INNOVATIVE APPROACH TO TIME SERIES FORECASTING: FROM TRADITIONAL METHODS TO THE CUTTING-EDGE MODEL TIMESFM

Abstract

Forecasting time series is crucial for companies as it enables them to make informed decisions using historical data and future predictions. Various approaches exist for time series forecasting, each tailored to different data characteristics and forecasting tasks. Currently, widely known models include ARIMA, LSTM networks, Prophet, and XGBoost. However, these models often require significant preprocessing time. To address this issue and leverage recent advancements in generative AI, Google introduced TimesFM, a decoder-only model specifically designed for time series forecasting. TimesFM utilizes transformer layers and a multi-layer perceptron block to transform time series fragments into tokens, enabling efficient forecasting with minimal generation steps. Synthetic and real-world data are combined for pretraining to capture fundamental temporal patterns and enhance model generalization. Evaluation demonstrates TimesFM's competitive performance across various benchmark time series datasets compared to traditional statistical methods and DL models.