arXiv subjects

Todd Kuffner

Publications and source records attributed to Todd Kuffner.

4 recordsLinked to original sources

arXiv2023

Online Bootstrap Inference with Nonconvex Stochastic Gradient Descent Estimator

In this paper, we investigate the theoretical properties of stochastic gradient descent (SGD) for statistical inference in the context of nonconvex optimization problems, which have been relatively unexplored compared to convex settings. Our study is the first to establish provable inferential procedures using the SGD estimator for general nonconvex objective functions, which may contain multiple local minima. We propose two novel online inferential procedures that combine SGD and the multiplier bootstrap technique. The first procedure employs a consistent covariance matrix estimator, and we establish its error convergence rate. The second procedure approximates the limit distribution using bootstrap SGD estimators, yielding asymptotically valid bootstrap confidence intervals. We validate the effectiveness of both approaches through numerical experiments. Furthermore, our analysis yields an intermediate result: the in-expectation error convergence rate for the original SGD estimator in nonconvex settings, which is comparable to existing results for convex problems. We believe this novel finding holds independent interest and enriches the literature on optimization and statistical inference.

stat.ML↗

arXiv2021

Conditional Randomization Rank Test

We propose a new method named the Conditional Randomization Rank Test (CRRT) for testing conditional independence of a response variable Y and a covariate variable X, conditional on the rest of the covariates Z. The new method generalizes the Conditional Randomization Test (CRT) of [CFJL18] by exploiting the knowledge of the conditional distribution of X|Z and is a conditional sampling based method that is easy to implement and interpret. In addition to guaranteeing exact type 1 error control, owing to a more flexible framework, the new method markedly outperforms the CRT in computational efficiency. We establish bounds on the probability of type 1 error in terms of total variation norm and also in terms of observed Kullback-Leibler divergence when the conditional distribution of X|Z is misspecified. We validate our theoretical results by extensive simulations and show that our new method has considerable advantages over other existing conditional sampling based methods when we take both power and computational efficiency into consideration.

stat.ME↗

arXiv2019

Bayesian Inference on Volatility in the Presence of Infinite Jump Activity and Microstructure Noise

Volatility estimation based on high-frequency data is key to accurately measure and control the risk of financial assets. A L\'{e}vy process with infinite jump activity and microstructure noise is considered one of the simplest, yet accurate enough, models for financial data at high-frequency. Utilizing this model, we propose a "purposely misspecified" posterior of the volatility obtained by ignoring the jump-component of the process. The misspecified posterior is further corrected by a simple estimate of the location shift and re-scaling of the log likelihood. Our main result establishes a Bernstein-von Mises (BvM) theorem, which states that the proposed adjusted posterior is asymptotically Gaussian, centered at a consistent estimator, and with variance equal to the inverse of the Fisher information. In the absence of microstructure noise, our approach can be extended to inferences of the integrated variance of a general It\^o semimartingale. Simulations are provided to demonstrate the accuracy of the resulting credible intervals, and the frequentist properties of the approximate Bayesian inference based on the adjusted posterior.

math.ST↗

arXiv2018

On model selection criteria for climate change impact studies

Climate change impact studies inform policymakers on the estimated damages of future climate change on economic, health and other outcomes. In most studies, an annual outcome variable is observed, e.g. agricultural yield, along with a higher-frequency regressor, e.g. daily temperature. Applied researchers then face a problem of selecting a model to characterize the nonlinear relationship between the outcome and the high-frequency regressor to make a policy recommendation based on the model-implied damage function. We show that existing model selection criteria are only suitable for the policy objective if one of the models under consideration nests the true model. If all models are seen as imperfect approximations to the true nonlinear relationship, the model that performs well in the normal climate conditions is not guaranteed to perform well at the projected climate that is different from the historical norm. We therefore propose a new criterion, the proximity-weighted mean-squared error (PWMSE), that directly targets precision of the damage function at the projected future climate. To make this criterion feasible, we assign higher weights to prior years that can serve as weather analogs to the projected future climate when evaluating competing models using the PWMSE. We show that our approach selects the best approximate regression model that has the smallest weighted error of predicted impacts for a projected future climate. A simulation study and an application revisiting the impact of climate change on agricultural production illustrate the empirical relevance of our theoretical analysis.

stat.ME↗