Dozhdikov, A.V. (2026). Forecasting Russian film box office performance with large language models: Enhancing industry resilience and mitigating risks. Nauka Televideniya—The Art and Science of Television, 22 (2), 183–217. https: ...

Текст статьи на сайте журнала URL: https://tv-science.online/journals/22-2-prognozirovanie-rezultatov-rossijskih-filmov-v-prokate-s-ispolzovaniem-bolshih-yazykovyh-modelej-v-tselyah-povysheniya-ustojchivosti-natsionalnoj-otrasli-i-snizheniya-riskov/ (дата обращения 24.08.2026)

Abstract

The article presents the results of applying the Low-Rank Adaptation (LoRA) technology to retrain large language models for predicting the performance of Russian films based on extended text descriptions. The study includes: (1) a theoretical review of modern approaches to film box office forecasting using natural language processing; (2) proof of the hypothesis that forecasting is possible using LLMs; (3) the practical implementation of a predictive model based on retraining the ru-enRoSBERTa LoRA model; and (4) a description of a preliminary experiment on retraining a generative model based on Qwen2.5—7B instruct for creating film project descriptions. The experimental data were obtained from 1,683 Russian national films released between 2004 and 2024. The results demonstrate the possibility of achieving a forecasting accuracy of 0.93 (ROC-AUC = 0.92) using LoRA adaptation and LLM prompting—comparable to the metrics of traditional machine learning methods that rely on historical data. The study confirms the effectiveness of LLMs for solving forecasting problems in the film industry and substantiates the prospects for creating a “cyberproducer,” a decision support system for the film industry. These findings can be used to improve the financial, economic, and sociopolitical sustainability of the Russian film industry; optimize investment strategies; and reduce the financial and regulatory risks for filmmakers. Unlike traditional ML approaches based on regression, historical distribution data, and creative team composition—which tend to disadvantage emerging authors, having no accomplished projects—LLMs enable direct work with ideas and drafts during the script pitching stage, without discriminating against novice creators. Future work will extend the analysis to more detailed synopses, treatments, and full film scripts, and will aim to create a hybrid forecasting model by combining boosted ensemble ML models with LLMs.