2.1 Sample

Initial data comprised daily box office records (January 1, 2022 - November 1, 2024) from the Korean Film Council’s KOBIS, supplemented by public holiday data. Preprocessing involved several exclusions, including films released before April 17, 2022 (the end of social distancing), box office data beyond 100 days post-release, films released after September 13, 2024 (to ensure sufficient observation), pre-release audience data, re-releases, and films with fewer than 100,000 cumulative viewers. This yielded 219 films. As online reviews were sourced from CGV, a major cinema chain, four films not released on this platform were further excluded, resulting in a candidate pool of 215 films (Choe & Kang, 2017; Kwon & Kim, 2021). Final sample determination based on Bass model estimation is detailed in Section 3.1.

2.2 Variables

2.2.2 Independent Variables

Emotion ratios from CGV online reviews served as independent variables. The initial aim was to collect the first 1,000 reviews per movie post-release. Emotion analysis utilized a KcELECTRA model (Lee, 2021) fine-tuned with the KOTE dataset (Jeon et al., 2022), which identifies 43 Korean emotion categories.

The process of constructing the emotion dataset for each movie involved several steps, visualized in Figure 1

. First, reviews were crawled from the CGV website. Second, using Google Colab, the fine-tuned KcELECTRA model was applied to each review to predict emotion labels and their corresponding probability scores. The model can assign multiple emotion labels to a single review, and these scores represent independent probabilities for each label, thus not summing to one. Initially, all predicted emotions with a score greater than .4 were recorded for each review. Third, for each review, these predicted emotions were sorted by score in descending order, and the top five emotions were provisionally retained. Fourth, any instances of the ‘No Emotion’ label were removed from consideration. Fifth, a stricter filtering criterion was applied: only emotions with a prediction score of .8 or higher were kept for the final analysis. This threshold was determined based on a qualitative assessment of prediction accuracy against review content to ensure a high confidence level in the identified emotions. Sixth, reviews that had no emotions meeting this .8 score threshold after the preceding filtering steps were excluded from the emotion dataset.

This rigorous filtering process resulted in a final baseline emotion dataset comprising 182,987 reviews across the analyzed films. By aggregating such a large volume of reviews per film, we leverage the Law of Large Numbers to mitigate the influence of individual-level noise (e.g., a viewer’s personal mood or viewing context). While individual reviews may be influenced by external factors, the aggregated emotion ratios effectively capture the common sentiment driven by the film’s intrinsic attributes, thus ensuring ecological validity. The number of reviews per movie varied, with a minimum of 411 and a maximum of 994. An example of how emotions are extracted from a review text and their associated scores can be seen in Figure 2

. After applying these filters, 39 distinct emotion categories (excluding ‘No Emotion’ and four other emotions that either never appeared or never met the .8 score threshold) were present in the dataset.

For each of the 215 films and for each of the 39 emotions, an emotion ratio ($R_{e,m}$) was calculated as

$$R_{e,m} = N_{e,m} / K_{m}$$

(2)

where $N_{e,m}$ is the number of reviews for movie $m$ that express emotion $e$ (with a score ≥ .8), and $K_{m}$ is the total number of valid reviews analyzed for movie $m$. These emotion ratios served as the primary independent variables. Given the large number of emotion variables, factor analysis was employed to reduce dimensionality, as detailed in Section 4.2. The resulting factor scores were then used in regression models to examine their relationship with the Bass diffusion parameters.

3.1 Bass Diffusion Model Estimation

Bass parameters (p, q, m) were estimated using NLS with constraints (p, q ≥ .00001, and m > observed maximum audience). Table 1

3.2 Factor Analysis of Emotional Variables

Factor analysis (Principal Component Analysis with Varimax rotation; loadings < .3 suppressed) on the 39 emotion ratios yielded nine factors. Table 2

Table 2 Factor Analysis Results for 39 Emotions (Rotated Component Matrix)

