Predicting Wine Quality Under Changing Climate: An Integrated Approach Combining Machine Learning, Statistical Analysis, and Systems Thinking

Abstract

Climate change poses significant challenges for viticulture, particularly in regions known for producing high-quality wines. Wine quality results from a complex interaction between climatic factors, regional characteristics, and viticultural practices. Methods: This study integrates statistical analysis, machine learning (ML) algorithms, and systems thinking to assess the extent to which wine quality can be predicted using monthly weather data and regional classification. The dataset includes average wine scores, monthly temperatures and precipitation, and categorical region data for Slovenia between 2011 and 2021. Predictive models tested include Random Forest, Support Vector Machine, Decision Tree, and linear regression. In addition, Causal Loop Diagrams (CLDs) were constructed to explore feedback mechanisms and systemic dynamics. Results: The Random Forest model showed the highest prediction accuracy (R² = 0.779). Regional classification emerged as the most influential variable, followed by temperatures in September and April. Precipitation did not have a statistically significant effect on wine ratings. CLD models revealed time delays in the effects of adaptation measures and highlighted the role of perceptual lags in growers’ responses to climate signals. Conclusions: The combined use of ML, statistical methods, and CLDs enhances understanding of how climate variability influences wine quality. This integrated approach offers practical insights for winegrowers, policymakers, and regional planners aiming to develop climate-resilient viticultural strategies. Future research should include phenological phase modeling and dynamic simulation to further improve predictive accuracy and system-level understanding.

1. Introduction

Wine quality is a complex and multidimensional concept resulting from the interplay of natural and human factors. It represents a fundamental indicator of market value, regional identity, and the sustainable distinctiveness of wine-growing areas [1,2]. While traditionally assessed through sensory and oenological parameters, an increasing number of studies are employing analytical methods to better understand the influence of environmental and geographic factors on wine quality and to develop predictive models [3]. Climatic conditions, particularly temperature and precipitation, play a vital role in grapevine development and wine quality. Their impact is increasingly relevant under current climate change trends, which challenge the stability of viticultural systems [4,5]. High-quality wine regions are especially vulnerable, as even minor thermal deviations can lead to undesirable shifts in sugar–acid balance, aromatic expression, and phenolic profiles [6]. Regional characteristics, often conceptualized under the term terroir, also play a key explanatory role. This concept encompasses climatic, soil, topographical, cultural, and technological attributes [7]. Due to its layered and composite nature, terroir is difficult to quantify and therefore rarely included in predictive models, even though regional classification can serve as a proxy for numerous hard-to-measure influences. In our study, we operationalized terroir using a categorical regional variable, acknowledging this as a reductionist approach. This choice balances the trade-off between capturing the explanatory richness of terroir and maintaining the interpretability and parsimony of predictive models.

Data-driven approaches, such as machine learning, have become essential tools in capturing complex, nonlinear relationships in agroecological systems. Recent studies [8] have successfully used machine learning models based on growing degree days (GDD) and regional climate data to predict grapevine phenological stages and wine characteristics, emphasizing the importance of biologically aligned climatic metrics. Moreover, GDD-based modeling has been successfully used to assess climate change impacts not only on grape phenology, but also on pest dynamics in vineyards, as demonstrated by Comte et al. (2024) [9]. Methods such as Random Forest, Support Vector Machines (SVM), and neural networks have proven effective in predicting sensory quality, ripening stages, and yields [10]. Nevertheless, few studies have simultaneously incorporated climatic variables, regional features, and methodological diversity, which would be necessary to more accurately interpret the drivers of wine quality.

Complementing these approaches is the use of systems thinking, which facilitates the understanding of interdependent processes, feedback loops, and time delays within viticultural systems [11,12,13]. Causal Loop Diagrams (CLDs) can help visualize interactions that are not easily captured by quantitative models and support the interpretation of why certain variables—such as region—emerge as consistently strong predictors.

Despite these advances, a clear methodological gap remains: most existing studies treat influencing factors in isolation and rarely integrate quantitative and systemic perspectives. The aim of this research is to bridge that gap by applying machine learning, statistical analyses, and systems modeling to explore to what extent wine quality can be predicted based on climatic variables and regional characteristics. The case of Slovenia—with its pronounced regional diversity and long-standing winemaking tradition—provides an ideal context for this type of integrative analysis.

2. Literature Review

Wine quality is a fundamental concept in viticultural and oenological science, reflecting both the objective characteristics of wine and the subjective preferences of consumers. In the scientific literature, wine quality is commonly defined as a combination of physicochemical parameters, sensory properties, and perceived market value [14,15,16]. In practice, quality is most often assessed using standardized scoring systems (e.g., the 20-point OIV scale, the 100-point Parker scale), where experts evaluate the wine based on color, aroma, taste, complexity, and balance [17].

Treating wine quality as a dependent variable in research presents a particular challenge, as it is not determined by a single factor, but rather arises from the interaction between grape variety, winemaking techniques, agroecological conditions, and market expectations [18,19,20,21]. The grape variety defines the genetic potential for aromatic profile, tannin content, and acid structure [22], while winemaking practices—from harvest timing to fermentation procedures—significantly shape the expression of that potential [23].

Environmental factors, particularly climatic conditions, influence the vine’s phenological stages, the synthesis of primary and secondary metabolites, and microbiological activity in the vineyard [24]. A comprehensive review by Williamson et al. (2024) [25] highlighted how diverse climate metrics—such as GDD, diurnal range, and heat stress indices—affect grapevine biophysical responses, calling for more precise temporal alignment of data with phenological stages. In recent decades, quality as perceived by consumers has also become a key part of the scientific understanding of wine, incorporating elements such as origin labels, sustainability certifications, price, and narratives about the region and producer [26,27,28,29,30,31,32].

Due to the complexity and multilayered nature of the quality construct, researchers often approximate wine quality using average scores from sensory panels, market ratings, or a combination of chemical and sensory indicators as a proxy for the dependent variable [33,34].

Climatic variables, particularly temperature and precipitation, are considered key ecological factors that influence wine quality through their impact on grapevine growth, development, and metabolic processes. An optimal average temperature during the growing season (between 16 °C and 22 °C) is associated with a balanced concentration of sugars, acids, and aromatic compounds in grapes [35,36,37]. Studies have shown that elevated temperatures during the ripening period—especially in August and September—accelerate sugar accumulation but often lead to reduced acidity and overripe phenolic profiles, which can affect the freshness and aging potential of the wine [38,39]. Precipitation exerts a more complex and indirect influence. While moderate rainfall prior to harvest can be beneficial, excessive precipitation increases the risk of fungal diseases and juice dilution [40]. Climatic anomalies—such as early spring frosts, summer heatwaves, and prolonged droughts—are strongly associated with vintage variability in wine quality, particularly in regions where grapevine microclimatic tolerance is narrow [41,42,43,44].

