There is growing interest in the role of sentiment in economic decision-making. However, most research on the subject has focused on positive and negative valence. Conviction Narrative Theory (CNT) places Approach and Avoidance sentiment (that which drives action) at the heart of real-world decision-making, and argues that it better captures emotion in financial markets. This research, bringing together psychology and machine learning, introduces new techniques to differentiate Approach and Avoidance from positive and negative sentiment on a fundamental level of meaning. It does this by comparing word-lists, previously constructed to capture these concepts in text data, across a large range of semantic features. The results demonstrate that Avoidance in particular is well defined as a separate type of emotion, which is evaluative/cognitive and action-orientated in nature. Refining the Avoidance word-list according to these features improves macroeconomic models, suggesting that they capture the essence of Avoidance and that it plays a crucial role in driving real-world economic decision-making.

Models based on the transformer architecture, such as BERT, have marked a crucial step forward in the field of Natural Language Processing. Importantly, they allow the creation of word embeddings that capture important semantic information about words in context. However, as single entities, these embeddings are difficult to interpret and the models used to create them have been described as opaque. Binder and colleagues proposed an intuitive embedding space where each dimension is based on one of 65 core semantic features. Unfortunately, the space only exists for a small data-set of 535 words, limiting its uses. Previous work (Utsumi, 2018, 2020; Turton et al., 2020) has shown that Binder features can be derived from static embeddings and successfully extrapolated to a large new vocabulary. Taking the next step, this paper demonstrates that Binder features can be derived from the BERT embedding space. This provides two things; (1) semantic feature values derived from contextualised word embeddings and (2) insights into how semantic features are represented across the different layers of the BERT model.

This paper provides a perspective on historical background, innovation and applications of Artificial Intelligence (AI) and Machine Learning (ML), data successes and systems challenges, national security interests, and mission opportunities for system problems. AI and ML today are used interchangeably, or together as AI/ML, and are ubiquitous among many industries and applications. The recent explosion, based on a confluence of new ML algorithms, large data sets, and fast and cheap computing, has demonstrated impressive results in classification and regression and used for prediction, and decision-making. Yet, AI/ML today lacks a precise definition, and as a technical discipline, it has grown beyond its origins in computer science. Even though there are impressive feats, primarily of ML, there still is much work needed in order to see the systems benefits of AI, such as perception, reasoning, planning, acting, learning, communicating, and abstraction. Recent national security interests in AI/ML have focused on problems including multidomain operations (MDO), and this has renewed the focus on a systems view of AI/ML. This paper will address the solutions for systems from an AI/ML perspective and that these solutions will draw from methods in AI and ML, as well as computational methods in control, estimation, communication, and information theory, as in the early days of cybernetics. Along with the focus on developing technology, this paper will also address the challenges of integrating these AI/ML systems for warfare.

In Communication Theory, intermedia agenda-setting refers to the influence that different news sources may have on each other, and how this subsequently affects the breadth of information that is presented to the public. Several studies have attempted to quantify the impact of intermedia agenda-setting in specific countries or contexts, but a large-scale, data-driven investigation is still lacking. Here, we operationalise intermedia agenda-setting by putting forward a methodology to infer networks of influence between different news sources on a given topic, and apply it on a large dataset of news articles published by globally and locally prominent news organisations in 2016. We find influence to be significantly topic-dependent, with the same news sources acting as agenda-setters (i.e., central nodes) with respect to certain topics and as followers (i.e., peripheral nodes) with respect to others. At the same time, we find that the influence networks associated to most topics exhibit small world properties, which we find to play a significant role towards the overall diversity of sentiment expressed about the topic by the news sources in the network. In particular, we find clustering and density of influence networks to act as competing forces in this respect, with the former increasing and the latter reducing diversity.