(labeled Factor 1 to Factor 9) for use in regression. Based on the factor loadings, nine emotional dimensions were identified and labeled accordingly. Factor 1 was labeled Irritation, reflecting emotions such as dissatisfaction and boredom, while Factor 2, characterized by distrust, despair, and anger, was labeled Anger. Factor 3 was labeled Distress as it encompassed emotions of disgust, sorrow, and shock, and Factor 4 was labeled Anxiety, capturing anxiety, fear, and surprise. Factor 5 was labeled Tension, indicating the contrast between expectancy and comfort. Factor 6, defined by exhaustion and realization, was labeled Fatigue. Factor 7 was labeled Gratitude, representing relief, respect, and gratitude, whereas Factor 8 was labeled Care and Factor 9 was labeled Pride. Because these emotional factors were extracted from the reactions of actual moviegoers, the factor loading values can be interpreted as indicating the extent to which each film possesses attributes capable of eliciting specific emotional responses.

The labeling of Factor 5 warrants particular theoretical justification, as it exhibits a bipolar dimension. Lehne and Koelsch (2015) define tension as “an emotional state characterized by a desire for change that is felt to eventually occur.” In music, this relates to the expectation-driven experience of listeners. Building on this, Factor 5’s bipolar structure—anticipation (+) versus contentment (−)—closely corresponds to foundational emotion models: Thayer’s (1989) energetic arousal dimension, Russell’s (1980) circumplex arousal axis, and Scherer’s (2001) arousal-based categorization. In these frameworks, high-arousal emotions (e.g., anticipation) contrast with low-arousal emotions (e.g., serenity). From a psychological perspective, anticipatory tension is neither pure surprise nor joy—it is an expectancy-driven state associated with heightened physiological arousal (heart rate, cortisol levels). This aligns with Scherer’s felt action tendency and Russell’s arousal-based differentiation of affective states. Therefore, for movies—where temporal unfolding and suspense are narrative cornerstones—Tension is both a conceptually valid and psychologically grounded label for this bipolar factor.

3.3 Regression Analysis Results

OLS regressions examined the impact of the nine emotion factor scores on p, q, and m. The variance inflation factors (VIFs) of all independent variables included in the regression models were below 6, indicating that multicollinearity was not a concern.

In the regression analysis using the innovation coefficient (p) of the Bass model as the dependent variable (Table 3

Among the control variables, the number of poster images, number of screens, and the Seollal dummy variable were found to have significantly negative effects.

These results indicate that films with stronger attributes likely to elicit irritation-related emotions tend to attract a higher proportion of consumers who choose to watch the movie independently. In contrast, films with stronger attributes associated with anger, tension, or fatigue are associated with a lower proportion of self-initiated moviegoers. Moreover, contrary to expectations, an increase in the number of posters and screens was found to decrease the innovation coefficient, suggesting that, unlike traditional explanations, greater marketing communication efforts and wider distribution channel accessibility may actually reduce the proportion of consumers who decide to watch a film on their own initiative.

In the regression analysis using the imitation coefficient (q) of the Bass model as the dependent variable,

These results suggest that films with stronger attributes likely to evoke Anger or Tension tend to attract a higher proportion of viewers who decide to watch the film under the influence of other consumers. Furthermore, consistent with expectations, films released during Seollal and Chuseok holidays exhibited higher imitation coefficients, indicating that during peak seasons, a larger number of low-involvement consumers—who are more susceptible to word-of-mouth—tend to watch movies.

In the regression analysis using the market potential (m) of the Bass model as the dependent variable,

These results indicate that films with stronger attributes likely to evoke irritation-related emotions tend to have lower market potential, whereas those that elicit tension-related emotions tend to exhibit higher market potential. In addition, consistent with previous findings on film-related determinants, the positive correlation between the number of screens and box office performance was reaffirmed.

