Abstract
Multi-variate time series forecasting is an important problem with a widerange of applications. Recent works model the relations between time-series asgraphs and have shown that propagating information over the relation graph canimprove time series forecasting. However, in many cases, relational informationis not available or is noisy and reliable. Moreover, most works ignore theunderlying uncertainty of time-series both for structure learning and derivingthe forecasts resulting in the structure not capturing the uncertaintyresulting in forecast distributions with poor uncertainty estimates. We tacklethis challenge and introduce STOIC, that leverages stochastic correlationsbetween time-series to learn underlying structure between time-series and toprovide well-calibrated and accurate forecasts. Over a wide-range of benchmarkdatasets STOIC provides around 16% more accurate and 14% better-calibratedforecasts. STOIC also shows better adaptation to noise in data during inference andcaptures important and useful relational information in various benchmarks.