Reproduce “DAM: Towards A Foundation Model for Time Series Forecasting” on OpenResearch

About this paper

Luke Darlow, Qiwen Deng, Ahmed Hassan, Martin Asenov, Rajkarn Singh, Artjom Joosen, Adam Barker, Amos Storkey

The Deep Data-Dependant Approximate Analytical Model (DAM) is presented as a neural model for universal time series forecasting. This model uniquely handles irregular and variable-length input data and produces continuous function forecasts, demonstrating competitive or superior performance across 25 diverse datasets in long-term and zero-shot forecasting tasks.

Key insights

• A single, broadly trained model can achieve competitive or superior performance across a wide range of time series tasks, often outperforming specialized models trained per dataset.
• The combination of an adaptive history sampling strategy and continuous function output via basis decomposition is effective for handling data irregularities and achieving flexible forecasting horizons.
• The explicit basis function composition provides inherent interpretability, revealing the temporal frequencies contributing to forecasts and increasing model transparency.

Summary by alphaXiv.