3.4 Robustness Check

To assess the temporal stability of emotion expressions in online reviews, an additional analysis was conducted. For this check, review data were collected at three specific time points: the day of release (Day 0), seven days post-release (Day 7), and fourteen days post-release (Day 14). A cap of 1,000 reviews per day per movie was set, and the analysis included 42 movies for which at least 300 reviews were available at each of these three time points. Chi-square tests were performed to compare the distribution of the 39 emotion ratios across these time points (Day 0 vs. Day 7, Day 7 vs. Day 14, and Day 0 vs. Day 14). As shown in Table 6

4.1 Theoretical Implications

The key contribution of this study lies in establishing an empirical link between the emotional attributes of films and their diffusion trajectories. Factor 1 (Irritation) and Factor 5 (Tension) showed opposite effects on the innovation coefficient and market potential—positive and negative for one, negative and positive for the other. This suggests that irritation-inducing attributes may attract innovators and boost early audience inflow but ultimately hinder box office performance, whereas tension-inducing attributes may have limited impact initially yet enhance long-term success. These contrasting patterns imply that innovators and the general audience evaluate emotional attributes using markedly different, even opposing, criteria.

Similarly, Factor 1 (Anger) and Factor 5 (Tension) negatively affected the innovation coefficient but positively influenced the imitation coefficient, indicating that films evoking clear anger targets or strong tension tend to be avoided by innovators but embraced by imitators. Collectively, these findings highlight that imitators play a decisive role in determining a film’s ultimate commercial performance in the Korean market. Notably, Factor 5 (Tension) significantly affected all three Bass model parameters. Positively associated with expectancy and negatively with comfort, attraction, and happiness, this factor represents a cinematic attribute that evokes suspense through unpredictability. In essence, films that elicit tension and suspense are more likely to become the subject of word-of-mouth communication and achieve commercial success in the Korean film market.

From a production standpoint, filmmakers should consider how a film’s emotional tone influences its diffusion trajectory. Since tension-related attributes promote long-term success while irritation-related ones mainly attract early audiences but deter broader viewership, creating balanced tension or controlled suspense may sustain audience engagement beyond the initial release. From a marketing perspective, the findings highlight that imitators play a decisive role in determining box office outcomes. Accordingly, post-release strategies should focus on word-of-mouth amplification and emphasize the emotional depth and suspenseful appeal of films, rather than mere novelty. Leveraging peak seasons and socially engaging content can further enhance diffusion among later audiences.

4.2 Limitations and Future Research

Several constraints warrant caution in interpreting the findings and point to opportunities for future research. First, the use of static summary statistics derived from early reviews—rather than dynamic, time-varying emotional signals—represents a methodological choice with important trade-offs. While this approach effectively captures the emotional tone perceived by the majority of viewers (especially due to the Law of Large Numbers effect with our large sample of 182,987 reviews), it does not account for temporal evolution in emotional expressions or shifts in audience composition over a film’s lifecycle. Future research could employ state-space models or time-varying coefficient frameworks (e.g., Kalman filters) to explicitly model how emotional signatures evolve over time and how such dynamics influence diffusion parameters. Second, while our large sample size (183,000 reviews) ensures ecological validity by leveraging the Law of Large Numbers to capture genuine common sentiment rather than idiosyncratic noise, sample representativeness and data collection strategies also present limitations. The dataset consists of 209 Korean films released within a particular period, and reviews were exclusively gathered from a single cinema website (CGV). Emotional expression patterns may differ on other platforms (e.g., Naver or international streaming services). Third, although we controlled for 19 genre dummies to isolate the marginal effect of a film’s specific emotional execution, perfect disentanglement of structural factors remains challenging. For instance, certain genres (e.g., Thriller) inherently predispose films toward higher Tension scores, making it difficult to fully separate genre-driven expectations from execution-driven emotional impact. Future studies could employ hierarchical or random-effects models to more rigorously decompose within-genre and between-genre variance. Fourth, the emotional classification relies on a KcELECTRA-based model fine-tuned with the KOTE dataset at a high prediction-score threshold (.8). Although this approach strengthens accuracy, similarities or ambiguities among certain emotions—such as boredom versus disappointment—remain challenging. Cultural and linguistic nuances specific to Korean online discourse further limit generalizability. Subsequent research could expand to multi-lingual contexts, refine labeling criteria, or consolidate highly similar emotions to reduce classification ambiguity.