Forecasting quantiles with reservoir computers and time-varying regimes of entropy
Abstract
I present a machine learning approach for predicting quantiles of time series. Reservoir computers are designed with the quantile loss function to forecast quantiles associated with different probability levels. I compare Reservoir computing approach with
a known statistical approach proposed by Engle and Manganelli (CAViaR, 2004). First, I do a comparison analysis for data simulated by a random logistic map. Then, I compare the mentioned methods on financial time series. I show that the machine learning approach
based on recurrent neural networks outperforms the statistical approach especially when a signal-noise ratio is relatively high.
In the second part of the presentation, I will give some results related to my thesis topic. Real-world systems may be in general non stationary, with an entropy value that is not constant in time. We propose a hypothesis testing procedure to test the null
hypothesis of constant Shannon entropy for time series, against the alternative of a significant variation of the entropy between two subsequent periods. We find the optimal length of the rolling window used for estimating the time-varying Shannon entropy
by optimizing a novel criterion.
Speaker
Andrey Shternshis is a PhD student at Scuola Normale Superiore of Pisa in the field of
Computational Methods and Mathematical Models for Sciences and Finance.
Best wishes,
Laura