced information technologies and biometric authentication in improving system resilience. Renewable energy technologies were addressed in (Manowska et al., 2024b), which analyzed photovoltaic system performance optimization, while environmental policy instruments were evaluated in (Rybak et al., 2022) using ARMAX-based models to assess the effectiveness of environmental taxes in reducing greenhouse gas emissions. Consumption patterns and energy usage trends among Polish households were examined in (Gajdzik et al., 2023), revealing increasing electricity and natural gas consumption alongside declining coal use. Recent studies have increasingly explored data-driven and machine learning approaches in energy market analysis. In (Shen et al., 2023), sentiment analysis combined with LSTM networks was applied to investigate the impact of investor sentiment on stock returns in China’s energy sector, demonstrating stronger effects for renewable energy companies. Methodological foundations for sentiment analysis were further discussed in (Mao et al., 2024) and (Principles of Sentiment Analysis, 2023), while (Manowska and Bluszcz, 2022) demonstrated the effectiveness of LSTM models for forecasting crude oil consumption under nonlinear and nonstationary conditions. Figure 3 illustrates Poland’s electricity import and export trends within the EU since 2016, highlighting increasing cross-border electricity exchange and growing integration with European energy markets (Venizelou and Poullikkas, 2024). Figure 4 presents key contributing factors commonly considered in sentiment analysis studies related to energy markets (Mao et al., 2024; Principles of Sentiment Analysis, 2023). Figure 5 shows the annual average electricity prices in Poland over the period 2016–2024, providing additional context for price variability under conditions of energy system transformation (Poland: electricity prices for households, 2024; Electricity price statistics, 2016).

Despite the extensive literature addressing energy-transition pathways, market interdependencies, technological development, and electricity-pricing mechanisms, limited evidence is available on whether separate measures of thematic news attention and news tone provide incremental predictive information beyond conventional market variables. Existing studies also do not always distinguish clearly between media tone, public opinion, market expectations, predictive association, and causal influence. The present study addresses these gaps through a country-specific application to the Polish electricity market.

Recent research broadens the measurement of information and sentiment beyond conventional lexicons. Koy et al. (2026) use Google Trends as an investor-attention proxy and document nonlinear relationships with cryptocurrency outcomes. VADER provides a deterministic and auditable rule-based benchmark for English-language text (Hutto and Gilbert, 2014), whereas contextual transformer approaches can represent domain-specific meaning more effectively. Lee and Anderl (2025), for example, apply business-news sentiment in an energy-market prediction setting, and Ren (2025) emphasises source representativeness, validation, and the distinction between media discourse and public opinion in energy research. These studies motivate the use of VADER as a transparent baseline rather than as the most accurate available method.

3 Materials and methods

This study follows a multi-stage research design integrating selected policy-news information with market-based energy, carbon, and financial variables. First, news articles are collected, screened, and processed. Second, two conceptually distinct text-derived indicators are constructed: TF–IDF-based thematic measures and a VADER-based news-tone index. Third, these indicators are aligned with the market variables and used in OLS and Elastic Net forecasting specifications. Finally, chronological out-of-sample performance is evaluated to determine whether the text-derived indicators provide incremental predictive information beyond the market-control variables. The framework is observational and predictive; it does not identify a causal effect of news tone on electricity prices.

3.1 Data sources and preprocessing

The presented research focuses on transforming information extracted from online news sources into quantitative variables that support the analysis of electricity price formation under the EU Emissions Trading System (ETS). To this end, the Global Database of Events, Language and Tone (GDELT) is employed as the primary source of textual data. Although GDELT aggregates multilingual news from a broad range of countries and publishers, the final analytical corpus is determined by the study-specific filtering and source-screening procedures. Consequently, the representativeness of the CCNI depends on the retained articles rather than on the full coverage of the GDELT database. The resulting index should therefore be interpreted as reflecting the selected European policy-news corpus and not the complete European information environment or Polish public opinion (Leetaru and Schrodt). This database has been widely used in previous studies aiming to extract sentiment-related indicators using natural language processing techniques for economic and energy market analyses (Alamro et al., 2019; Galla et al., 2018; Guidolin and Pedio, 2021). In addition to textual data, market-based variables relevant to the ETS were incorporated into the analysis. Daily return series based on closing prices of European Union Allowance (EUA) futures (EUR/ton) were obtained from ember-climate.org and included as a supporting variable due to their role in reflecting carbon pricing dynamics within the ETS framework. Furthermore, additional explanatory variables related to energy and financial markets were considered, as summarized in Table 1 (Böhringer et al., 2022).

