Przejdź do treści

1. Artificial Intelligence in Finance

Welcome to the Artificial Intelligence (AI) section. This area explores machine learning and AI applications in quantitative finance, from data processing to predictive modeling. The information displayed in this section is coming from different sources such as books, articles or even certification courses I've taken like GARP-RAI.

1.1 Overview

Artificial Intelligence and Machine Learning are transforming quantitative finance, enabling sophisticated pattern recognition, prediction, and decision-making at scale.

The term AI was coined in 1956 by John MacCarthy together with Marvin Minsky of Harvard, Nathan Rochester of IBM, and Claude Shannon of Bell. All together echoed Turing's notion 1 stating that learning-related process and other feature of intelligence can be in principle desscribed such that a machine may simulate it.

Turing Machine

The Turing Machine is an ideal, abstract computational model rather than a physical device. Conceived by Alan Turing in 19363, and it serves as a mathematical framework for understanding what computation fundamentally means. Some characteristics are:

  • Abstract Nature: A theoretical device designed to formalize the concept of computability without physical constraints
  • Simple Components: Consists of an infinite tape divided into cells, a read/write head, and a finite state control
  • Universal Principle: Capable of simulating any algorithm or computational process, establishing the foundation for modern computer science
  • Deterministic Operation: Follows precise rules based on the current state and input symbol

The Turing Machine demonstrated that complex computations could be reduced to elementary mechanical operations by reading symbols, writing symbols, and moving along a tape. This formalization proved that any algorithmic process could theoretically be executed by such a machine, a concept known as Turing Completeness.

Though never built as originally described, the Turing Machine's principles directly influenced the architecture of programmable digital computers that emerged in the following decades.

1.1.1 Classical AI

Classical Artificial Intelligence, also called symbolic AI or Good Old-Fashioned AI (GOFAI), is an approach to building intelligent systems based on explicit rules, logic, and symbolic reasoning. Instead of learning patterns from large datasets, classical AI systems represent knowledge in structured forms such as symbols, rules, and logical statements, by means manipulating them through well-defined algorithms.

One of the most influential figures in the foundations of classical AI was Alan Turing. Although he worked before AI became an established field, Turing laid the theoretical groundwork for modern computing. However, mathematical logicians explored before Turing the solution to the question "What are the fundamental capabilities and limitations of computers?". Then, in his 1936 paper introducing the concept of the Turing machine, he formalized the idea of computation as a mechanical process, as a revolutionary idea beyond the work of other mathematitians on computability theory (See Sipser 4, chapter 0). This abstract model demonstrated that any calculation could, in principle, be performed by a sufficiently powerful machine following simple symbolic rules. This insight became fundamental to the development of programmable digital computers, which later made AI research possible.

Turing also contributed directly to philosophical and practical discussions about machine intelligence. In his 1950 paper “Computing Machinery and Intelligence,” he proposed what is now known as the Turing Test. The test suggests that if a machine can engage in a conversation indistinguishable from that of a human, it could be considered intelligent. This idea shifted the debate about AI from abstract definitions of “thinking” to observable behavior, influencing generations of AI researchers.

Classical AI systems that emerged in the decades after Turing’s work—such as logic-based theorem provers and early expert systems—were built on the assumption that intelligence could be achieved through symbolic manipulation, an idea closely aligned with Turing’s vision of computation. His pioneering theories about algorithms, machine processes, and the possibility of machine intelligence provided both the conceptual and technical foundation for the classical AI movement.

Artificial General Intelligence

Artificial General Intelligence (AGI), also known as strong AI, refers to a hypothetical AI system capable of understanding, learning, and applying knowledge across any intellectual domain with human-level performance. Unlike current narrow AI systems, which excel at specific tasks, AGI would possess advanced features like flexibility, adaptability, and general reasoning abilities.

Artificial General Intelligence would demonstrate several distinguishing characteristics. It would possess domain agnosticism, enabling the transfer of knowledge and skills across diverse fields without requiring task-specific retraining. Additionally, AGI systems would exhibit self-directed learning capabilities, identifying relevant information and improving performance independently. Such systems would be capable of abstract reasoning, working with concepts, analogies, and novel problems without explicit instruction. A key feature would be contextual understanding—grasping nuance, intent, and real-world implications of decisions. Finally, AGI would achieve results with energy efficiency comparable to biological intelligence, requiring computational resources on par with human cognition.

While AGI remains theoretical and is a subject of significant debate in AI research, it represents the ultimate goal of many AI researchers. Current systems, including large language models and deep learning architectures, are considered narrow AI—powerful within their domains but lacking the generalized capabilities of AGI. The timeline for achieving AGI, and whether it is fundamentally achievable, remains an open question in both academia and industry.

1.1.1.1 Deep Blue and Symbolic Game Playing

Deep Blue, developed by IBM and first deployed in 1997, represents a landmark achievement in classical AI. The system famously defeated world chess champion Garry Kasparov, demonstrating that machines could outperform humans in complex strategic domains. However, Deep Blue's approach remained fundamentally symbolic and rule-based, aligning perfectly with classical AI principles.

Deep Blue's architecture relied on explicit game rules, heuristic evaluation functions, and brute-force search algorithms. Rather than learning patterns from data, the system used hand-crafted evaluation functions based on chess principles—material value, piece positioning, and tactical patterns—all expressed as symbolic rules. The engine evaluated millions of board positions per second using minimax algorithms with alpha-beta pruning, techniques rooted in formal logic and game theory.

The symbolic nature of Deep Blue's intelligence is crucial to understand. Every evaluation was deterministic and traceable to explicit rules programmed by human experts. The system did not learn or adapt during matches; instead, it applied pre-defined knowledge and computational brute force. This exemplifies classical AI's philosophy that intelligence emerges from manipulating well-defined symbolic representations according to logical rules, without requiring the learning mechanisms that characterize modern machine learning approaches.

1.2 Topics Covered

  • Deep Learning — Neural network fundamentals, backpropagation, and modern architectures for representation learning

  • Reinforcement Learning — Sequential decision-making algorithms including policy gradient methods (A2C, PPO) and value-based approaches for learning optimal strategies

Leveraging AI for intelligent financial decision-making

  1. See chapter 3 in Parsing the Turing Test book 2 ↩

  2. Robert Epstein, Gary Roberts, and Grace Beber. Parsing the Turing Test. Springer, 2008. URL: https://link.springer.com/book/10.1007/978-1-4020-6710-5. ↩

  3. Alan M. Turing. On computable numbers, with an application to the entscheidungsproblem. Proceedings of the London Mathematical Society, 42:230–265, 1936. Reprinted in 1937. ↩

  4. Michael Sipser. Introduction to the Theory of Computation. Cengage Learning, 3rd edition, 2012. ISBN 978-1133187790. ↩