The concept of terroir refers to the combination of natural and cultural factors that define the character of wine from a specific geographical location. It encompasses climate, microclimate, soil composition, topography, grape variety distribution, cultivation practices, and historical traditions, all of which are expressed in the unique sensory profile of the wine [45,46,47,48,49,50]. Due to its multidimensional nature, terroir is extremely difficult to quantify directly, which often leads to its exclusion from predictive models. Nevertheless, studies have shown that even the simple inclusion of regional affiliation as a categorical variable can capture important environmental and socio-cultural influences that significantly contribute to explaining differences in wine quality across viticultural zones [51,52,53].

Machine learning (ML) methods are increasingly used in viticulture to predict complex phenomena such as wine quality, yield, disease occurrence, and grape ripening. Their key advantage lies in the ability to detect nonlinear and high-dimensional patterns that traditional statistical approaches often fail to capture [54,55,56]. In recent years, various ML architectures have been developed for wine quality prediction, particularly using Random Forest, Support Vector Machines (SVMs), and Artificial Neural Networks (ANNs).

Over the past decade, there has been a marked rise in the application of artificial intelligence and ML methods for predicting the sensory characteristics of wine, including overall quality. For instance, Tiwari [57] developed a mathematical model to predict the quality of Pinot Noir wines based on chemical properties, validating its performance using machine learning algorithms. The model achieved a high degree of alignment with professional tasters’ scores. Similarly, Jain [58] analyzed publicly available physicochemical data from red wines and applied algorithms such as Random Forest and XGBoost to achieve high prediction accuracy, identifying key predictors such as alcohol content, volatile acidity, and pH.

A literature review by Cardoso Schwindt [59] further emphasizes the importance of integrating chemical descriptors with sensory analysis. Methods such as Quantitative Structure–Odor Relationships (QSORs) are opening new avenues for a more refined quantitative understanding of complex aromatic profiles in wine. A common thread among these studies is the drive to develop models that incorporate multimodal data—chemical, sensory, agroclimatic, and geographic—in order to provide a more holistic prediction of wine quality. Despite these promising developments, few studies to date have simultaneously integrated climatic variables, regional characteristics, and mixed-method approaches (e.g., ML combined with statistics), highlighting a critical gap in the current scientific literature.

Although quantitative approaches such as statistical analysis and machine learning models are essential for predicting complex agroecological phenomena, they do not always capture the full range of dynamic interactions, feedback loops, and time delays that characterize viticultural systems. To address this gap, systems thinking—particularly through the use of Causal Loop Diagrams (CLDs)—has gained increasing traction in recent years. These diagrams offer a qualitative visualization of complex processes, where climatic influences, plant physiology, and human practices are interwoven into interdependent and often delayed effects [60,61,62,63,64,65,66]. In viticulture, such an approach is particularly useful for explaining why region may emerge as the most powerful predictive factor, as it encompasses microclimatic, soil-related, and socio-economic elements that are not fully captured by individual variables. As a result, there is a growing body of research that employs methodological synthesis, combining machine learning, statistical testing, and systems modeling—either to enhance predictive performance or to provide a more nuanced explanation of observed patterns [67,68,69,70]. These hybrid approaches enable a more holistic understanding of viticultural systems by linking empirical findings with the system’s dynamic logic. In doing so, they open new pathways for more effective climate adaptation strategies and the strategic development of high-quality wine regions. While ML methods offer powerful predictive tools, previous studies often face notable limitations. Small or regionally restricted datasets increase the risk of overfitting and limit generalizability. Moreover, vineyard heterogeneity—in terms of microclimate, management, and varietals—is difficult to reflect with aggregated data. As highlighted by Colibaba et al. [8] and Comte et al. [9], achieving robust predictions requires high-resolution, phenologically aligned data and better integration of domain knowledge. These challenges emphasize the need for careful validation and contextual interpretation in ML-based viticulture research.

The following section presents the methodology of this study, which is based on a combination of quantitative and qualitative analytical approaches.

3. Materials and Methods

The aim of this study was to assess to what extent wine quality—expressed as the average wine score—can be predicted based on climatic variables and regional characteristics using machine learning methods complemented by statistical analyses. To achieve this, we compiled a comprehensive dataset that includes average wine scores, detailed meteorological data, and geographical features of individual Slovenian wine-growing regions.

3.1. Wine Scores and the Dependent Variable

Data on wine quality were derived from more than 72,000 wine sample evaluations from vintages produced between 2011 and 2021. The scores were obtained from the Agricultural and Forestry Institute of Maribor (Kmetijsko gozdarski zavod Maribor), which conducts professional wine assessments as part of annual quality evaluations. For the purpose of this study, we used average annual wine scores aggregated by region, which allowed for direct comparison with meteorological and geographical data.

The average wine score (hereafter referred to as avg. score) served as the dependent variable in both statistical and predictive models.

Table 1. Descriptive statistics—average temperature and average wine score.

Variable	N	Range	Minimum	Maximum	Mean	Std. Deviation	Variance
avg. T [°C] January	93	14.6	−5.2	9.4	2.437	2.6803	7.184
avg. T [°C] February	93	12.8	−3.0	9.8	3.973	2.9404	8.646
avg. T [°C] March	93	7.5	3.3	10.8	7.804	1.984	3.936
avg. T [°C] April	93	6.8	8.5	15.3	12.29	1.424	2.028
avg. T [°C] May	93	6.5	12.8	19.3	15.937	1.5123	2.287
avg. T [°C] June	93	5.2	19.1	24.3	21.108	1.3636	1.859
avg. T [°C] July	93	5.9	20.0	25.9	22.856	1.3781	1.899
avg. T [°C] August	93	6.0	18.7	24.7	22.172	1.4957	2.237
avg. T [°C] September	93	6.8	13.9	20.7	17.282	1.5681	2.457
avg. T [°C] October	93	6.7	9.2	15.9	12.303	1.7097	2.923
avg. T [°C] November	93	10.4	2.6	13.0	8.092	2.1157	4.476
avg. T [°C] December	93	8.9	−0.1	8.8	3.389	1.9382	3.933
Avg. Score	93	1.464819611	16.09227053	17.55709014	17.05754438	0.3269151454	0.107

presents the descriptive statistics of average temperatures alongside the average wine score, while

Table 2. Descriptive statistics—average precipitation.

Variable	N	Minimum	Maximum	Mean	Std. Deviation
precipitation [mm] January	93	8.7	214.5	69.744	41.0083
precipitation [mm] February	93	2.6	322.9	66.485	59.6276
precipitation [mm] March	93	8.1	311.7	95.719	75.6143
precipitation [mm] April	93	0.8	305.6	59.376	59.924
precipitation [mm] May	93	8.7	214.5	70.117	40.6936
precipitation [mm] June	93	25.4	265.9	119.65	53.1897
precipitation [mm] July	93	3.4	199.5	90.333	52.08
precipitation [mm] August	93	5.2	264.7	92.744	56.7478
precipitation [mm] September	93	2.0	230.3	89.205	65.4562
precipitation [mm] October	93	27.9	300.9	137.978	73.9767
precipitation [mm] November	93	19.4	360.4	115.977	68.8845
precipitation [mm] December	93	0.0	452.4	136.898	103.4163

displays the average amount of precipitation.