TABLE 1. Different Resource prices, own study based on (Böhringer et al., 2022).

TABLE 2. Variance inflation factors for the return-space predictors, source: own study.

Table 1 presents the market variables used in the econometric analysis, including electricity, natural gas, coal, crude oil, carbon allowances, and the European equity-market indicator. The underlying series differ in their starting dates, trading calendars, and data availability; consequently, the nominal source period does not imply that a complete observation is available for every calendar day. The Polish day-ahead electricity-price series contains 3,217 calendar-day observations from 2 January 2016 to 22 October 2024. The electricity, fuel, carbon, equity, and text-derived series were subsequently aligned using a common calendar-date key. Daily logarithmic returns and the required forecasting lags were then constructed, and observations with missing values in the target or predictors required for model estimation were removed. After these alignment, transformation, lag-construction, and complete-case filtering steps, 1,011 observations remained in the final common econometric sample. Thus, the 1,011 observations do not represent the total number of calendar days originally available; they represent the complete common-date observations for which all variables required by the forecasting models were simultaneously available. Figure 6 illustrates the overall workflow of the proposed sentiment-analysis and electricity-price forecasting framework, including data collection, text processing, index construction, market-data alignment, and econometric modelling.

For the primary forecasting exercise, the final analytical sample was divided chronologically rather than randomly. All complete observations preceding 31 January 2023 formed the initial estimation sample, while the primary out-of-sample evaluation covered 31 January 2023 to 22 October 2024 and produced 448 one-day-ahead forecasts. The estimation sample was subsequently expanded as new observations became available. The 50- and 100-observation windows reported in Table 3 are terminal sensitivity windows used to assess the incremental contribution of the text-derived variables and are not separate randomly selected test samples.

TABLE 3. Out-of-sample forecast comparison across the TF–IDF, market-control, and full specifications.: own study.

The market variables represent economic and financial conditions associated with Polish electricity prices. The specification does not directly include load forecasts, wind and solar generation forecasts, residual load, outages, congestion, curtailment, or weather conditions. The text-derived variables are therefore treated as supplementary forecasting inputs rather than substitutes for these physical and operational predictors. Their empirical contribution is assessed conditional on the selected market controls and not relative to a fully specified operational electricity-price forecasting model.

3.2 Sentiment extraction and index construction

The textual-data collection process relied on the Global Database of Events, Language and Tone (GDELT) to identify event-linked news URLs associated with European Union actors and energy-related developments. GDELT provides large-scale news coverage and has been employed in previous economic, financial, and energy-market studies to construct text-derived information indicators (Leetaru and Schrodt; Alamro et al., 2019; Galla et al., 2018; Guidolin and Pedio, 2021). Candidate URLs were filtered for European relevance and subsequently screened according to source reliability. The Media Bias/Fact check news, 2026 Check factual-reporting classification was used as an additional source-screening criterion, and articles originating from sources assessed as doubtful or unreliable were excluded (Media Bias/Fact check news, 2026). The exact retrieval parameters, thematic terms, screening rules, and extraction procedure are reported in Supplementary Appendix SA.

The resulting corpus is interpreted as a selected European policy-news dataset rather than as a representative measure of public opinion. Although GDELT aggregates content from multiple countries and languages, the final analytical corpus is determined by the study-specific geographical, topical, accessibility, and source-screening procedures. The Climate Change News Index (CCNI) therefore reflects the tone of the retained news coverage and does not directly observe the attitudes of consumers, traders, policymakers, or the general population.

The textual analysis was divided into two separate and conceptually distinct branches. In the first branch, article text was converted to lowercase and cleaned through the removal of punctuation, non-alphabetic characters, and English stop words. The text was then tokenised and lemmatised. Term Frequency–Inverse Document Frequency (TF–IDF) was applied to quantify the relative prominence of selected energy, emissions, fuel, renewable-energy, and policy terms. TF–IDF is commonly used in text mining to reduce the influence of frequently occurring but weakly informative terms and to assign greater weight to terms that distinguish individual documents within a corpus (Mao et al., 2024; Principles of Sentiment Analysis, 2023).

