Invited Talks | Spring 2016 AMS/CAFIN Workshop and Distinguished Lecture

AMS/CAFIN Distinguished Lecture

Bayesian Predictive Synthesis, Mike West, Arts and Sciences Professor of Statistics and Decision Sciences, Duke University. Abstract: I discuss Bayesian reasoning and resulting methodology for model and forecast comparison, calibration, and combination. Beginning with a foundational perspective that in part responds to recent "pragmatic" modelling developments in macroeconomic forecasting, I revisit a theoretical framework for model and forecast uncertainty assessment that links back to 1980s/90s literatures cutting across statistics, economics, management science and other fields. Fast-forwarding and updating to 2016 sees the emergence of Bayesian predictive synthesis (BPS)-- a coherent theoretical basis for combining multiple forecast densities, whether from models, individuals, or other sources, and extending existing forecast pooling and Bayesian model mixing methods. Time series extensions are implicit dynamic latent factor models, allowing adaptation to time-varying biases, mis-calibration, and dependencies among models or forecasters. Bayesian simulation-based computation enables implementation. A macroeconomic time series study highlights insights into dynamic relationships among synthesized forecast densities, as well as the potential for improved forecast accuracy at multiple horizons.

Invited Talks

Relative Tick Size and Price Discovery, Eric Aldrich, Department of Economics. UCSC. Abstract: We develop a theoretical model to highlight a previously unexplored mechanism of price discovery: relative minimum price increments for equivalent assets trading on distinct financial exchanges. Although conventional wisdom dictates that futures market assets lead equities equivalents in terms of price formation, our model predicts that the opposite should be true when particular relative price conditions hold for the bids and offers of each asset. We develop a new empirical measure of price discovery which is suited to asynchronous, high-frequency transaction and quotation data, and apply it to the highly liquid E-mini/SPY pair in order to test the predictions of the model. Empirical evidence strongly supports the model and further demonstrates that relative minimum contract size plays an additional role in the formation of prices.
Dynamic Linear Models in hostile environments, Ricardo Lemos, The Climate Corporation. Abstract: In fisheries stock assessment, researchers routinely need to estimate probabilistic trajectories of annual stock abundance, together with parameters that define the stock's resilience to exploitation, the carrying capacity of the environment, and the efficiency of fishing gear. Such tasks are carried out in the framework of surplus production models (SPMs). In data-richer fisheries, information regarding the relative abundance of fish with different ages is also available, making it possible to construct age-structured models (ASMs). Because of the canonical assumptions regarding the surplus production function (in SPMs) and the stock-recruitment function (in ASMs), these models are non-linear and can become difficult to fit to data. In this work I provide a brief overview of the pitfalls of canonical SPMs and AGMs, and present alternatives, based on the Dynamic Linear Modeling framework. I compare four SPMs applied to the Namibian hake fishery, and discuss how an ASM can still provide insights even if little information is available about the steepness parameter.
A Biased Sample of Data Science At Facebook. Dan Merl, Data Scientist, Facebook. Abstract: At Facebook there are approximately 1200 data scientists, researchers and analysts whose mission is to develop new ways to harness Facebook data to build better products and inform responsible decisions. In this talk, I will discuss a biased sample of some ongoing, mostly-statistical research being performed by a similarly biased sample of 20 of these scientists. This set of 20, in addition to providing a fine example of the breadth and depth of data science activities across Facebook, also uniquely satisfy an optimization problem characterized by the extent to which I am minimally unqualified to present their work on their behalf by virtue of being their manager.