Online social networks provide users with unprecedented opportunities to engage with diverse opinions. At the same time, they enable confirmation bias on large scales by empowering individuals to self-select narratives they want to be exposed to. A precise understanding of such tradeoffs is still largely missing. We introduce a social learning model where most participants in a network update their beliefs unbiasedly based on new information, while a minority of participants reject information that is incongruent with their preexisting beliefs. This simple mechanism generates permanent opinion polarization and cascade dynamics, and accounts for the aforementioned tradeoff between confirmation bias and social connectivity through analytic results. We investigate the model’s predictions empirically using US county-level data on the impact of Internet access on the formation of beliefs about global warming. We conclude by discussing policy implications of our model, highlighting the downsides of debunking and suggesting alternative strategies to contrast misinformation.

Online social networks provide users with unprecedented opportunities to engage with diverse opinions. At the same time, they enable confirmation bias on large scales by empowering individuals to self-select narratives they want to be exposed to. A precise understanding of such tradeoffs is still largely missing. We introduce a social learning model where most participants in a network update their beliefs unbiasedly based on new information, while a minority of participants reject information that is incongruent with their preexisting beliefs. This simple mechanism generates permanent opinion polarization and cascade dynamics, and accounts for the aforementioned tradeoff between confirmation bias and social connectivity through analytic results. We investigate the model's predictions empirically using US county-level data on the impact of Internet access on the formation of beliefs about global warming. We conclude by discussing policy implications of our model, highlighting the downsides of debunking and suggesting alternative strategies to contrast misinformation.

We present the results of a preliminary study to test the hypothesis that it is possible to automatically identify opinions, in the form of conviction narratives, as they emerge in text data, and to measure and monitor how actors in the online news media influence others in the media to adopt similar narratives to their own. Narratives are represented in the form of sentiment that online news sources express about various topics. Our results suggest that there is evidence of specific news sources acting as opinion leaders, determining the narratives that others in the online media adopt.

At a time when economics is giving intense scrutiny to the likely impact of artificial intelligence (AI) on the global economy, this paper suggests the two disciplines face a common problem when it comes to uncertainty. It is argued that, despite the enormous achievements of AI systems, it would be a serious mistake to suppose that such systems, unaided by human intervention, are as yet any nearer to providing robust solutions to the problems posed by Keynesian uncertainty. Under the radically uncertain conditions, human decision-making (for all its problems) has proved relatively robust, while decision making relying solely on deterministic rules or probabilistic models is bound to be brittle. AI remains dependent on techniques that are seldom seen in human decision-making, including assumptions of fully enumerable spaces of future possibilities, which are rigorously computed over, and extensively searched. Discussion of alternative models of human decision making under uncertainty follows, suggesting a future research agenda in this area of common interest to AI and economics.

The financial crisis of 2008 was unforeseen partly because the academic theories that underpin policy making do not sufficiently account for uncertainty and complexity or learned and evolved human capabilities for managing them. Mainstream theories of decision making tend to be strongly normative and based on wishfully unrealistic idealized modeling. In order to develop theories of actual decision making under uncertainty, we need new methodologies that account for how human (sentient) actors often manage uncertain situations well enough. Some possibly helpful methodologies, drawing on digital science, focus on the role of emotions in determining people's choices; others examine how people construct narratives that enable them to act; still others combine qualitative with quantitative data.

Making decisions under uncertainty is at the core of human decision-making, particularly economic decision-making. In economics, a distinction is often made between quantifiable uncertainty (risk) and un-quantifiable uncertainty (Knight, 1921). However, this distinction is often ignored by, in effect, the quantification of unquantifiable uncertainty, through the assumption of subjective probabilities in the mind of the human decision makers (Savage, 1954). This idea has is also reflected in developments in artificial intelligence (AI). However, there are serious reasons to doubt this assumption, which are relevant to both AI and economics. Some of the reasons for doubt relate directly to problems that AI has faced historically, that remain unsolved, but little regarded. AI can proceed on a prescriptive agenda, making engineered systems that aid humans in decision-making, despite the fact that these problems may mean that the models involved have serious departures from real human decision-making, particularly under uncertainty. However, in descriptive uses of AI and similar ideas (like the modelling of decisionmaking agents in economics), it is important to have a clear understanding of what has been learned from AI about these issues. This paper will look at AI history in this light, to illustrate what can be expected from models of human decision-making under uncertainty that proceed from these assumptions. Alternative models of uncertainty are discussed, along with their implications for examining in vivo human decision-making uncertainty in economics.