Table 1 and Table 2 provide an overview of the descriptive statistics for wine scores and monthly climatic variables. Notably, the standard deviation of wine scores remains relatively low, suggesting moderate interannual variation, while temperature variables exhibit expected seasonal fluctuations. These initial distributions support the use of both parametric and non-parametric modeling approaches.

Temperature and precipitation data were obtained from the publicly available archive of the Slovenian Environment Agency (ARSO). The dataset includes average monthly temperatures and total monthly precipitation from January to December for each year individually. Additionally, data were aggregated by season (spring, summer, autumn, winter), which enabled a seasonal analysis of the impact of climatic factors on wine quality.

For each region, data from the nearest meteorological station were used, as presented in

Table 3. Wine-growing district, wine region, corresponding meteorological stations, and categorical variable assignment.

Number–Categorical Variable	Wine-Growing District	Wine-Growing Region	Weather Station
1.	Štajerska Slovenija	Podravje	Maribor
2.	Prekmurje	Podravje	Murska Sobota
3.	Bizeljsko-Sremič	Posavje	Cerklje
4.	Dolenjska	Posavje	Novo mesto
5.	Bela krajina	Posavje	Novo mesto
6.	Vipavska dolina	Primorska	Bilje
7.	Goriška brda	Primorska	Bilje
8.	Kras	Primorska	Bilje
9.	Slovenska Istra	Primorska	Portorož

and the corresponding map in Figure 1.

Average values indicated the highest precipitation levels during the autumn months (September, November) and the highest temperatures during the summer period (June–August). The greatest variability in precipitation was recorded during the winter months.

Climatic variables were treated as independent variables and were included in statistical correlation and regression models, as well as in machine learning algorithms.

3.2. Geographical Classification of Wine-Growing Regions

As shown in Table 3, Slovenia is divided into three major wine regions: Podravje, Posavje, and Primorska. Each of these includes several wine-growing districts. In this study, data were aggregated at the level of nine distinct regions (as listed in the table). Each region was assigned a numerical code (1–9), which was used as a categorical variable in subsequent statistical analyses and machine learning models.

In addition to regional classification, we also considered the temporal dimension—the production year (2011–2021)—which enabled the identification of long-term climatic trends.

Table 4. Summary of key variables.

Variable Type	Name in Models	Description
Dependent variable	pov. ocena	Average annual wine rating for each region
Independent variables	temp_1–12, pad_1–12	Monthly temperatures and precipitation (January–December)
	temp_sez, pad_sez	Seasonal temperatures and precipitation (spring, summer, autumn, winter)
Categorical variable	regija (1–9)	Wine-growing area
Time variable	leto	Year of wine production

summarizes additional key variables used in the analysis. The period 2011–2021 was selected based on complete data availability and was not filtered for extreme events, allowing natural climatic variability to be reflected in the analysis.

This methodological framework enables the integration of climatic, geographical, and quality-related data, serving as the foundation for the subsequent application of predictive models and statistical analyses presented in the following sections.

To examine the relationship between climatic variables and wine scores, we employed a range of statistical methods using the SPSS software (version 28). These methods included distribution checks, trend analysis, and tests for regional differences, as well as correlation and factor analyses. Following the descriptive analysis of average monthly precipitation and temperatures for the period 2011–2021—presented in the introductory section—we conducted normality tests. In the results section, we also present the findings obtained through machine learning methodologies and the CLD (Causal Loop Diagram) model.

CLDs were primarily developed as qualitative tools to synthesize and visualize the key feedback structures, delays, and dynamic interactions present in viticulture systems affected by climate change. These diagrams are grounded in both the relevant literature and the patterns revealed by our empirical analyses [36,60,64,65].

It is worth emphasizing that the ability to capture and represent such causal relationships is also a key advantage of contemporary artificial intelligence approaches, especially large language models (LLMs), which are increasingly able to encode, infer, and utilize causal structures in complex systems [72]. The inclusion of causality not only supports robust explanation and prediction but also enhances the interpretability and applicability of AI-driven analyses in domains such as climate-resilient viticulture.

3.3. Statistical Analysis: Software and Methods

To explore the relationships between climatic conditions and average wine scores, several statistical analyses were conducted using IBM SPSS Statistics. The distribution of data was assessed with the Kolmogorov–Smirnov and Shapiro–Wilk tests. Long-term temperature trends were examined using a General Linear Model (GLM), while regional differences in temperature were analyzed with one-way ANOVA and Tukey’s HSD post hoc test. Differences in wine scores across regions were tested using the non-parametric Kruskal–Wallis test. Correlations between monthly climatic variables and wine scores were calculated using Spearman’s rho and Kendall’s tau-b coefficients. To identify underlying patterns among seasonal variables, a Principal Component Analysis (PCA) with Varimax rotation was performed, and the resulting components were included in a multiple regression model to assess their predictive value. These procedures provided a robust statistical basis for interpreting the impact of climate variability on wine quality.

4. Results

4.1. Results of Statistical Methodological Steps Performed in the SPSS Program

4.1.1. Checking the Normality of the Distribution

To test the assumptions of normality, we applied the Kolmogorov–Smirnov and Shapiro–Wilk tests (

Table 5. Normality of data distribution—temperature. Tests of Normality.

Variable	KS a Statistic	KS a df	KS a Sig.	SW Statistic	SW df	SW Sig.
precipitation [mm] January	0.153	93	<0.001	0.881	93	<0.001
precipitation [mm] February	0.164	93	<0.001	0.74	93	<0.001
precipitation [mm] March	0.155	93	<0.001	0.889	93	<0.001
precipitation [mm] April	0.205	93	<0.001	0.758	93	<0.001
precipitation [mm] May	0.156	93	<0.001	0.878	93	<0.001
precipitation [mm] June	0.12	93	0.002	0.957	93	0.004
precipitation [mm] July	0.11	93	0.008	0.963	93	0.010
precipitation [mm] August	0.142	93	<0.001	0.932	93	<0.001
precipitation [mm] September	0.157	93	<0.001	0.924	93	<0.001
precipitation [mm] October	0.148	93	<0.001	0.899	93	<0.001
precipitation [mm] November	0.14	93	<0.001	0.877	93	<0.001
precipitation [mm] December	0.155	93	<0.001	0.877	93	<0.001

^a Lilliefors Significance Correction.

and

Table 6. Normality of data distribution—precipitation. Tests of Normality.

