PDM: Time-series forecasting models with 6M–119M parameters (www.flashmind.com)

Flashmind, based in the UAE, has unveiled PDM, a new family of AI models optimized for time-series forecasting, which includes three variants: Nano, Pico, and Femto, ranging from 6 million to 119 million parameters. These models are capable of predicting future values based solely on past data without requiring any additional model training, a process known as zero-shot forecasting. At launch, PDM-Nano topped the leaderboard for zero-shot foundation models by normalized MASE, demonstrating its superior predictive capabilities for various applications, including demand forecasting and energy utilization. What sets PDM apart is its ability to incorporate covariates, or additional contextual information, enhancing its predictive accuracy. For instance, including data on promotions or operational changes can significantly reduce forecast errors—showing improvements in accuracy metrics like mean absolute error from 29.50 to 18.00 drinks in a café stock forecasting scenario. The tailored architecture allows all three model variants to perform multivariate forecasting, addressing multiple related time series simultaneously while maintaining lower errors compared to larger models in the market. This innovation positions PDM as a leading solution for businesses and researchers seeking intuitive and effective forecasting tools in the rapidly evolving AI/ML landscape.