Class Incremental Learning for Financial Data Streams and Two Classes of Methodologies

Larry Milner

doi:10.20944/preprints202409.1846.v1

Submitted:

24 September 2024

Posted:

24 September 2024

You are already at the latest version

Abstract

Class Incremental Learning (CIL) has emerged as a vital machine learning paradigm, particularly suited for environments with continuous data streams, such as those encountered in the financial industry. CIL enables models to progressively acquire and accumulate new knowledge without retraining the entire dataset, making it highly efficient for dynamic financial markets. In this paper, we explore the application of CIL to financial data streams, where high-frequency, high-dimensional data, such as stock prices and cryptocurrency transactions, evolve rapidly. The challenges inherent in CIL, including catastrophic forgetting—where the model loses previously learned knowledge—and intransigence—where it resists learning new classes—are addressed through both regularization-based techniques and dynamic architecture expansion. We present a comprehensive analysis of the algorithms designed to mitigate these challenges and highlight their relevance in financial applications like algorithmic trading, portfolio management, risk assessment, and fraud detection. By demonstrating how CIL enables models to adapt to new market conditions and asset classes while retaining past knowledge, we illustrate its critical role in advancing real-time financial decision-making. Through a detailed examination of the methodologies and their applications, this paper aims to provide insights into the future of adaptive financial models using Class Incremental Learning. We have also summarized methodologies into a model optimization-based class and a data focused class.

Keywords:

Class-incremental Learning

;

Financial Learning

;

Machine Learning

;

Optimization

Subject:

Computer Science and Mathematics - Artificial Intelligence and Machine Learning

Introduction

In recent years, financial data streams have grown in volume, complexity, and velocity, creating significant challenges for machine learning models. Traditional learning approaches often fall short in such dynamic environments due to their reliance on static training data and their inability to adapt to evolving patterns without constant retraining. Class Incremental Learning (CIL), a branch of incremental learning, provides a robust framework to tackle these challenges. CIL allows machine learning models to acquire and accumulate new knowledge progressively while retaining and optimizing previously learned information. Its potential for real-time adaptation makes it highly valuable in the financial industry, where markets are constantly shifting, and new patterns frequently emerge.

Incremental learning's ability to adjust progressively without forgetting previously learned information offers a distinct advantage in financial data streams, which are often subject to non-stationarity, unexpected changes, and the introduction of new financial instruments or market dynamics. This characteristic is crucial for financial models that need to respond quickly to new data, such as evolving market trends or the introduction of new asset classes, while still maintaining their predictive power for older market conditions.

Review on Past Literatures

Incremental learning has been studied extensively in [1,2,3]. Incremental learning aims to efficiently aggregate novel classes with prior classes in a comprehensive model. Class Incremental Learning (CIL) is a subfield of incremental learning designed to enable models to progressively handle an increasing number of classes over time. Unlike traditional learning systems, which typically require a fixed dataset with all classes present during training, CIL allows a model to update its knowledge base when new classes are introduced, without having to retrain the entire model from scratch. In financial applications, this is particularly important as new financial products, regulatory changes, and emerging economic factors continuously reshape the landscape.

CIL addresses two key issues inherent in its framework: catastrophic forgetting and intransigence. Catastrophic forgetting refers to the phenomenon where a model loses its ability to recall previous knowledge after learning new information. In financial systems, this could be disastrous, as losing insights into past market behavior can lead to significant financial losses. Intransigence, on the other hand, is the model's resistance to learning new classes while attempting to retain old knowledge, which is detrimental in dynamic environments like finance, where new data streams constantly introduce unfamiliar variables.

To overcome these challenges, numerous algorithms and methodologies have been developed to enhance the adaptability of CIL models. The primary focus is on ensuring that models can integrate new information while maintaining stable performance across both old and new classes. Approaches such as replay-based methods, where the model periodically re-learns a subset of the previous data, and regularization-based techniques, which penalize significant changes in model weights, have gained prominence. In the financial sector, these methods are particularly valuable for maintaining the predictive capacity of algorithms over time as they encounter a continuous flow of new financial data.

Class incremental learning is sometimes combined with graph neural network development [4]. Incremental learning is expected to conduct the current learning tasks without accessing the information of earlier tasks. An important assumption for incremental learning is that tasks are constructed into several subtasks without overlapping. This also makes the future machine unlearning tasks for privacy issues easier as show in [5].

Financial Data Streams and Incremental Learning

The class-incremental learning aims to learn from an evolved data sequence drawn from a non-stationary distribution. Most models assume a series of infinite data sequence. Financial data streams are characterized by high-frequency, high-dimensional, and often noisy data points that can originate from various sources, including stock prices, cryptocurrency transactions, macroeconomic indicators, and social sentiment analysis. One of the most prominent examples of data streams in finance is stock market trading data, where prices, volumes, and order book activity change rapidly, often within milliseconds. Additionally, data from alternative financial sources, such as cryptocurrency exchanges, exhibit similar patterns of volatility and complexity. For models tasked with predicting market movements, a static approach quickly becomes obsolete.