Variable	KS a Statistic	KS a df	KS a Sig.	SW Statistic	SW df	SW Sig.
avg. T [°C] January	0.109	93	0.008	0.958	93	0.004
avg. T [°C] February	0.14	93	<0.001	0.948	93	0.001
avg. T [°C] March	0.11	93	0.007	0.937	93	<0.001
avg. T [°C] April	0.104	93	0.015	0.956	93	0.003
avg. T [°C] May	0.085	93	0.091	0.983	93	0.280
avg. T [°C] June	0.107	93	0.011	0.947	93	<0.001
avg. T [°C] July	0.076	93	0.200 *	0.981	93	0.196
avg. T [°C] August	0.104	93	0.015	0.962	93	0.009
avg. T [°C] September	0.115	93	0.004	0.969	93	0.026
avg. T [°C] October	0.06	93	0.200 *	0.974	93	0.060
avg. T [°C] November	0.045	93	0.200 *	0.992	93	0.824
avg. T [°C] December	0.106	93	0.012	0.976	93	0.086
avg. Score	0.154	93	<0.001	0.868	93	<0.001

* This is a lower bound of the true significance. ^a Lilliefors Significance Correction.

). The results indicate that most of the climatic variables do not follow a normal distribution (p < 0.05), which justifies the use of non-parametric tests in subsequent analyses.

The most pronounced deviations from normality were observed in precipitation data for April and December, as well as in temperature data for February. All months showed p-values < 0.05 in at least one of the tests (Kolmogorov–Smirnov or Shapiro–Wilk), indicating that precipitation data for none of the months follow a normal distribution.

4.1.2. Trend Analysis (GLM)

To determine whether weather conditions have changed significantly over time, we conducted a General Linear Model (GLM) analysis with year as the independent variable. The results indicate statistically significant changes in temperature over time, particularly in the spring and summer months (April, May, June), confirming the effects of climate change.

To explore whether temperatures have changed systematically over time, we applied multivariate tests to assess the temporal effect of year on monthly temperature patterns. The results strongly indicate a consistent warming trend, particularly during spring and early summer months. The results of the multivariate tests (Pillai’s Trace, Wilks’ Lambda, Hotelling’s Trace, and Roy’s Largest Root) (

Table 7. Results of multivariate tests. Multivariate Tests ^a.

Effect	Test	Value	F	Hypothesis df	Error df	Sig.	Partial Eta Squared
Intercept	Pillai’s Trace	1.000	58,056.833 b	12.000	71.000	<0.001	1.000
Intercept	Wilks’ Lambda	0.000	58,056.833 b	12.000	71.000	<0.001	1.000
Intercept	Hotelling’s Trace	9812.422	58,056.833 b	12.000	71.000	<0.001	1.000
Intercept	Roy’s Largest Root	9812.422	58,056.833 b	12.000	71.000	<0.001	1.000
Year	Pillai’s Trace	7.972	26.208	120.000	800.000	<0.001	0.797
Year	Wilks’ Lambda	0.000	118.032	120.000	565.767	<0.001	0.920
Year	Hotelling’s Trace	251.369	144.956	120.000	692.000	<0.001	0.962
Year	Roy’s Largest Root	91.882	612.549 c	12.000	80.000	<0.001	0.989

^a Design: Intercept + Year. ^b Exact statistic. ^c The statistic is an upper bound on F that yields a lower bound on the significance level.

) show that the variable year has a statistically significant effect on monthly temperatures, as all tests yielded p-values < 0.001. The high values of Partial Eta Squared (ranging from 0.797 to 0.989) further confirm that the effect of year on temperature variation is very strong, indicating that average monthly temperatures changed significantly over the analyzed period.

In addition, the analysis of between-subjects effects (

Table 8. Analysis of between-subjects effects. Tests of Between-Subjects Effects.

Source	Dependent Variable	Type III Sum of Squares	df	Mean Square	F	Sig.	Partial Eta Squared
Corrected Model	avg. T [°C] January	396.033 a	10	39.603	12.259	<0.001	0.599
Corrected Model	avg. T [°C] February	565.252 b	10	56.525	20.137	<0.001	0.711
Corrected Model	avg. T [°C] March	253.036 c	10	25.304	19.018	<0.001	0.699
Corrected Model	avg. T [°C] April	153.828 d	10	15.383	38.535	<0.001	0.825
Corrected Model	avg. T [°C] May	160.876 e	10	16.088	26.639	<0.001	0.765
Corrected Model	avg. T [°C] June	125.114 f	10	12.511	22.367	<0.001	0.731
Corrected Model	avg. T [°C] July	88.716 g	10	8.872	8.485	<0.001	0.508
Corrected Model	avg. T [°C] August	98.504 h	10	9.85	7.527	<0.001	0.479
Corrected Model	avg. T [°C] September	78.214 i	10	7.821	4.339	<0.001	0.346
Corrected Model	avg. T [°C] October	131.229 j	10	13.123	7.815	<0.001	0.488
Corrected Model	avg. T [°C] November	160.845 k	10	16.084	5.256	<0.001	0.391
Corrected Model	avg. T [°C] December	118.156 l	10	11.816	3.976	<0.001	0.327
Intercept	avg. T [°C] January	1224.664	1	1224.664	378.661	<0.001	0.643
Intercept	avg. T [°C] February	1224.464	1	1224.464	378.654	<0.001	0.842
Intercept	avg. T [°C] March	5004.539	1	5004.539	3761.365	<0.001	0.979
Intercept	avg. T [°C] April	12,750.539	1	12,750.539	8367.54	<0.001	0.997
Intercept	avg. T [°C] May	2132.681	1	2132.681	856.5	<0.001	0.998
Intercept	avg. T [°C] June	4125.384	1	4125.384	935.684	<0.001	0.999
Intercept	avg. T [°C] July	4325.881	1	4325.881	1136.423	<0.001	0.998
Intercept	avg. T [°C] August	41,127.188	1	41,127.188	31,428.609	<0.001	0.997
Intercept	avg. T [°C] September	25,320.764	1	25,320.764	1404.753	<0.001	0.994
Intercept	avg. T [°C] October	12,144.59	1	12,144.59	7392.857	<0.001	0.989
Intercept	avg. T [°C] November	5189.539	1	5189.539	1695.892	<0.001	0.954
Intercept	avg. T [°C] December	962.627	1	962.627	323.913	<0.001	0.798
Year	avg. T [°C] January	396.033	10	39.603	12.259	<0.001	0.599
Year	avg. T [°C] February	565.252	10	56.525	20.137	<0.001	0.711
Year	avg. T [°C] March	253.036	10	25.304	19.018	<0.001	0.699
Year	avg. T [°C] April	153.828	10	15.383	38.535	<0.001	0.825
Year	avg. T [°C] May	160.876	10	16.088	26.639	<0.001	0.765
Year	avg. T [°C] June	125.114	10	12.511	22.367	<0.001	0.731
Year	avg. T [°C] July	88.716	10	8.872	8.485	<0.001	0.508
Year	avg. T [°C] August	98.504	10	9.85	7.527	<0.001	0.479
Year	avg. T [°C] September	78.214	10	7.821	4.339	<0.001	0.346
Year	avg. T [°C] October	131.229	10	13.123	7.815	<0.001	0.488
Year	avg. T [°C] November	160.845	10	16.084	5.256	<0.001	0.391
Year	avg. T [°C] December	118.156	10	11.816	3.976	<0.001	0.327
Error	avg. T [°C] January	264.902	82	3.232
Error	avg. T [°C] February	230.171	82	2.807
Error	avg. T [°C] March	109.102	82	1.331
Error	avg. T [°C] April	32.733	82	0.399
Error	avg. T [°C] May	49.52	82	0.604
Error	avg. T [°C] June	45.951	82	0.56
Error	avg. T [°C] July	86.013	82	1.049
Error	avg. T [°C] August	107.304	82	1.309
Error	avg. T [°C] September	147.824	82	1.803
Error	avg. T [°C] October	137.7	82	1.679
Error	avg. T [°C] November	250.96	82	3.06
Error	avg. T [°C] December	243.693	82	2.972
Total	avg. T [°C] January	1213.06	93
Total	avg. T [°C] February	2263.49	93
Total	avg. T [°C] March	6026.5	93
Total	avg. T [°C] April	14,234.0	93
Total	avg. T [°C] May	23,289.97	93
Total	avg. T [°C] June	41,605.14	93
Total	avg. T [°C] July	48,757.26	93
Total	avg. T [°C] August	45,924.56	93
Total	avg. T [°C] September	28,001.22	93
Total	avg. T [°C] October	14,346.28	93
Total	avg. T [°C] November	6502.2	93
Total	avg. T [°C] December	1430.14	93
Corrected Total	avg. T [°C] January	660.936	92
Corrected Total	avg. T [°C] February	795.423	92
Corrected Total	avg. T [°C] March	362.138	92
Corrected Total	avg. T [°C] April	186.561	92
Corrected Total	avg. T [°C] May	210.396	92
Corrected Total	avg. T [°C] June	171.065	92
Corrected Total	avg. T [°C] July	174.729	92
Corrected Total	avg. T [°C] August	205.807	92
Corrected Total	avg. T [°C] September	226.039	92
Corrected Total	avg. T [°C] October	268.929	92
Corrected Total	avg. T [°C] November	361.805	92
Corrected Total	avg. T [°C] December	361.849	92