Recursive Bayesian estimation using sequential Monte Carlos methods is a powerful numerical technique to understand latent dynamics of nonlinear non-Gaussian dynamical systems. It enables us to reason under uncertainty and addresses shortcomings underlying deterministic systems and control theories which do not provide sufficient means of performing analysis and design. In addition, parametric techniques such as the Kalman filter and its extensions, though they are computationally efficient, do not reliably compute states and cannot be used to learn stochastic problems. We review recursive Bayesian estimation using sequential Monte Carlo methods highlighting open problems. Primary of these is the weight degeneracy and sample impoverishment problem. We proceed to detail synergistic computational intelligence sequential Monte Carlo methods which address this. We find that imbuing sequential Monte Carlos with computational intelligence has many advantages when applied to many application and problem domains.

Conviction narrative theory (CNT), a social psychological approach to the way economic agents take deisions under Knightian uncertainty, together with the new methodology of directed algorithmic text analysis (DATA), provide the opportunity for a theory of economic sentiment or animal sprits grounded in empirical facts. Applying DATA to the full text of the daily Reuters news feeds from January 1996 through November 2013, we derive an “animal spirits” series for both the US and the UK economy. Both series inform the movements in real GDP over the period. For example, in both countries there is a marked downturn in animal spirits in June 2007, well in advance of other indicators of the coming recession. The series may also explain why the subseqeunt recovery has been exceptionally weak from a historical perspective.

We develop social network and relative sentiment shift analysis techniques to study how financial narratives influence financial markets. First, we analyze Reuters News articles focusing on narratives about Fannie Mae. Second, we analyze Broadband and Energy narratives in the Enron Corporation email database. Combining datasets we show that phantastic object narratives can be detected and tracked as they develop and spread through networks to lead to a disconnect between narrative and underlying reality. The methods may be applicable to other text datasets to create early warnings.

Multi-objective optimization problems consist of numerous, often conflicting, criteria for which any solution existing on the Pareto front of criterion trade-offs is considered optimal. In this paper we present a general-purpose algorithm designed for solving multi-objective prob- lems (MOPS) on graphics processing units (GPUs). Specifically, a purely asynchronous multi-populous genetic algorithm is introduced. While this algorithm is designed to maximally utilize consumer grade nVidia GPUs, it is feasible to implement on any parallel hardware. The GPU's mas- sively parallel architecture and low latency memory result in +125 times speed-up for proposed parametrization relative to single threaded CPU implementations. The algorithm, NSGA-AD, consistently solves for so- lution sets of better or equivalent quality to state-of-the-art methods.

Much research has been conducted in recent years applying support vector machines (SVMs) for financial forecasting. Financial time series have been shown to be very noisy and difficult to generalize: directional accuracies are often close to the class distribution. In order to improve the accuracy of our predictions, we look at applying rejection to the decision outputs of the SVM model. We study whether accuracies can be improved by rejecting based on a) the distance from the separating hyperplane, b) the probabilistic SVM output, and c) the magnitude of the support vector regression output. We test on the gold future financial contract with 6125 out of sample points covering 2010. The results show small but insignificant accuracy gains but substantial economic improvements when applying the non-probabilistic reject methods. Further, for the margin based rejection method, we observe a strong relationship between the rejection rate and the classification accuracy; This helps with the heuristic that distance-from-the-margin can be associated with confidence.