arXiv subjects

Taylor R. Brown

Publications and source records attributed to Taylor R. Brown.

6 recordsLinked to original sources

arXiv2022

A Case-Study of Sample-Based Bayesian Forecasting Algorithms

For a Bayesian, real-time forecasting with the posterior predictive distribution can be challenging for a variety of time series models. First, estimating the parameters of a time series model can be difficult with sample-based approaches when the model's likelihood is intractable and/or when the data set being used is large. Second, once samples from a parameter posterior are obtained on a fixed window of data, it is not clear how they will be used to generate forecasts, nor is it clear how, and in what sense, they will be ``updated" as interest shifts to newer posteriors as new data arrive. This paper provides a comparison of the sample-based forecasting algorithms that are available for Bayesians interested in real-time forecasting with nonlinear/non-Gaussian state space models. An applied analysis of financial returns is provided using a well-established stochastic volatility model. The principal aim of this paper is to provide guidance on how to select one of these algorithms, and to describe a variety of benefits and pitfalls associated with each approach.

stat.AP↗

arXiv2022

The Most Difference in Means: A Statistic for the Strength of Null and Near-Zero Results

Statistical insignificance does not suggest the absence of effect, yet scientists must often use null results as evidence of negligible (near-zero) effect size to falsify scientific hypotheses. Doing so must assess a result's null strength, defined as the evidence for a negligible effect size. Such an assessment would differentiate strong null results that suggest a negligible effect size from weak null results that suggest a broad range of potential effect sizes. We propose the most difference in means ($δ_M$) as a two-sample statistic that can both quantify null strength and perform a hypothesis test for negligible effect size. To facilitate consensus when interpreting results, our statistic allows scientists to conclude that a result has negligible effect size using different thresholds with no recalculation required. To assist with selecting a threshold, $δ_M$ can also compare null strength between related results. Both $δ_M$ and the relative form of $δ_M$ outperform other candidate statistics in comparing null strength. We compile broadly related results and use the relative $δ_M$ to compare null strength across different treatments, measurement methods, and experiment models. Reporting the relative $δ_M$ may provide a technical solution to the file drawer problem by encouraging the publication of null and near-zero results.

stat.ME↗

arXiv2021

Approximating Posterior Predictive Distributions by Averaging Output From Many Particle Filters

This paper introduces the {\it particle swarm filter} (not to be confused with particle swarm optimization): a recursive and embarrassingly parallel algorithm that targets an approximation to the sequence of posterior predictive distributions by averaging expectation approximations from many particle filters. A law of large numbers and a central limit theorem are provided, as well as an numerical study of simulated data from a stochastic volatility model.

stat.ME↗

arXiv2020

PF: A C++ Library for Fast Particle Filtering

Particle filters are a class of algorithms that are used for "tracking" or "filtering" in real-time for a wide array of time series models. Despite their comprehensive applicability, particle filters are not always the tool of choice for many practitioners, due to how difficult they are to implement. This short article presents PF, a C++ header-only template library that provides fast implementations of many different particle filters. A tutorial along with an extensive fully-worked example is provided.

stat.CO↗

arXiv2019

A Pseudo-Marginal Metropolis-Hastings Algorithm for Estimating Generalized Linear Models in the Presence of Missing Data

The missing data issue often complicates the task of estimating generalized linear models (GLMs). We describe why the pseudo-marginal Metropolis-Hastings algorithm, used in this setting, is an effective strategy for parameter estimation. This approach requires fewer assumptions, it provides joint inferences on the parameters in the likelihood, the covariate model, and the parameters of the missingness-mechanism, and there is no logical inconsistency of assuming that there are multiple posterior distributions. Moreover, this approach is asymptotically exact, just like most other Markov chain Monte Carlo techniques. We discuss computing strategies, conduct a simulation study demonstrating how standard errors change as a function of percent missingness, and we use our approach on a "real-world" data set to describe how a collection of variables influences the car crash outcomes.

stat.ME↗

arXiv2019

A Factor Stochastic Volatility Model with Markov-Switching Panic Regimes

The use of factor stochastic volatility models requires choosing the number of latent factors used to describe the dynamics of the financial returns process; however, empirical evidence suggests that the number and makeup of pertinent factors is time-varying and economically situational. We present a novel factor stochastic volatility model that allows for random subsets of assets to have their members experience non-market-wide panics. These participating assets will experience an increase in their variances and within-group covariances. We also give an estimation algorithm for this model that takes advantage of recent results on Particle Markov chain Monte Carlo techniques.

stat.AP↗