ART MARKET PREDICTIONS THROUGH DEEP LEARNING

Abstract [English]

The international art market lends complex dynamics that interact with aesthetic perception, the cycle of economic activities and the mood of the investor, making it a difficult task to forecast prices. The paper presents a deep learning architecture that combines visual, contextual, and temporal data to predict the valuations of artworks with further accuracy. The study proposes a data engineering pipeline that is multimodal that includes curated image collections and structured information, including artist background, sales history, medium, and dimensions. A convolutional neural network (CNN) is used to produce high-level structure of artistic style and quality, whereas transformer and Long Short-Memory (LSTM) structures discover the temporal dynamics of price tendencies in the past. These modalities are amalgamated into a single embedding that is a combination of both visual and economic cues a fusion layer. This model is hyperparameter tuned and transferred learned with the help of pretrained encoders to optimize prediction accuracy and prevent overfitting with regularization measures. Findings indicate better performance compared to traditional econometric and regression models and better correlation with the real market trends and overall generalization across genres and time series. In addition to predictive ability, the framework offers interpretable information suggesting the impact of artistic qualities on valuation trends, therefore, linking computational intelligence to the art economics. The suggested system provides possible solutions in market analytics, auction prediction, and digital art investment platforms, which add to the development of data-driven decision-making in the creative economy.