The aim of the project is to develop recursive methods for probabilistic time series forecasting using signatures of stochastic process trajectories.

Typical applications include electricity consumption, price, renewable production, and the objective is multivariate forecasting (forecast of different time series at the same time). Multivariate forecasting allows also different time horizons as short term (day), medium term (week, month) and long term (year) forecasts. The forecast needs to adapt to a constantly changing environment. Recursive approaches are favoured in this project as they provide forecasts that are automatically updated when new data arrive.

We propose to combine the recent concept of randomized signatures with one of the classical recursive forecasting methods: state-space models and the associated recursive algorithms.

PhD student: Nina DROBAC

PhD supervisors: Pr Olivier WINTENBERGER (Director), Dr Yannig GOUDE (Co-advisor)

Research laboratory: LPSM – UMR 8001 – CNRS – Sorbonne Universite – Statistiques, données, algorithmes