For term j in document d, the TF–IDF weight is defined as in Equation 1.where TF j , d denotes the frequency of term j in document d, N is the total number of documents, and DF j is the number of documents containing the term. The resulting TF–IDF values were used to evaluate thematic attention and to construct the bipolar keyword representation presented in Figure 7. They were not used as inputs to VADER.

TFIDF j , d = TF j , d ⁡ ln N DF j (1)

In the second branch, the Valence Aware Dictionary and Sentiment Reasoner (VADER) was applied directly to the cleaned but un-lemmatized article bodies. Retaining the original linguistic structure for VADER allowed its lexical, negation, intensification, and punctuation rules to operate on text closer to the retrieved article content. Each article received a continuous compound score bounded between − 1 and + 1, where more negative values indicate negative news tone and more positive values indicate positive news tone.

The daily CCNI was calculated as the arithmetic mean of the compound scores of the articles published on day t in Equation 2:where s i , t denotes the VADER compound score of article i published on day t, and N t is the number of scored articles available on that date. The continuous compound score was retained for index construction. Positive, negative, and neutral classifications were used for descriptive corpus analysis according to the polarity thresholds documented in Supplementary Appendix SA.

C C N I t = 1 N t ∑ i = 1 N t s i , t , (2)

The CCNI measures the average polarity of the retained news articles. It does not automatically incorporate the number of articles published on a given date. Article volume was therefore treated separately as a corpus-coverage indicator and is reported in Supplementary Figure A2. This distinction prevents changes in publication intensity from being interpreted incorrectly as changes in sentiment polarity.

The TF–IDF thematic variables and the daily CCNI were aligned with the market variables through an explicit calendar-date key. Previous studies have similarly combined textual information with energy, commodity, carbon, and financial-market variables (Böhringer et al., 2022; Media Bias/Fact check news, 2026; Wagner et al., 2010; Coyne et al., 2017; Lubis et al., 2021; Mittermayer, 2004; Nikfarjam et al., 2010). The resulting merged dataset was used in the econometric forecasting framework described in Section 3.5.

3.3 Corpus construction and filtering

The initial retrieval produced 6,348 candidate documents. A sequential quality-control procedure was applied before the construction of the CCNI. First, 23 duplicate URLs were removed. One record with an unparsable publication date was then excluded. A further 818 records published before 1 January 2010 were removed to maintain the intended analysis period. Subsequently, 119 documents whose retrieved body text was empty or contained fewer than 20 tokens were excluded. The 20-token rule represents a study-specific quality-control threshold intended to remove empty, incomplete, or non-substantive webpage content; it is not presented as a universal requirement for valid VADER scoring. Finally, nine documents that produced no VADER lexicon match were removed.

After these filtering stages, 5,378 documents remained in the final CCNI corpus, representing 84.7% of the retrieved candidate pool. The complete filtering cascade is reported in the corpus-construction table in Supplementary Appendix SA, while the extraction parameters, 49-term topical dictionary, thematic grouping, and preprocessing operations are documented in Sections A.1–A.8.

The corpus is concentrated in English-language, Brussels-oriented policy journalism, with Politico Europe accounting for the predominant share of the accessible articles. Accordingly, the CCNI should not be interpreted as a comprehensive EU-wide media index or as a direct measure of Polish-language public sentiment. It represents the tone of a selected and influential European policy-news source covering energy, climate, fuel, and regulatory developments.

This source concentration is particularly important for interpreting the Polish application. EU-level policy reporting may contain information relevant to Polish electricity prices through emissions regulation, fuel-market disruptions, energy-security decisions, and European market reforms. Nevertheless, Polish-language media could generate different thematic and polarity patterns because it may place greater emphasis on domestic tariffs, household electricity bills, mining-sector employment, district heating, national political debates, and country-specific regulatory interventions. This limitation is considered explicitly in Sections 5 and 6.

3.4 Choice of sentiment lexicon

VADER was selected as a transparent and reproducible baseline rather than as the most accurate available method for professional energy-policy journalism. Its principal methodological advantage is that it is deterministic and can be applied without a labelled training corpus. Its scoring process is also inspectable, allowing the contribution of individual lexical terms and linguistic rules to be examined. These properties support reproducibility and comparability with earlier lexicon-based sentiment studies (Hutto and Gilbert, 2014).

