arXiv subjects

Riccardo Marcaccioli

Publications and source records attributed to Riccardo Marcaccioli.

6 recordsLinked to original sources

arXiv2026

A Random-Matrix Criterion for Initializing Gated Recurrent Neural Networks

Proper weight initialization prior to training has historically been one of the key factors that helped kick off the deep learning revolution. Initialization is even more crucial in "reservoir computing", where the weights of a readout layer are learned linearly while the reservoir weights are fixed and largely determine the richness, stability and memory of the resulting dynamics. In the infinite-width limit it has been shown that meaningful initializations are those sitting at an effective critical point of the randomly initialized model. The phase transition is controlled by the weight variance $g^2$ and separates an ordered phase from a chaotic one where information progressively degrades. Here we derive a simple criterion to estimate the critical $g_c$ for a broad class of recurrent architectures and we show that it closely tracks the gain at which a gated-RNN reservoir achieves peak performance on a chaotic forecasting task. Finally, we argue that our criterion can serve as a design principle for future initialization schemes.

cs.LG↗

arXiv2023

Deep Reinforcement Learning for Active High Frequency Trading

We introduce the first end-to-end Deep Reinforcement Learning (DRL) based framework for active high frequency trading in the stock market. We train DRL agents to trade one unit of Intel Corporation stock by employing the Proximal Policy Optimization algorithm. The training is performed on three contiguous months of high frequency Limit Order Book data, of which the last month constitutes the validation data. In order to maximise the signal to noise ratio in the training data, we compose the latter by only selecting training samples with largest price changes. The test is then carried out on the following month of data. Hyperparameters are tuned using the Sequential Model Based Optimization technique. We consider three different state characterizations, which differ in their LOB-based meta-features. Analysing the agents' performances on test data, we argue that the agents are able to create a dynamic representation of the underlying environment. They identify occasional regularities present in the data and exploit them to create long-term profitable trading strategies. Indeed, agents learn trading strategies able to produce stable positive returns in spite of the highly stochastic and non-stationary environment.

cs.LG↗

arXiv2021

Exogenous and Endogenous Price Jumps Belong to Different Dynamical Classes

Synchronising a database of stock specific news with 5 years worth of order book data on 300 stocks, we show that abnormal price movements following news releases (exogenous) exhibit markedly different dynamical features from those arising spontaneously (endogenous). On average, large volatility fluctuations induced by exogenous events occur abruptly and are followed by a decaying power-law relaxation, while endogenous price jumps are characterized by progressively accelerating growth of volatility, also followed by a power-law relaxation, but slower than for exogenous jumps. Remarkably, our results are reminiscent of what is observed in different contexts, namely Amazon book sales and YouTube views. Finally, we show that fitting power-laws to {\it individual} volatility profiles allows one to classify large events into endogenous and exogenous dynamical classes, without relying on the news feed.

q-fin.TR↗

arXiv2020

Correspondence between temporal correlations in time series, inverse problems, and the Spherical Model

In this paper we employ methods from Statistical Mechanics to model temporal correlations in time series. We put forward a methodology based on the Maximum Entropy principle to generate ensembles of time series constrained to preserve part of the temporal structure of an empirical time series of interest. We show that a constraint on the lag-one autocorrelation can be fully handled analytically, and corresponds to the well known Spherical Model of a ferromagnet. We then extend such a model to include constraints on more complex temporal correlations by means of perturbation theory, showing that this leads to substantial improvements in capturing the lag-one autocorrelation in the variance. We apply our approach on synthetic data, and illustrate how it can be used to formulate expectations on the future values of a data generating process.

cond-mat.stat-mech↗

arXiv2020

Maximum Entropy approach to multivariate time series randomization

Natural and social multivariate systems are commonly studied through sets of simultaneous and time-spaced measurements of the observables that drive their dynamics, i.e., through sets of time series. Typically, this is done via hypothesis testing: the statistical properties of the empirical time series are tested against those expected under a suitable null hypothesis. This is a very challenging task in complex interacting systems, where statistical stability is often poor due to lack of stationarity and ergodicity. Here, we describe an unsupervised, data-driven framework to perform hypothesis testing in such situations. This consists of a statistical mechanical approach - analogous to the configuration model for networked systems - for ensembles of time series designed to preserve, on average, some of the statistical properties observed on an empirical set of time series. We showcase its possible applications with a case study on financial portfolio selection.

q-fin.ST↗

arXiv2019

A parametric approach to information filtering in complex networks: The Pólya filter

The ever increasing availability of data demands for techniques to extract relevant information from complex interacting systems, which can often be represented as weighted networks. In recent years, a number of approaches have been proposed to extract network backbones by assessing the statistical significance of links against null hypotheses of random interaction. Yet, it is well known that the growth of most real-world networks is highly non-random, as past interactions between nodes typically increase the likelihood of further interaction. Here, we propose a network filtering methodology based on a family of null hypotheses that can be calibrated to assess which links are statistically significant with respect to a given network's own heterogeneity. We design such family of null hypotheses by adapting the Pólya urn, a simple one-parameter combinatorial model driven by a self-reinforcement mechanism, to a network setting. We provide a full analytical description of our filter, and show that it retains or discards links based on a non-trivial interplay between their own local importance and the importance of the nodes they belong to. We prove that the widely used disparity filter can be recovered as a large-strength approximation for a specific value of the Pólya filter's parameter, and illustrate our findings by applying the filter to two large network datasets.

physics.soc-ph↗