Legend: ^a R Squared = 0.599 (Adjusted R Squared = 0.550) ^b R Squared = 0.711 (Adjusted R Squared = 0.675) ^c R Squared = 0.699 (Adjusted R Squared = 0.662). ^d R Squared = 0.825 (Adjusted R Squared = 0.803). ^e R Squared = 0.765 (Adjusted R Squared = 0.736). ^f R Squared = 0.731 (Adjusted R Squared = 0.699). ^g R Squared = 0.508 (Adjusted R Squared = 0.448). ^h R Squared = 0.479 (Adjusted R Squared = 0.415). ⁱ R Squared = 0.346 (Adjusted R Squared = 0.266). ^j R Squared = 0.488 (Adjusted R Squared = 0.426). ^k R Squared = 0.391 (Adjusted R Squared = 0.316). ^l R Squared = 0.327 (Adjusted R Squared = 0.244).

) revealed that year has a statistically significant impact on temperatures in all months, with p-values < 0.001 for each individual month. The highest explained variance (R²) values were observed for April (0.825), May (0.765), and June (0.731), suggesting that these months experienced the most pronounced temperature changes over time. The lowest R² value was recorded for December (0.327), indicating that this winter month showed the least variation in temperature over the years compared to other periods. These findings suggest that spring and summer months are the most affected by warming trends.

4.1.3. Spatial Differences Between Regions (ANOVA and Kruskal–Wallis)

To examine differences in climatic variables between regions, we conducted a one-way analysis of variance (ANOVA), supplemented by Tukey’s HSD post hoc test (Figure 2). The analysis revealed significant regional differences, with Goriška Brda (Region 7) and Slovenska Istra (Region 9) standing out for their higher temperatures.

The results of post hoc analyses revealed significant temperature differences between geographic regions in most months, with the most pronounced differences observed between Region 2 (the coldest) and Region 9 (the warmest). During winter months such as January and February, temperatures in Region 2 were significantly lower than those in Region 9. In March, Regions 2, 4, and 6 recorded significantly lower temperatures compared to Regions 8 and 9. No statistically significant differences were found between regions in transitional months such as April and May. In June, regions showed highly diverse temperature profiles—Region 2 had the lowest and Region 9 the highest temperatures. Similar patterns were observed in July and August, where Regions 2, 4, and 6 had significantly lower temperatures compared to Regions 7 and 9. In autumn, September temperatures were lowest in Regions 2, 4, and 6 and highest in Regions 8 and 9, while the most notable October difference was observed between Regions 2 and 9. In November, Regions 2, 5, and 6 were significantly colder than Regions 8 and 9. December also exhibited sharp contrasts, with the lowest temperatures in Region 2 and the highest in Region 9.

The results of the Tukey HSD tests confirm that statistically significant differences in average monthly temperatures exist between regions in most months, with the exception of April and May. The largest differences were observed during the summer months (June, July, August), where Regions 7 and 9 were notably warmer than the others.

To test for differences in wine scores across regions, the Kruskal–Wallis test was conducted (

Table 9. Testing differences in wine scores between regions. Test Statistics ^a,b.

Region_Code	N	Mean Rank
BEL. KRAJ.	10	74.5
BIZ. SREM.	10	24.1
DOL.	10	5.5
GOR BRDA	10	59.6
KRA	10	41.3
PREKM	11	52.09
SLO. ISTRA	10	55.6
ŠT. SLO.	11	68.36
VIP. DOL.	11	40.0
Total	93
Test		Kruskal–Wallis H
Value		52.884
Degrees of Freedom (df)		8
Asymptotic Significance (Asymp. Sig.)		<0.001

^a Kruskal–Wallis Test. ^b Grouping Variable: Region_Code.

). The results indicate that wine scores are not evenly distributed across regions (Asymp. Sig. < 0.001), highlighting the importance of local (microclimatic) factors.

4.1.4. Correlation Analysis

To examine the relationship between monthly climatic variables (temperature and precipitation) and the average wine score, we used Spearman’s correlation and Kendall’s tau-b, which are suitable for ordinal data and do not require a normal distribution.

Relationship Between Temperature and Wine Score

The results of the correlation analysis showed that certain temperature variables in specific months are statistically significantly associated with the average wine score. The strongest positive association was observed for September (Spearman’s ρ = 0.601; p < 0.001; Kendall’s τ = 0.466; p < 0.001), suggesting that higher temperatures during the grape ripening period contribute significantly to wine quality. Moderate positive correlations were also found in April (ρ = 0.291; p = 0.005; τ = 0.310; p < 0.001) and December (ρ = 0.308; p = 0.003; τ = 0.258; p = 0.003), indicating the influence of early spring development and the end of the vegetative cycle.