Nevertheless, VADER was originally developed and calibrated primarily for social-media-style English text. Its valence dictionary and rules for intensification, punctuation, capitalisation, and negation may be less effective for professionally edited policy journalism, in which tone is often communicated through framing, conditional statements, technical terminology, and implicit evaluation. VADER may also misinterpret domain-specific expressions such as “negative electricity prices,” “price support,” “emissions cap,” “capacity reduction,” or “fuel-price correction,” because their technical meanings may differ from their general lexical polarity.

The model is also limited in its ability to identify sarcasm, implicit stance, long-range contextual relationships, and subtle differences between the author’s position and a quoted statement. Its performance may differ across languages and translated material. Consequently, the CCNI is treated as an interpretable lexicon-based news-tone indicator rather than as a complete representation of editorial stance or market expectations.

Contextual transformer models such as BERT, RoBERTa, FinBERT, and multilingual XLM-R can potentially capture domain-specific and context-dependent meanings more effectively. Recent studies have applied transformer-derived news sentiment to energy-market forecasting and have emphasised the importance of source representativeness and model validation (Lee and Anderl, 2025; Ren, 2025). Such approaches, however, require appropriate labelled data, domain adaptation, validation against human annotations, and greater computational resources. The use of VADER in the present study is therefore justified by transparency, auditability, computational efficiency, and comparability, while a domain-adapted transformer comparison is identified as an important direction for future research.

3.5 Econometric modeling framework

The econometric analysis focuses on modeling electricity price dynamics using sentiment-based and market-related explanatory variables within a regression-based framework. Electricity prices were transformed into daily logarithmic returns in order to reduce non-stationary and stabilize variance, which is a common practice in energy and financial market studies (Weron, 2006; Higgs and Worthington, 2008). The dependent variable is defined as in Equation 3:where P t denotes the electricity price at time t, and r_t represents the corresponding daily return.

r t = ln P t − ln P t − 1 (3)

The baseline regression model is specified as in Equation 4:where S_t denotes the sentiment-based index constructed in Section 3.2, EU A_t represents the return on European Union Allowance futures, X_t is a vector of additional control variables including fuel prices and financial market indicators, β is a vector of parameters to be estimated, and ε_t is an error term.

r t = β 0 + β 1 S t + β 2 E U A t + β 3 T X t + ε t (4)

As a benchmark approach, the model parameters were first estimated using Ordinary Least Squares (OLS). OLS provides an interpretable baseline for assessing the relationship between electricity price returns and explanatory variables and is widely applied in empirical energy economics (Wooldridge, 2016). The OLS estimator is given by Equation 5:where X denotes the matrix of explanatory variables and y is the vector of observed electricity price returns.

β ^ = X T X − 1 X T y (5)

To account for potential multicollinearity among explanatory variables and to improve model robustness, an Elastic Net regression approach was additionally employed. Elastic Net combines the properties of LASSO and Ridge regression through joint L₁ and L₂ regularization and has been shown to perform well in high-dimensional and correlated predictor settings, including energy market applications (Zou and Hastie, 2005; Ziel et al., 2015; Lahmiri, 2016). The Elastic Net estimator is obtained by solving the following optimization problem Equation 6:where N denotes the number of observations, λ is the regularization parameter controlling the overall penalty strength, and α ϵ 0 , 1 determines the balance between LASSO and Ridge penalties.

β ^ = a r g m i n β 1 / 2 N Σ i = 1 N y i − x i T β 2 + λ α β 1 + 1 − α β 2 2 (6)

Both modeling approaches were applied using the same set of explanatory variables to ensure comparability. Model estimation and subsequent performance evaluation were conducted using out-of-sample validation procedures described in Section 3.4.

3.5.1 Forecasting and validation procedure

All forecasting experiments were done in the chronological order of the observations. The final complete econometric sample contains 1,011 observations after calendar alignment, return construction, lag generation, and removal of incomplete rows. The primary out-of-sample evaluation covers 448 forecast observations from 31 January 2023 to 22 October 2024. At each forecast origin, the model is estimated using all complete observations available before the forecasted observation, after which the estimation window is expanded by the newly observed value. The 50- and 100-observation terminal windows reported separately are sensitivity checks used to examine the incremental contribution of the text variables and should not be interpreted as separate primary test samples.

