Nuwa: Evaluation-Grounded Agentic Construction of Time Series Forecasting Systems

Time series forecasting has recently moved beyond static, single-pass prediction toward agentic methods that reason, use tools, execute workflows, and accumulate memory. However, most existing forecasting agents still optimize predictions or pipelines under largely implicit configurations, leaving the construction of the forecasting system itself underexplored. This limits their ability to attribute performance changes to specific decisions, such as which module, exogenous variable, or hyperparameter caused an improvement. We propose Nuwa, an evaluation-grounded agentic framework for constructing time series forecasting systems. Nuwa defines a structured design space over modules, exogenous variables, and hyperparameters, and performs Evaluation-Budgeted Construction that uses data meta-information, performance diagnostics, and memory to propose candidate sets, select budgeted subsets for controlled evaluation, attribute local effects, and decide whether to stop, accept, retry, switch subspace, or continue construction. The resulting design-level evidence is stored in a Design Memory, enabling reusable knowledge about which configurations work under which data conditions. Across approximately 15,000 experimental runs, Nuwa achieves consistent performance gains across module, exogenous-variable, and hyperparameter construction, with an overall average MSE reduction of 8.0\% over the corresponding reference settings. By turning forecast-level feedback into attributable and transferable design evidence, Nuwa makes forecasting system construction an explicit objective for agentic time series forecasting. The anonymous repository is available at https://anonymous.4open.science/r/NUWA/.