Probabilistic Forecast of Time Series with Transformers and Normalizing Flows

Today, machine learning has many applications and is used in different fields and industries. One of its applications is demand forecasting. Future demands play a vital role in any industry. It can help the organization from planning to making the stock levels available to the customer when required. This thesis aims to understand normalizing flows and do multivariate probabilistic forecasting using normalizing flows conditioned on autoregressive models like GRUs and Transformers. The normalizing flow scan be effective in modelling complex distributions. The built models are tested on different available time-series data sets and H&Ms data. TheH&Ms dataset consists of the weekly sales of various articles. These articles are supplied to the stores from different warehouses. The time series in fashion retail is often ephemeral; hence shorted time steps are there. Similar is the case with the H&Ms data. The data acquired is first preprocessed and then passed to the model. The results are obtained and compared using different evaluation metrics