The forecasting design uses one-period lags to prevent look-ahead bias. To forecast the electricity-price return for day, the explanatory information consists of values available by day: the lagged electricity-price return, lagged TTF natural-gas return, Rotterdam-coal return, Brent-oil return, EUA-carbon return, STOXX Europe 600 return, lagged CCNI, and lagged TF–IDF thematic-attention variables. Thus, no predictor dated after the forecast origin enters model estimation or prediction.

For Elastic Net, the regularization hyperparameters were re-selected at each model refit using five-fold non-shuffled cross-validation conducted exclusively within the corresponding estimation sample. The mixing parameter was selected from the candidate set. For each candidate mixing value, the overall regularization parameter was evaluated over a 100-point geometrically spaced grid generated from the current estimation sample, with the upper bound defined by, the penalty at which all penalized coefficients are shrunk to zero. The coordinate-descent solver used a convergence tolerance of, a maximum of 5,000 iterations, and a fixed random seed of 42. Hyperparameters were re-estimated at each of the 23 refits. Across these refits, the selected mixing parameter was 0.10 in every case, while the selected had a median of 0.1089 and ranged from 0.0769 to 0.1546. Importantly, the forecast observation was excluded from model fitting and hyperparameter selection at every refit. Predictor standardization parameters were estimated from the corresponding estimation sample and then applied to the forecast observation, preventing information from the held-out forecast point from entering the fitting procedure.

The models are estimated in logarithmic-return space. The observed price at is therefore used as the price-level anchor. MAE, RMSE, MAPE, and reported for the price forecasts are calculated after this inverse transformation and are consequently expressed at the electricity-price level rather than in return space.

3.6 Multicollinearity diagnostics

The fundamental price series display substantial correlation in levels, with pairwise correlations between approximately 0.58 and 0.84 among oil, coal, carbon, and natural gas as presented in Table 2. Because the forecasting models are estimated using daily returns, multicollinearity is assessed using the transformed variables. In return space, the predictors exhibit weak pairwise correlations: the largest correlation is 0.23, between carbon allowances and the equity index. All variance inflation factors are below 1.08, and the condition number of the standardized design matrix is 1.33. These diagnostics indicate that severe multicollinearity is not present in the transformed predictor set.

Multicollinearity does not by itself make OLS coefficients biased under exogeneity, but it may increase coefficient variance and reduce estimate stability. The low VIF and condition-number values therefore reduce concerns about instability caused by correlated predictors. They do not establish causality, eliminate omitted-variable bias, or rule out other forms of model misspecification. Elastic Net is retained as a regularized predictive model rather than as a procedure that converts estimated coefficients into structural effects.

3.7 Evaluation metrics

The predictive performance of the econometric models was evaluated using complementary measures commonly applied in electricity-price forecasting (Hyndman and Koehler, 2006; Weron, 2014). The metrics were calculated from reconstructed electricity-price levels rather than directly from logarithmic returns. MAE measures the average absolute forecast error, whereas RMSE gives greater weight to large errors and is therefore particularly relevant for spike-prone electricity prices.

The Mean Absolute Error is defined as in Equation 7:

M A E = 1 N ∑ t = 1 N ∣ P t − P ^ t ∣ . (7)

The Root Mean Squared Error in Equation 8:

R M S E = 1 N ∑ t = 1 N P t − P ^ t 2 . (8)

The Mean Absolute Percentage Error in Equation 9:

M A P E = 100 N ∑ t = 1 N ∣ P t − P ^ t P t ∣ . (9)

MAPE was calculated using reconstructed positive price levels rather than logarithmic returns because percentage errors are not appropriate for variables that may be negative or close to zero.

The Pearson correlation coefficient is defined in Equation 10:

ρ P , P ^ = C o v P , P ^ σ P σ P ^ . (10)

The metrics were interpreted jointly because improvement in one measure does not necessarily imply improvement in all aspects of forecasting performance. In particular, a reduction in MAE may coexist with an increase in RMSE when a model improves typical forecasts but performs less effectively during extreme observations.

