Searcharxiv⌕ Search

arXiv · 2611.22257

The Online LSTM Prediction Models and The Term Structure of Predictions

Abstract

We examine and illustrate the properties of the Long Short Term Memory (LSTM) prediction models applied online to stationary inputs. We discuss the effects of: a) the network architecture that defines the admissible prediction functions, b) the objective function used to measure the prediction accuracy, and c) the Stochastic Gradient Descent (SGD) algorithm. We also study the prediction optimality, the term structure of predictions, the filtered memory dynamics. We also introduce a set of diagnostic tools for memory-based prediction models that focus on the estimated dynamics of the hidden states.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Christian Gourieroux, Joann Jasiak, Cheng Zhong. 2026-11-17. The Online LSTM Prediction Models and The Term Structure of Predictions. https://searcharxiv.com/abs/2611.22257

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Difference-in-differences for mediation analysis using double machine learning

We propose a difference-in-differences (DiD) framework with mediation for possibly multivalued discrete or continuous treatments and mediators, aimed at identifying the direct effect of the treatment on the outcome (net of effects operating through the mediator), the indirect effect via the mediator, and the joint effects of treatment and mediator, consistent with the framework of dynamic treatment effects. Identification relies on a conditional parallel trends assumption imposed on the mean potential outcome across treatment and mediator states, or (depending on the causal parameter) additionally on the mean potential outcomes and potential mediator distributions across treatment states. We propose ATET estimators for repeated cross sections and panel data within the double/debiased machine learning framework, which allows for data-driven control of covariates, and we establish their asymptotic normality under standard regularity conditions. We investigate the finite-sample performance of the proposed methods in a simulation study and illustrate our approach in an empirical application to the US National Longitudinal Survey of Youth, estimating the direct effect of health care coverage on general health as well as the indirect effect operating through routine checkups.

econ.EM↗

Estimation and Model Selection in Factor Models with Long Memory

Panels in economics, finance and climate science are often driven by common components that differ in memory, such as common trends, cycles, or joint shocks. We study approximate factor models whose common factors are fractionally integrated with heterogeneous and unknown memory and develop the asymptotic theory of principal components for the panel in levels. Our theory builds on the eigenvalues of the sample covariance matrix being ordered by memory asymptotically: a nonstationary factor group with memory $d>1/2$ generates eigenvalues of order $NT^{2d-1}$, whereas stationary factors generate eigenvalues of order $N$. Principal components estimate the factors consistently up to rotation, while adjacent factor groups are separated at a rate given by their memory gap. The information criteria of Bai and Ng remain consistent for the total number of factors, with their penalties acting as thresholds on eigenvalues that are asymptotically ordered by factor memory. When the nonstationary factors lose their variance advantage, estimation in levels remains consistent, whereas after differencing the factors can become undetectable. A Monte Carlo study shows that factor-number selection in levels tolerates more idiosyncratic noise than selection after first differencing. For monthly temperatures at $308$ European weather stations, our methods select two factors. The first is a persistent common warming factor. The estimated common component implies that 2021-2025 was $2.15^\circ\mathrm{C}$ warmer than 1960-1990, with the strongest warming in Eastern Europe. After first differencing, the criterion selects one factor, and loses regional heterogeneity as compared to the model in levels.

econ.EM↗

Nested logit: model structure and normalisation

Nested logit remains a widely used model, especially in applied work. However, there remains substantial confusion surrounding the existence of different versions of the model, different normalisations for the same version, and different software implementations. This paper argues that the labels in common use, RU1 and RU2, and the name non-normalised nested logit (NNNL) itself, have contributed to the confusion. The model is written in two forms: one in which utilities are divided by the structural parameter of the nest that contains them, which is consistent with utility maximisation by construction, and one without that division, known as NNNL. The difference between these forms is not one of normalisation. Normalising at the top or at the bottom gives the same model, and NNNL is itself normalised. What distinguishes the forms is whether the division is performed. In short, normalisation changes the units, whereas division changes the model. The undivided form coincides with the utility-consistent one only under specific conditions, and otherwise is consistent with utility maximisation only under utilities other than those specified, in which generic coefficients become nest-specific. The paper highlights that RU1 has been used for three different specifications, compares how widely used software packages normalise and report their parameters, and illustrates the consequences with calculated examples. The recommendation is that models be described by what they do, rather than by these labels.

econ.EM↗