Sequential learning versus no learning in Bayesian regression models

We consider two regression models: linear and logistic. The dependent variable is observed periodically and in each period a Bayesian formulation is used to generate updated forecasts of the dependent variable as new data is observed. One would expect that including new data in the Bayesian updates results in improved forecasts over not including the new data. Our findings indicate that this is not always true. We show there exists a subset of the independent variable space that we call the “region of no learning.” If the independent variable values for a given period in the future are in this region, then the forecast does not change with any new data. Moreover, if the independent variable values are in a neighborhood of the region of no learning, then there may be little benefit to wait for the new data and update the forecast. We propose a statistical approach to characterize this neighborhood which we call the “region of little learning.” Our results provide insights into the trade‐offs that exist in situations when the decision maker has an incentive to make an early decision based on an early forecast versus waiting to make a later decision based on an updated forecast. © 2014 Wiley Periodicals, Inc. Naval Research Logistics 61: 532–548, 2014     

A logistic regression/Markov chain model for NCAA basketball

Each year, more than $3 billion is wagered on the NCAA Division 1 men's basketball tournament. Most of that money is wagered in pools where the object is to correctly predict winners of each game, with emphasis on the last four teams remaining (the Final Four). In this paper, we present a combined logistic regression/Markov chain model for predicting the outcome of NCAA tournament games given only basic input data. Over the past 6 years, our model has been significantly more successful than the other common methods such as tournament seedings, the AP and ESPN/USA Today polls, the RPI, and the Sagarin and Massey ratings. © 2006 Wiley Periodicals, Inc. Naval Research Logistics, 2006.     

On the bivariate negative binomial distribution of mitchell and paulson

The bivariate negative binomial distribution of Mitchell and Paulson [17] for the case b = c = 0 is shown to be equivalent to the accident proneness model of Edwards and Gurland [4] and Subrahmaniam [19,20]. The diagonal series expansion of its joint probability function is then derived. Two other formulations of this distribution are also considered: (i) as a mixture model, which showed how it arises as the discrete analogue to the Wicksell-Kibble bivariate gamma distribution, and (ii) as a consequence of the linear birth-and-death process with immigration.     

A reliability model with states of partial operation

This article generalizes the classical dichotomic reliability model to include states of partial operation. The generalized model can be considered as a special case of a general jump process. Both continuous and discrete state spaces are included. The relationship to cumulative damage shock models is discussed. Properties of the model are investigated and these are illustrated via examples. The equivalence of three forms of component independence is proved, but this equivalence does not generalize to the property of zero covariance. Alternative forms of series and parallel connections and the effect of component replacement are discussed.     

An application of empirical bayes techniques to the simultaneous estimation of many probabilities

Consider the following situation: Each of N different combat units is presented with a number of requirements to satisfy, each requirement being classified into one of K mutually exclusive categories. For each unit and each category, an estimate of the probability of that unit satisfying any requirement in that category is desired. The problem can be generally stated as that of estimating N different K-dimensional vectors of probabilities based upon a corresponding set of K-dimensional vectors of sample proportions. An empirical Bayes model is formulated and applied to an example from the Marine Corps Combat Readiness Evaluation System (MCCRES). The EM algorithm provides a convenient method of estimating the prior parameters. The Bayes estimates are compared to the ordinary estimates, i.e., the sample proportions, by means of cross validation, and the Bayes estimates are shown to provide considerable improvement.     

Evaluation of costs of ordering policies in large machine repair problems

This article examines a version of the machine repair problem where failures may be irreparable. Since the number of machines in the system keeps decreasing, we impose a fixed state-dependent ordering policy of the type often encountered in inventory models. Although the system is Markovian, the number of states becomes very large. The emphasis of the article, therefore, is on deriving computationally tractable formulas for the steady-state probabilities, the long-run average cost per unit time, and the vector of expected discounted costs. When the state space is so large that exact computations may be infeasible, we propose approximations which are relatively quick and simple to compute and which yield very accurate results for the test problems examined.     

On a sequential test procedure with delayed observations

This paper considers sequential test procedures to decision problems where there exists time delays in obtaining observations.     

War reserve spares kits supplemented by normal operating assets

In peacetime, base stock levels of spares are determined on the assumption of normal resupply from the depot. In the event of war, however, a unit must be prepared to operate from stock on hand for a period of time without being resupplied from the depot. This paper describes a mathematical model for determining such war reserve spares (WRS) requirements. Specifically, the model solves the following kind of optimization problem: find the least-cost WRS kits that will keep the probability of a stockout after K cannibalizations less than or equal to some target objective α. The user of the model specifies the number of allowable cannibalizations, and the level of protection that the kit is supposed to provide. One interesting feature of this model is that in the probability computation it takes into account the possiblility of utilizing normal base operating assets. Results of a sensitivity analysis indicate that if peacetime levels were explicitly taken into account when designing a WRS kit, a cost saving of nearly 40 percent could be effected without degrading base supply performance in wartime.     

A maximum utility solution to a vehicle constrained tanker scheduling problem

A modification to the Dantzig and Fulkerson Tanker Scheduling Problem is described. An insufficient number of vehicles and a utility associated with each vehicle delivery are assumed. The new problem is shown to be equivalent to a Transshipment Problem, the solution of which is the same as the maximal utility solution of the modified Tanker Scheduling Problem. An example is given.