Where forecast-loss series were compared, Diebold–Mariano tests were used to examine whether the mean out-of-sample loss differential was statistically distinguishable from zero. The persistence comparison evaluates the complete forecasting model, whereas the market-control-versus-full comparison isolates the incremental contribution of the text-derived indicators.

4 Results

This section presents the empirical results obtained from the sentiment analysis and econometric modeling framework described in Section 3. The results are organized to first characterize the temporal dynamics and internal structure of the constructed sentiment index, followed by an examination of the thematic content underlying positive and negative sentiment. Subsequently, the performance of sentiment-augmented econometric models is assessed and compared with benchmark specifications, and the implications for electricity price dynamics in the Polish energy market are analyzed. The results are reported in a descriptive and comparative manner, focusing on observable patterns and quantitative performance measures, while their broader interpretation and policy implications are discussed in Section 5.

4.1 Sentiment dynamics and properties of the constructed index

Figure 8 presents the monthly mean CCNI, the corresponding standard-error band, and the 12-month moving average. The compound score is bounded between − 1 and + 1, with higher values indicating more positive average news tone and lower values indicating more negative average news tone. Monthly aggregation is used for visualisation, whereas the forecasting models use the daily CCNI.

The retained corpus is positive on average over much of the sample, but its mean tone declines substantially from early 2022 and remains comparatively subdued through 2023. This timing coincides with the European energy crisis, although the co-movement is descriptive and does not establish causality. The monthly series displays substantial variation in both its central tendency and dispersion. Periods of greater dispersion coincide with intensified policy-news coverage concerning energy prices, emissions trading, geopolitical developments, fuel-market disruptions, and regulatory intervention. The increasing number of observations in later years reflects changes in corpus coverage and digital-news availability and should be distinguished from changes in sentiment polarity.

The CCNI measures the average polarity of the retained news articles and does not automatically incorporate the number of articles published on a given date. Article volume is therefore treated separately as a corpus-coverage indicator and is reported in Supplementary Figure A2. Distinguishing tone from volume prevents changes in publication intensity from being interpreted incorrectly as changes in sentiment polarity.

Figure 9 presents the aggregate numbers of articles classified as positive, negative, and neutral. Positive classifications constitute the largest category in the complete corpus, followed by negative and neutral classifications. This aggregate imbalance does not imply that positive news tone dominates every subperiod, because the monthly composition changes over time, as illustrated by Figure 8. The distribution is therefore interpreted as a descriptive characteristic of the retained corpus rather than as evidence of balanced sentiment categories.

4.2 Keyword-based structure of positive and negative sentiment

Figure 7 presents the keyword-based structure underlying the constructed sentiment index using a bipolar representation of TF–IDF-weighted terms. Keywords are positioned along a common horizontal axis according to their mean TF–IDF weights computed separately for documents classified as positive and negative. This representation allows for a direct comparison of the thematic profiles associated with opposing sentiment polarities within the energy-related news corpus.

The results indicate a clear and systematic differentiation between the content driving positive and negative sentiment. Terms located on the positive side of the bipolar scale are predominantly associated with forward-looking narratives related to renewable energy deployment, energy transition strategies, policy instruments, and system modernization. These keywords reflect informational content emphasizing structural transformation, long-term planning, and regulatory support mechanisms within the energy sector.

In contrast, keywords positioned on the negative side of the scale are primarily linked to conventional energy carriers, fuel price dynamics, regulatory uncertainty, and short-term market pressures. The prominence of these terms suggests that negative sentiment is largely driven by news related to price volatility, supply constraints, and external risk factors affecting energy markets.

The spatial separation of keywords across the bipolar axis demonstrates that sentiment polarity is not randomly distributed but is instead aligned with distinct thematic clusters. Positive sentiment is characterized by narratives centered on stability, development, and transformation, whereas negative sentiment is associated with risk-oriented and disruption-related discourse. This structured differentiation supports the internal coherence of the constructed sentiment index and confirms that sentiment decomposition captures substantive differences in information content rather than superficial linguistic variation.

