GNNs have seen great success in natural language processing and signal processing. We have investigated whether such results can be generalized to a broader range of time series by proposing a novel method of graph construction which do not rely on preexisting spatial or relational data and Time-GNN was developed. This method dynamically learns relationships between timesteps based on the behavior of the input time series. TimeGNN is able to achieve comparable forecasting performance while being far more scalable than existing GNN based approaches. However, improvements to forecasting quality and more benchmark comparisons still need to be done.