In addition, May and November also showed statistically significant associations with wine quality, although the correlations were somewhat weaker (ρ ≈ 0.2–0.3; p < 0.05). These findings confirm that temperature conditions during key phenological stages play an important role in shaping the final quality of wine.

Relationship Between Precipitation and Wine Score

The analysis of the impact of precipitation on wine score did not reveal any statistically significant associations. Spearman’s and Kendall’s correlation coefficients were generally low, ranging between −0.2 and 0.2, and p-values were mostly above the threshold of statistical significance (p > 0.05). The highest, yet still non-significant, correlations were observed in November and December (ρ ≈ 0.15–0.17), indicating a possible weak but inconsistent relationship.

These results suggest that monthly precipitation alone is not a reliable predictor of wine quality. Any potential influence of precipitation may only manifest indirectly or in interaction with other variables—such as high temperatures, humidity, or time-sensitive phenological phases—as also supported by the findings of the factor and regression analyses. A similar conclusion was drawn in the study by Jones and Goodrich [70] who found that while climate variability, including precipitation, can influence vintage quality in regions such as Napa Valley, precipitation patterns alone (e.g., winter or ripening period rainfall) were not consistently associated with higher or lower wine ratings. Instead, the combination of temperature-related variables and teleconnection patterns (e.g., PDO phases) played a more decisive role in explaining vintage variability [67].

4.1.5. Factor Analysis of Climatic Patterns

To identify latent patterns among climatic variables, we performed a Principal Component Analysis (PCA) using Varimax rotation (

Table 10. Factor analysis of climatic patterns. (A) Rotated Component Matrix ^a. (B) Component Transformation Matrix.

(A)
Variable	Component 1	Component 2	Component 3
avgT_summer	0.243	0.221	0.831
avgT_autumn	0.635	0.543	0.424
avgT_winter	0.540	0.576	0.103
avgT_spring	0.223	0.804	0.221
winter_precipitation	0.856	−0.001	0.053
spring_precipitation	0.866	0.070	−0.142
summer_precipitation	0.299	0.162	−0.777
autumn_precipitation	−0.248	0.772	−0.317
(B)
Component	1	2	3
1	0.770	0.574	0.279
2	0.259	0.119	−0.959
3	−0.583	0.810	−0.057

Extraction Method: Principal Component Analysis. Rotation Method: Varimax with Kaiser Normalization. ^a Rotation converged in six iterations.

). The analysis included eight variables: average seasonal temperatures and precipitation (spring, summer, autumn, winter). The aim was to identify meaningful clusters of variables that could better summarize key climatic influences and be used in subsequent regression analysis.

Three principal components were extracted, together explaining 75.5% of the total variance:

• Factor 1: Winter and Spring Influence

High loadings: avgT_autumn (0.919), avgT_winter (0.775), winter_precipitation (0.673), spring_precipitation (0.667)

Interpretation: Influences from the colder part of the year and early vegetation development

• Factor 2: Seasonal Transitions

High loadings: avgT_spring, autumn_precipitation

Interpretation: Patterns characteristic of spring and autumn variability

• Factor 3: Summer Influence (“Hot and Dry Summer”)

Highest loadings: avgT_summer (0.831), summer_precipitation (−0.777)

Interpretation: Combination of high temperatures and low precipitation during the summer period.

Subsequently, the extracted factors were used in a multiple regression analysis (Figure 3) to assess their predictive power for the dependent variable—the average wine score. The results showed the following:

• The overall model was not statistically significant (R² = 0.058; F(3,89) = 1.836; p = 0.146)
• Only Factor 3 (“hot and dry summer”) was a statistically significant predictor (β = 0.078; p = 0.023)

This indicates that higher temperatures and lower precipitation during the summer months may modestly but significantly contribute to higher wine scores. The other two factor patterns did not exhibit meaningful predictive power.

4.2. Machine Learning Models

To complement traditional statistical approaches and assess the predictive power of climatic variables and regional characteristics, we applied machine learning (ML) methods in this study. These approaches enable the modeling of complex and nonlinear relationships among multiple input variables, which are often present when analyzing the effects of climate conditions on wine quality.

4.2.1. Data Preparation and Software Environment

The analyses were conducted using the open-source platform Orange 3.38.1, which allows for easy data integration, model development, and visual comparison of different algorithms. The input dataset included average monthly temperatures (12 variables), average monthly precipitation (12 variables), and geographical region as a categorical variable (encoded 1–9). The target variable was the average wine score by region and vintage, obtained from the KGZS Maribor database.

All quantitative variables were pre-standardized, and categorical variables were appropriately encoded. To validate the models, we used stratified 10-fold cross-validation.

4.2.2. Algorithms Used

Four machine learning models were included in the analysis, representing varying degrees of complexity and interpretability:

• Random Forest—an ensemble model based on multiple decision trees,
• Decision Tree—a single decision tree, useful for rule explanation,
• Support Vector Machine (SVM)—an efficient model for handling high-dimensional data,
• Linear Regression—a baseline model for comparison with ML approaches.

4.2.3. Modeling Results

A comparison of model performance (

Table 11. Scores—sampling type—stratified 10-fold cross-validation.

Model	MSE	RMSE	MAE	MAPE	R2
Random Forest	0.023	0.153	0.119	0.007	0.779
Decision Tree	0.032	0.178	0.133	0.008	0.701
SVM	0.029	0.171	0.134	0.008	0.725

) showed that the Random Forest model performed best across most metrics (e.g., lowest MSE and highest R² = 0.779), followed by Linear Regression (R² = 0.772). The Decision Tree model was slightly less accurate (R² = 0.701), while the SVM model showed the weakest performance (R² = 0.725), likely due to the limited sample size and data complexity.

We specifically examined the impact of including region as a predictive variable (Figure 4). The variable importance analysis within the Random Forest model showed that geographical region had the highest predictive power. This variable surpassed all climatic parameters in importance, indicating the influence of local microclimatic, pedological, and agronomic factors that are implicitly captured by the region.

The analysis of individual variable importance (based on “Decrease in R²”) showed that, in addition to region, the most influential predictors were (1) average temperature in August and September, (2) average temperature in April, and (3) precipitation in May and March. However, the influence of precipitation remained relatively weak, which is consistent with the results of classical correlation analyses. The Random Forest model outperformed the others due to its ability to capture non-linear relationships and complex interactions among variables, which are common in viticultural and climatic systems. Default hyperparameters were used, which performed well in this case. Future studies could explore parameter tuning for further improvement.

4.3. Systems Approach and Development of CLD Models

The quantitative analysis of predictive factors affecting wine quality—based on statistical methods and machine learning—is complemented in this study by a qualitative approach grounded in systems thinking. The aim of this addition was to shed light on complex interactions between environmental influences, biological processes, and growers’ responses, which may appear as hidden or underrepresented effects in predictive models. Both CLDs were constructed based on relevant scientific literature, expert knowledge in viticulture and climate adaptation, and patterns emerging from our statistical and ML analyses. While no formal validation of the feedback structures was performed, we acknowledge this as a methodological limitation.