Overall, the bipolar keyword representation indicates that articles assigned positive and negative VADER scores contain different thematic profiles. This provides internal descriptive validation that the polarity classifications correspond to meaningful differences in textual content. However, the thematic separation does not establish that the CCNI improves electricity-price forecasts or causally affects market outcomes. The forecasting contribution of the text-derived indicators is evaluated separately in Section 4.3.

4.3 Econometric model performance

This subsection reports the empirical performance of the econometric models used to forecast electricity price returns, with particular emphasis on the contribution of sentiment-enhanced variables. Model performance is evaluated using out-of-sample forecasts and standard accuracy measures, allowing for a direct comparison between baseline specifications and models incorporating sentiment-based information.

Table 3 compares the out-of-sample forecasting performance of the TF–IDF-only, market-control, and full specifications across the 50- and 100-observation test windows. The lower error levels reported for the 100-observation window should be interpreted as descriptive differences between the two evaluation samples rather than as direct evidence that a longer test window improves parameter stability, because the windows contain different observations and market conditions.

The incremental contribution of the text-derived variables is small and depends on the model, evaluation metric, and test window. For the 50-observation window, the full specification reduces MAE from 22.64 to 22.15 for OLS and from 22.69 to 22.20 for Elastic Net relative to the market-control specification. These changes correspond to improvements of approximately 2.16% in both cases. However, RMSE increases from 30.03 to 30.15 for OLS and from 30.01 to 30.33 for Elastic Net, indicating that the inclusion of the text-derived variables does not improve the prediction of larger errors in this window.

For the 100-observation window, all differences between the full and market-control specifications remain below approximately 0.3%. The OLS full specification produces a slightly higher MAE and RMSE than the control model. The Elastic Net full specification produces a marginally lower MAE, while its RMSE is unchanged. The results therefore indicate limited and specification-dependent incremental predictive information rather than systematic forecasting superiority of the text-enhanced models.

The approximately 2.16% reduction in MAE should be interpreted as a modest improvement in predictive accuracy rather than as evidence of operationally significant value. The present study does not conduct a bidding, dispatch, balancing-cost, hedging, or trading-value back test, and therefore does not quantify the economic consequences of this reduction in forecast error. The result is therefore treated as limited and metric-dependent incremental predictive information.

Figure 10 compares observed and predicted electricity prices during the expanding-window evaluation period. The predicted series captures the broad central tendency and medium-term direction of prices but is visibly smoother than the observed series. Several abrupt upward and downward movements are underestimated, and the predicted-versus-actual panel indicates compression toward the center of the price distribution. The figure therefore illustrates both the usefulness and the limitations of the conditional-mean specification, particularly during extreme-price episodes.

Over the 448-day expanding-window evaluation from 31 January 2023 to 22 October 2024, the full Elastic Net specification attains an MAE of 55.57 PLN/MWh, an RMSE of 76.71 PLN/MWh, a MAPE of 14.34%, and an R 2 of 0.622. Relative to persistence, the complete Elastic Net specification produces modest but statistically detectable improvements. However, this comparison evaluates the complete model and does not isolate the contribution of the text-derived variables. The incremental text contribution is evaluated through the market-control-versus-full comparison in Table 3 and remains small and metric-dependent. The framework is therefore presented as a complementary analytical tool rather than as a stand-alone operational forecasting system.

Figure 11 separates the return series into individually scaled panels and presents their rolling-volatility behavior. Polish electricity returns exhibit substantially greater dispersion and more frequent extreme observations than the fuel, carbon, and equity-market series. The rolling-volatility panel indicates substantial time variation and clustering in electricity-price volatility. The shaded period identifies the European energy crisis for descriptive orientation and does not establish a causal effect of the crisis on any individual series.

Further insights into price dynamics are provided in Figure 11, which illustrates the evolution of electricity prices alongside prices of selected energy resources over the period 2020–2024. The results highlight pronounced co-movement between electricity prices and fossil fuel prices, particularly natural gas and coal. Periods of elevated fuel prices coincide with increased electricity price levels, reflecting the continued importance of conventional energy carriers in price formation despite the growing share of renewable energy sources.

Figure 12 reports the continuous return series for electricity and related energy resources. Figure 12 provides a combined representation of the continuous return series. Most observations are concentrated near zero, while extreme electricity-price movements dominate the common scale. The figure illustrates the heavy-tailed and spike-prone character of electricity returns and the clustering of extreme observations during periods of market stress. It should be interpreted together with the individually scaled panels in Figure 11.