The advantages of CIL are evident in this context. A model trained on historical financial data might perform well initially but could become less effective as market conditions change. By implementing CIL, such a model can incorporate new patterns, asset classes, or market events without forgetting previously acquired knowledge. This capability is essential in the financial industry, where staying competitive requires adapting to both historical trends and new market dynamics.

For instance, in the case of high-frequency trading (HFT), where decisions are made in milliseconds based on live data, incremental learning allows models to update continuously with fresh information without halting operations for retraining. As new types of assets, such as cryptocurrencies, gain prominence, CIL models can seamlessly adjust to incorporate these new classes into their decision-making processes while still retaining their understanding of traditional asset classes like stocks, bonds, and commodities.

Challenges in Class Incremental Learning

Despite its potential, Class Incremental Learning faces significant challenges, particularly when applied to financial data streams. One of the most critical issues is catastrophic forgetting, where the model, upon learning new classes or patterns, forgets previously learned classes. In financial applications, this would translate into the inability of a model to recognize important historical market patterns after encountering new data. For instance, a stock market prediction model may forget the behavior of traditional equities markets when trained on cryptocurrency data, leading to suboptimal performance.

In response, researchers have developed various strategies to mitigate catastrophic forgetting. One popular technique is rehearsal, where a subset of previously encountered data is stored and replayed during training to remind the model of past information. In financial settings, this could involve replaying historical data from older financial instruments or time periods to ensure that the model retains an understanding of how traditional assets behave. Another method involves regularization techniques, where the model is penalized for making drastic changes to its weights when learning new information. This allows it to learn new data while preserving its existing knowledge base.

Another challenge in CIL is intransigence, where the model becomes resistant to learning new information to avoid forgetting old knowledge. In the financial industry, this could be a significant drawback, as markets frequently evolve, and new asset classes and economic indicators need to be incorporated. If a model is unable to adapt to these changes, its usefulness is diminished. To combat intransigence, researchers have focused on methods that increase the plasticity of the model while maintaining its stability, such as dynamic architectures that grow to accommodate new classes without overloading the system.

Moreover, the financial industry imposes unique constraints on CIL systems, such as the need for real-time processing, scalability, and handling high-dimensional data. The increasing volume of financial data, combined with the real-time nature of financial markets, requires CIL models to be both efficient and scalable. Techniques such as ensemble learning, where multiple models are trained on different subsets of the data and their predictions are aggregated, have shown promise in handling the complexity and size of financial data streams.

Model Optimization Methodology

In the domain of class-incremental learning (CIL) algorithms, particularly focusing on model construction, there are two predominant perspectives regarding how models adapt to new classes. These approaches deal with the challenge of learning from evolving data streams while ensuring minimal forgetting of previously learned information.

The first approach involves integrating newly acquired classes into an existing model with a fixed architecture. This approach has been seen in [6,7,8].This method often relies on regularization techniques to manage potential drifts in the model's weights and functions. Regularization helps in maintaining the stability of previously learned classes by imposing constraints on how much the model's parameters can change when learning new information. For instance, techniques like Elastic Weight Consolidation (EWC) and Learning without Forgetting (LwF) impose penalties on changes in weights to prevent significant forgetting. Moreover, this fixed-architecture approach leverages knowledge from previously encountered classes, using mechanisms such as distillation or rehearsal, to facilitate model updates without retraining from scratch. In financial applications, this could involve gradually incorporating new market data or asset classes while retaining knowledge of older, traditional financial products, ensuring consistent performance across all classes. The first approach can also be potentially combined with LSTM approach for financial portfolio management as shown in [9].

The second approach involves dynamically expanding the model structure to accommodate newly encountered classes. In this idea, we need to dynamically expanding the model structure to capture distinctive features that are newly encountered as shown in [10,11,12]. This approach recognizes that some new classes may introduce entirely different patterns or features that cannot be adequately represented by the existing model architecture. As a result, the model grows by adding perceptron, extending blocks, or even expanding entire networks to better capture the unique characteristics of the new data. This method offers a flexible way to ensure that the model can handle new complexity without compromising the performance of the previously learned classes. However, it requires careful management to avoid overfitting and computational inefficiency. In financial scenarios, dynamic expansion could be used to address the introduction of novel asset classes like cryptocurrencies, which may exhibit distinct market behaviors compared to traditional assets. By expanding the model, financial systems can adapt to these new dynamics while still retaining a strong understanding of more established financial products.

Both approaches regularization-based fixed architectures and dynamically expandable models play critical roles in developing robust class-incremental learning systems, particularly in environments such as financial data streams, where constant adaptation to new classes is crucial.