To this end, we developed two Causal Loop Diagrams (CLDs) that visually illustrate key feedback mechanisms and dynamic relationships within the viticulture system under the influence of climate change.

The first CLD model focuses on vine biomass growth, where global warming acts as an external pressure on key physiological processes such as photosynthesis, transpiration, and ripening. The second model represents vineyard management and growers’ responses to changes in wine quality, highlighting time delays, adaptive mechanisms, and socio-economic influences.

Both models were developed based on system dynamics principles [60], scientific literature on grapevine responses to thermal stress (as reviewed in the literature section), and interpretations of the findings obtained through correlation and predictive analyses in this study.

From a methodological perspective, CLD models allow for the identification of reinforcing and balancing loops in the viticultural system, the understanding of nonlinearities and time delays, and support for the interpretation of statistically observed phenomena that numbers alone cannot explain (e.g., why region as a variable captures more than just average temperature). Integrating a systemic perspective enables a more comprehensive explanation of the effects of climatic factors on wine quality and contributes to the design of more resilient adaptation strategies in practice.

It can be observed that the rate of biomass growth depends on the growth factor, the share of available resources, and the current biomass quantity through a balancing feedback loop (Figure 5). Biomass growth also creates a reinforcing feedback loop, where more biomass leads to increased growth—provided that resources are available. Figure 5 illustrates these systemic interactions through a Causal Loop Diagram (CLD), which highlights the feedback mechanisms between global temperature, biosystem efficiency, and viticultural biomass production. It shows how rising temperatures can alter the efficiency of biological processes and resource availability, ultimately affecting the capacity of the system to sustain biomass growth over time.

A balancing loop is established via environmental carrying capacity, which limits resource availability as biomass increases. Global temperature affects the growth factor, which in turn influences biomass accumulation. This relationship illustrates the system’s sensitivity to climatic conditions. As biomass increases, the share of available resources decreases, since expanding biomass consumes those inputs.

The causal loop diagram in Figure 5 represents the system structure that drives the dynamics of vineyard biomass growth, emphasizing the interconnections between biomass, resource availability, and environmental constraints. The model also incorporates the impact of global temperature on growth, highlighting an ecological and environmental feedback system. Findings indicate that extremely high temperatures—especially those above 35 °C—can significantly disrupt grapevine physiological processes, particularly photosynthesis, thereby affecting the growth and development of vineyard biomass.

In addition to its effects on biomass growth, climate change also plays a critical role in determining wine quality. To better understand these effects and develop suitable strategies to preserve wine quality under changing climatic conditions, we further developed a causal loop diagram (CLD) as shown in Figure 6. The purpose of this model is to explore the interactions between viticultural practices, environmental factors, and their influence on grape ripening and wine quality. The CLD incorporates key variables that affect grape ripening and wine production and illustrates the complex dynamics introduced by climate change.

The development of this model was necessary because it became evident that a purely quantitative analysis of biomass growth dynamics is insufficient for a comprehensive understanding of the impacts of climate change on the entire viticultural system—especially on the final product: wine. The causal loop diagram enables a more holistic approach by linking changes in environmental conditions with vineyard practices and highlighting time delays that can significantly affect outcomes. These observed delays emphasize that changes in certain variables do not produce immediate effects, implying that long-term planning and patience are required to achieve desired results in viticulture.

Figure 6 provides a strategic overview—a vineyard management approach focused on wine quality. The CLD includes variables related to viticultural practices, environmental conditions, and their effects on grape ripening and wine quality. This diagram illustrates how vineyard management attempts to close the quality gap in response to changing climate conditions, emphasizing the importance of adaptive strategies and the systemic nature of decision-making in viticulture. Marked time delays (represented by two parallel lines) highlight lags within the system and indicate that changes in certain variables do not lead to immediate results. These delays can be critical. As global temperature rises, the growing season lengthens, allowing grapes to ripen more fully—this is indicated by a positive causal link (+). A longer growing season contributes to better grape maturity and thus higher wine quality. Higher wine quality, in turn, reduces the quality gap, i.e., the difference between desired and actual wine quality (indicated by a negative causal link). A larger quality gap (i.e., lower actual quality compared to the desired one) motivates improvements in vineyard practices. This positive link illustrates that the need to close the quality gap drives better management.

At this point, a critical perceptual delay is encountered within the management subsystem. On the other hand, transitioning to appropriate viticultural practices often takes time. Improved practices can positively affect the length of the growing season—for example, by optimizing irrigation, pruning, and other measures that support grape development. This relationship is also marked by a delay, meaning that the benefits of improved practices become evident only after a certain period.

The core feedback loop of the system is balancing: as actual wine quality approaches the desired level, the quality gap narrows, reducing the pressure for further improvements. This loop ensures that the system stabilizes once the target wine quality is achieved.

5. Discussion

One of the key findings of this study is that regional affiliation exhibited the highest predictive power in explaining differences in average wine ratings. In machine learning models, particularly Random Forest, the inclusion of region significantly increased the explained variance (R²), confirming previous findings on the multidimensional nature of the terroir concept [19]. However, we recognize that the reduction of terroir to a single regional code inevitably omits important microclimatic, pedological, and cultural subtleties. Although region was included in the analysis as a categorical variable, it implicitly captures a complex combination of microclimatic, pedological, agronomic, and cultural factors that significantly affect the expression of varietal potential and wine quality.

The importance of region is not only related to natural conditions, but also to adaptive practices of winegrowers, which vary depending on experience, tradition, and local knowledge. This further justifies the inclusion of region as an explanatory variable in future ML models, despite its limited interpretability.

Among climatic variables, temperatures in April, September, and December showed the strongest positive correlations with wine quality, consistent with agronomically sensitive periods in vine development [24,36]. High April temperatures may be associated with favorable conditions for budburst and early growth, while September—a key month for ripening—directly influences the accumulation of sugars, acids, and aromatic compounds. The role of December temperatures, however, remains less clear. Although a positive correlation was observed, this finding may reflect indirect or delayed effects of milder winters on vine dormancy, disease pressure, or residual physiological activity. Further research is needed to clarify the mechanisms behind this association and assess its consistency across datasets.

Physiological Impacts of Heat on Vine Functioning and Grape Quality

Recent research has highlighted the importance of heat thresholds in understanding how extreme temperatures physiologically impair grapevine function and, ultimately, wine quality. When daily maximum temperatures exceed approximately 35 °C, photosynthetic efficiency declines due to stomatal closure and enzyme inhibition, particularly affecting sugar accumulation in berries and the synthesis of aromatic precursors. Sustained exposure to high temperatures can also lead to oxidative stress, sunburn, and degradation of organic acids, particularly during sensitive phenological stages such as véraison and ripening, when metabolic activity is most intense. These stress responses alter not only primary metabolism but also the phenolic and sensory profile of the resulting wine [9,25,35,36,37,40,42]. A clearer understanding of these threshold effects is critical for interpreting nonlinear responses to climatic extremes, which cannot be captured solely through average temperature correlations.