Overall, the empirical results indicate that text-derived information produces small and specification-dependent changes in forecasting accuracy. The evidence does not support systematic improvement across all models, metrics, and evaluation windows. The CCNI and TF–IDF indicators are therefore interpreted as complementary information variables whose predictive contribution is limited and conditional on the forecasting specification.

5 Discussion

A fundamental electricity-price forecasting depends on load, wind and solar forecasts, residual load, fuel and carbon costs, outages, interconnection constraints, and the merit order. The present forecasting is additional variables. Its narrower contribution is to represent the contemporaneous policy-news environment surrounding capacity mechanisms, support schemes, grid rules, and emission policy. This is relevant during renewable transition because market design and flexibility requirements evolve alongside the physical system.

Cross-correlations with daily electricity returns at leads and lags of up to ten trading days remain below 0.003 in absolute value, providing no evidence of useful linear leading power in the tested daily specification. This does not rule out nonlinear, monthly, or event-specific relationships, but those relationships are not established here. The workflow may be transferable to other markets, whereas the numerical results are not: market design, generation mix, interconnection, language, media structure, and crisis exposure differ across countries. Operational use would also require load and renewable forecasts, decision-based economic back testing, and a broader local-language information set.

6 Limitation and future research

This study has some limitations. First, its observational design does not identify a causal effect of news tone on electricity prices. Second, the CCNI reflects the tone of selected policy-news articles rather than representative public opinion, observed trader expectations, or behavioral responses. Fourth, VADER is a lexicon- and rule-based model originally developed for social-media-style English text and may not fully capture contextual polarity, sarcasm, technical terminology, implicit evaluation, or multilingual variation. Fifth, the forecasting gains attributable specifically to the text variables are small and metric-dependent. Sixth, OLS and Elastic Net model the conditional mean and do not fully represent time-varying volatility, asymmetry, and extreme price spikes. Seventh, the single-market and crisis-dominated sample limits external validity. Future research should compare VADER with domain-adapted transformer models, construct Polish-language and EU-level indices separately, incorporate physical system variables, and combine conditional-mean forecasts with nonlinear or GARCH-type volatility models.

7 Conclusion

This study evaluates whether thematic attention and news tone provide incremental information for forecasting Polish electricity prices. The proposed framework separates TF–IDF-based thematic analysis from VADER-based news-tone measurement and combines these indicators with fuel, carbon, and financial-market variables in OLS and Elastic Net models.

The corpus analysis indicates that positive and negative news classifications contain different thematic structures. Positive news tone is associated mainly with investment, transition, policy incentives, clean technologies, and system-resilience narratives, whereas negative news tone is associated mainly with fuel shortages, price volatility, emissions costs, import dependence, regulatory uncertainty, and geopolitical risk.

Market fundamentals remain the principal predictive inputs. The contribution of the text-derived indicators is small and specification-dependent. Some MAE values improve marginally, while RMSE remains unchanged or increases slightly in several specifications. The findings therefore indicate limited complementary predictive information rather than systematic forecasting superiority.

The reported relationships are associative and predictive rather than causal. The CCNI should not be interpreted as representative Polish public opinion, a direct measure of trader expectations, or a verified leading indicator of regulatory change. It is more appropriately understood as a proxy for the selected European policy-news environment.

The workflow may be applied to other electricity markets, but the numerical findings are specific to Poland, its generation structure, market institutions, fuel exposure, selected news corpus, and crisis-intensive sample period. Future research should incorporate Polish-language sources, domain-adapted transformer models, physical power-system predictors, probabilistic forecasting, and dedicated nonlinear and conditional-volatility specifications.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Generative AI tools were used to support language editing, improve clarity and coherence of the manuscript. The use of generative AI was limited to editorial and linguistic support and did not influence the research design, data collection, data analysis, interpretation of results, or scientific conclusions. All content was reviewed, verified, and approved by the authors, who take full responsibility for the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fenrg.2026.1780848/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Solomon Giwa, Olabisi Onabanjo University, Nigeria

Reviewed by: Andaç Batur Çolak, Niğde Ömer Halisdemir University, Türkiye

Atif Maqbool Khan, AGH University of Science and Technology, Poland