Data focused methodologies

In addition to model optimization techniques, a crucial stream of methodologies for Class Incremental Learning (CIL) focuses on data-focused methods. These methods emphasize leveraging earlier observations to enhance model performance, addressing the challenges of forgetting and intransigence by making strategic use of historical data. Data-focused approaches can be categorized into two main subsets: feature-based methods and replay-based methods. Both offer distinct advantages in maintaining model performance across evolving data streams, particularly in the context of financial data.

Feature-based methods focus on exploring the features or embeddings extracted from earlier training observations to gain a deeper understanding of the underlying patterns in the data. By analyzing these features, the model can better generalize and recognize common structures across new and old classes, reducing the risk of catastrophic forgetting. Feature-based techniques often involve transferring knowledge from earlier learned representations to assist the model in learning new classes.

In financial applications, this approach is especially useful for analyzing high-dimensional data such as stock prices, trading volumes, or macroeconomic indicators. For instance, a model trained to predict stock prices based on features like price-to-earnings ratios, moving averages, and sentiment analysis from social media can extract feature embeddings that are relevant to multiple asset classes. When new financial products, such as cryptocurrencies, are introduced, the model can rely on these previously learned features to help understand the new asset class without requiring complete retraining. This approach ensures that the model retains its ability to predict the behavior of traditional stocks while adapting to new data streams with minimal disruption.

One example of a feature-based CIL method is Learning without Forgetting (LwF), which uses earlier learned feature representations to help the model adapt to new classes. Instead of retaining actual data samples from previous classes, LwF uses the outputs of the model on previous tasks to guide learning in new tasks, ensuring that previously learned information is not overwritten.

Replay-based methods take a different approach by explicitly incorporating earlier data samples or pseudo-exemplars into the learning process for new classes. These methods combine old and new data streams, allowing the model to continually adapt and refine its predictions. The core idea is to periodically replay a subset of previously encountered samples alongside the new data to prevent the model from completely forgetting past knowledge.

In financial data streams, replay-based methods are highly effective for maintaining performance over time as market conditions evolve. For example, in algorithmic trading, a model might use replay-based techniques to replay a subset of historical trading data from past years or months while learning from new data streams in real time. By revisiting old data during the learning process, the model maintains its awareness of historical market trends, ensuring that it can still recognize patterns like seasonal effects, market crashes, or bull runs, even as it learns new trends introduced by current market data.

An example of a widely used replay-based method is Experience Replay, originally developed for reinforcement learning but now commonly applied in CIL contexts. In Experience Replay, a buffer stores a subset of previous experiences (or data samples), which are periodically replayed during the learning process. This method helps to smooth out learning and avoid catastrophic forgetting by reinforcing the model's memory of old data.

Another variant is Generative Replay, which generates synthetic (or pseudo) samples of previously learned classes using a generative model, such as a Generative Adversarial Network (GAN) or a Variational Autoencoder (VAE). Instead of storing actual data points, which could become computationally expensive, the model generates pseudo-exemplars that resemble past observations. These pseudo-exemplars are then combined with new data to ensure the model retains its previous knowledge while learning new classes. In financial applications, this method could be used to generate synthetic trading data that mimics past market conditions, ensuring the model maintains its ability to predict traditional asset behavior while adapting to new asset classes or regulatory environments.

Applications of Class Incremental Learning in the Financial Industry

Class Incremental Learning has several potential applications in the financial industry. Some of the most relevant areas include algorithmic trading, portfolio management, risk assessment, and fraud detection. By leveraging CIL, financial institutions can develop models that evolve with changing market conditions, enhancing both their decision-making capabilities and operational efficiency.

One of the most prominent applications of CIL is in algorithmic trading, where financial models make real-time trading decisions based on live market data. In this domain, the ability to learn from new data while retaining the knowledge of past trends is essential for maintaining competitiveness. CIL can help trading algorithms adjust to new market conditions, such as regulatory changes, new financial products, or shifts in market sentiment, without losing their understanding of historical market behavior. This is particularly important in high-frequency trading environments, where models need to process vast amounts of data quickly and accurately.

Portfolio Management: Another application of CIL is in portfolio management, where financial models must continuously adapt to changing market conditions and asset classes. With the rise of new asset classes like cryptocurrencies, CIL models can integrate these new assets into their analysis while retaining their knowledge of traditional financial instruments. This allows for more effective portfolio diversification and risk management, as the model can assess the relationships between new and old assets without requiring complete retraining.

Financial institutions also rely on CIL for risk assessment, where models are used to evaluate the risk associated with different investments, loans, or market conditions. As new risk factors emerge—such as geopolitical events, pandemics, or technological disruptions—CIL allows models to incorporate these new risks into their analysis without forgetting previously known risk factors. This leads to more accurate and comprehensive risk assessments, which are crucial for financial decision-making.

