Pattern Recognition and Prediction in Time Series Data Through Retrieval-Augmented Techniques

Time series forecasting is an essential technique used in various domains, with finance being one of the primary users. There are several techniques for time series forecasting, with machine learning being one of the most common practices. Machine learning models rely on numerical input with time encoded in the input for prediction. However, visual pattern-based time series forecasting is also a popular method practiced by professionals in various fields. Traditional numerical input-based models succeed in recognizing patterns but often fail when the patterns are skewed or stretched compared to those in the training data. The aim of the work is to address this issue using a visual approach. In this work, OpenAI’s CLIP model has been used to perform a visual similarity search among visual time series patterns to achieve forecasting using the top k most similar patterns. The performance of the model shows it can be used effectively for time series analysis.