The highlighted importance of specific months supports the decision to use monthly temperature data rather than seasonal or annual averages, as this increases the sensitivity and accuracy of predictive models.

Contrary to expectations, precipitation did not emerge as a significant predictive factor. Despite some weak positive correlations (e.g., in November and December), most precipitation variables lacked statistical significance. Several reasons may explain this: (1) precipitation often acts indirectly—through soil moisture and disease pressure—which may not be directly reflected in wine ratings; (2) vineyard management practices (e.g., drainage, canopy management, use of protective treatments) may buffer the effects of precipitation; and (3) monthly aggregated data may overlook the intensity or timing of specific rainfall events that could have greater influence.

Despite the lack of statistical significance in our models, precipitation plays a well-documented role in viticulture, affecting vine water status, canopy development, disease incidence, and berry composition. The use of monthly aggregates may have obscured the influence of short-term events such as heavy rainfall or drought during sensitive growth stages. Moreover, the effects of precipitation are often mediated through complex interactions with temperature, soil, and vineyard practices. Future studies should consider integrating more specific indicators—such as drought indices or rainfall intensity—and exploring interaction effects to better capture precipitation’s role in wine quality formation. Another limitation lies in the use of aggregated regional data, which may obscure important microclimatic variations at the vineyard level. Such aggregation limits the detection of localized weather extremes or site-specific vineyard responses. The inclusion of high-resolution data from vineyard sensors, remote sensing platforms, or microclimate monitoring systems could improve model sensitivity and spatial accuracy. Future studies should aim to incorporate such data to better reflect the heterogeneity of viticultural environments.

Given the limitations of quantitative models, the results were complemented by the development of Causal Loop Diagrams (CLDs), which highlight key feedback mechanisms within the viticultural system. The model focused on biomass as an indirect indicator of quality illustrates that many viticultural responses—such as varietal changes, pruning strategies, or shading interventions—operate with delays, meaning that the effects of such actions often become evident only after several years. Particularly important is the notion of perceptual delay—the lag between systemic degradation and growers’ awareness—which can lead to untimely or inadequate responses. Perceptual delay refers to the time it takes for growers to recognize systemic changes, while temporal delay denotes the lag between actions taken and their observable effects in the vineyard. This has serious implications for climate change adaptation, underscoring the need for proactive rather than reactive vineyard management. CLD models thus offer a complementary explanation to ML outputs and point to structural reasons why regions with more flexible response capacities are often more successful in maintaining wine quality. While these developed CLD models provide valuable insight into the structure of feedback loops and systemic behavior in viticulture, they were constructed as qualitative conceptual tools and do not include quantified values for loop strength or explicit temporal delay durations. A full quantification of such feedback structures would require the development of a parameterized stock-and-flow simulation model, which was beyond the scope of this research. However, we acknowledge this as an important limitation and identify it as a future research direction. Dynamic simulation based on system dynamics modeling could further enhance understanding of time-sensitive responses in vineyard systems and help assess the long-term impact of adaptive strategies under climate variability.

The findings of this study confirm the value of a combined methodological approach that integrates statistical analysis, machine learning, and systems thinking, allowing for a deeper understanding of agroecological phenomena such as wine quality. Statistical tests (correlation, factor analysis, regression) provided a preliminary understanding of relationships between climate variables and wine ratings and allowed the formulation of basic hypotheses. Machine learning methods such as Random Forest and SVM enabled robust prediction based on complex, multidimensional data, where classical methods often fall short.

Despite this integrative methodological design, several limitations must be acknowledged that affect the interpretation of results and point to future research opportunities. First, the temporal aggregation of data—since all variables were analyzed on an annual basis—may obscure key phenological stages of the vine (e.g., budburst, flowering) that respond differently to climatic factors and may significantly contribute to quality variation. Future research should consider incorporating phenologically disaggregated data and linking specific developmental phases to predictive modeling. In this context, the use of Growing Degree Days (GDD), heat stress thresholds (e.g., days >35 °C), or cumulative heat indices could improve the biological relevance and temporal alignment of climatic predictors, providing a more accurate reflection of vine physiology under changing environmental conditions.

The second limitation concerns the temporal lag of effects, also identified in the CLD model—many grower responses (e.g., variety change, practice adjustments) occur only after a delay, while climate effects take place beforehand. This temporal mismatch complicates current predictive models and highlights the potential for integrating time-series approaches or system dynamics modeling with ML. Although the model was developed using data from Slovenia, the underlying methodological framework is adaptable to other wine-producing regions, provided that similar climatic and regional datasets are available.

The third issue relates to the limited dimensionality of the regional factor, which was encoded as a simple categorical variable. While it proved to be a key predictor, future studies could enrich the regional component with additional data—such as solar radiation indices, soil characteristics, varietal composition, vineyard age, or land use data. In addition, given the relatively small number of regional categories (n = 9) and the modest dataset size, the risk of overfitting in machine learning models must be acknowledged. Future work should address this by validating model robustness with expanded datasets and finer-grained regional or vineyard-level data. Moreover, although our current models relied on monthly aggregated temperature data, future research could benefit from the use of proxy phenological periods (e.g., April–May for budburst, August–September for ripening) as intermediate variables to approximate GDD and better reflect vine physiological processes. Such proxies may improve the temporal alignment between climatic predictors and sensitive developmental stages without requiring complete phenological datasets.

Finally, a promising future direction involves the use of multi-year sensory assessments conducted by professional panels, which could offer a more standardized and stable dependent variable. Nevertheless, the results demonstrate that even with a combined approach, significant predictive and interpretative insights can be achieved.

6. Conclusions

This study demonstrated that wine quality can be reliably predicted based on climatic and regional factors through the combined use of machine learning, statistical analysis, and systems modeling. Regional affiliation emerged as the strongest predictor, confirming the importance of microclimatic and pedological characteristics, as well as human practices encapsulated in the concept of terroir. Temperature—particularly during key months such as April and September—showed a significant influence on wine quality, whereas precipitation was not found to be statistically significant.

Machine learning methods, particularly Random Forest, enabled robust prediction and highlighted the relevance of regional differences. The use of Causal Loop Diagrams (CLDs) complemented the quantitative findings by explaining dynamic delays, systemic feedback loops, and the impact of adaptive strategies employed by winegrowers. This integrative approach provides a deeper understanding of viticultural systems under the influence of climate change and offers a foundation for designing more resilient wine production strategies. While developed on Slovenian data, the methodological framework presented here—combining machine learning, statistical analysis, and systems thinking—offers a flexible and transferable approach for other wine-producing regions. With appropriate local datasets, it can inform climate-resilient vineyard planning and contribute to regional or national policy strategies aimed at agricultural adaptation under climate change.