In high frequency trading, where models process vast amounts of data in milliseconds, replay-based methods can be used to replay historical data on specific trading strategies or market events. This ensures the model maintains knowledge of past market behaviors, such as volatility patterns, while learning new strategies based on real-time data streams. Feature-based methods could also extract key trading signals that remain consistent across different market regimes, allowing the model to generalize across varying market conditions.

Replay-based methods can be leveraged in portfolio management models to revisit old market data while adapting to new financial assets. For instance, a portfolio management model might replay data from traditional asset classes like stocks and bonds while learning from newer asset classes such as cryptocurrencies or decentralized finance (DeFi) products. Feature-based methods would assist by extracting common financial indicators, such as volatility or Sharpe ratio, to help the model make informed portfolio decisions across both old and new assets.

Financial fraud evolves rapidly, with new schemes emerging over time. Replay-based methods can be particularly valuable in fraud detection systems, where a subset of past fraud cases is periodically replayed during model training to ensure that the system retains its ability to detect earlier fraud patterns. Feature-based methods can extract important characteristics of fraudulent activities, allowing the model to generalize across different types of fraud, from identity theft to money laundering, even as new schemes emerge.

Future Directions and Conclusion

Class Incremental Learning holds tremendous potential for revolutionizing the way machine learning models are applied to financial data streams. As financial markets become more dynamic and complex, the need for models that can adapt to new data without forgetting old knowledge becomes increasingly important. By addressing the challenges of catastrophic forgetting and intransigence, CIL offers a promising approach to improving the adaptability and efficiency of financial models.

Future research in this area is likely to focus on further refining the techniques used to balance stability and plasticity in CIL models, as well as improving the scalability of these models to handle ever-increasing volumes of financial data. Additionally, the integration of CIL with other advanced machine learning techniques, such as reinforcement learning and deep learning, may provide even more powerful tools for financial institutions to navigate the complexities of modern markets.

In conclusion, Class Incremental Learning is a critical development in the field of machine learning, offering valuable solutions to the challenges posed by financial data streams. By enabling models to learn progressively and adapt to new information without forgetting the past, CIL provides a foundation for more robust and responsive financial models. As the financial industry continues to evolve, the application of CIL will undoubtedly play a key role in shaping the future of financial decision-making.

References

E. Belouadah, A. Popescu and I. Kanellos, "A comprehensive study of class incremental learning algorithms for visual tasks," Neural Networks, vol. 135, p. 38–54, 2021. [CrossRef]
G. M. Van de Ven, T. Tuytelaars and A. S. Tolias, "Three types of incremental learning," Nature Machine Intelligence, vol. 4, p. 1185–1197, 2022. [CrossRef]
D.-W. Zhou, Q.-W. Wang, Z.-H. Qi, H.-J. Ye, D.-C. Zhan and Z. Liu, "Deep class-incremental learning: A survey," arXiv preprint arXiv:2302.03648, 2023. [CrossRef]
Z. Wang, Y. Zhu, Z. Li, Z. Wang, H. Qin and X. Liu, "Graph neural network recommendation system for football formation," Applied Science and Biotechnology Journal for Advanced Research, vol. 3, p. 33–39, 2024. [CrossRef]
C. Qi, "Graph Unlearning: Mechanism and Future Direction for Machine Unlearning with Complex Relationships," Preprints, September 2024. [CrossRef]
Chaudhry, P. K. Dokania, T. Ajanthan and P. H. S. Torr, "Riemannian walk for incremental learning: Understanding forgetting and intransigence," in Proceedings of the European conference on computer vision (ECCV), 2018.
J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska and others, "Overcoming catastrophic forgetting in neural networks," Proceedings of the national academy of sciences, vol. 114, p. 3521–3526, 2017. [CrossRef]
S.-W. Lee, J.-H. Kim, J. Jun, J.-W. Ha and B.-T. Zhang, "Overcoming catastrophic forgetting by incremental moment matching," Advances in neural information processing systems, vol. 30, 2017.
Z. Li, B. Wang and Y. Chen, "Incorporating economic indicators and market sentiment effect into US Treasury bond yield prediction with machine learning," Journal of Infrastructure, Policy and Development, vol. 8, p. 7671, 2024. [CrossRef]
O. Ostapenko, M. Puscas, T. Klein, P. Jahnichen and M. Nabi, "Learning to remember: A synaptic plasticity driven framework for continual learning," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019.
J. Yoon, E. Yang, J. Lee and S. J. Hwang, "Lifelong learning with dynamically expandable networks," in 6th International Conference on Learning Representations, ICLR 2018, 2018.
J. Xu and Z. Zhu, "Reinforced continual learning," Advances in neural information processing systems, vol. 31, 2018.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.