This study analyzed high-frequency tick-by-tick data of carbon emission allowances in the European emissions trading system, resampled into 5 -minute intervals, to identify innovative price patterns and extract key features for anticipatory insights. A dataset of pattern sequences and features was constructed, and a GAN model with LSTM and Self-Attention layers was used to generate synthetic sequences, producing artificial interval data. The synthetic data was validated using key financial metrics, demonstrating it effectively emulates real market behavior. The results support applications in risk management, algorithmic trading, market simulation, and stress testing under varying conditions.
It is challenging to consistently outperform benchmark indices and provide attractive long-run risk-return ratios under different economic scenarios. In this work, newly developed machine learning techniques are proposed to build robust portfolios that potentially can perform well under different stock market dynamics. Hierarchical clustering is used to assign stocks weights by minimizing the overall portfolio volatility (i.e. risk parity). The resulting portfolios present low stock correlations, and the invested stocks tend to neutralize each others potential downside expectation, yielding more reliable out of sample results. A new scheme is proposed to improve the hierarchical allocation method by including a dynamic orthogonal projections method, which tends to reduce outliers and noise, while obtaining subspaces that potentially can discriminate the individual stocks contributions. The preliminary experimental results suggest that attractive portfolios can be obtained, improving on the traditional market benchmarks, while maintaining a high degree of diversification and lower volatility. These experiments focus on the out-of-sample backtesting performance of the proposed scheme on a developing stock market, and on its comparison with traditional market indices to evaluate its potential.
Stock market prediction is a complex and dynamic area of fintech, involving the assessment of risk, projecting returns, and analyzing trends. Researchers have demonstrated that signal decomposition can facilitate all of these tasks by untangling intricate data streams. This paper presents a novel approach to predicting short-term stock market returns. The proposed scheme integrates Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and the Mixture of Experts (MoE) method within an ensemble learning framework. Decomposed signals are encoded into embeddings via attention mechanisms to facilitate feature extraction, and an innovative training target tailored for the unique MoE architecture was shown to significantly boost prediction accuracy. In simulations using two real-world stock markets, the proposed method outperformed existing state-of-the-art methods on all relevant metrics.
This article presents a conventional asset penalizing game model in green finance. We use a Mean Field Game (MFG) approach to investigate how a large population of small interacting conventional asset holders influences green asset prices, resulting in a green asset premium. Firstly, we formulate the green premium for green assets in a decentralized exchange pool setting. Using a virtual curve to model how collective behavior affects green asset prices, we analyze green premium cost implications for agents. Then, we develop our MFG model for the transition of green assets. To address the MFG problem, we use the Pontryagin maximum principle to find the fix-point solution in a deterministic conventional asset price setting. In the numerical experiment, we expand it to the stochastic price case and use Monte Carlo simulations to obtain the equilibrium collective asset transfer rate under certain market conditions. For empirical analysis, we examine the explanatory power of our model concerning the real market's green premium.
The one-period trinomial option pricing model is well known in the literature as it considers three possible movement directions of the asset price. However, by equating the price of the option with the self-financing hedging portfolio at maturity, this yields a linear system of three equations with two unknowns that correspond to the coefficients for the delta hedging portfolio. Hence, the trinomial model is said to be incomplete, that is, there exists an infinite number of equivalent martingale measures. To deal with this incompleteness, this paper aims to price options via some robust linear regression techniques in order to mainly handle the problem of outliers that the least squares fails to consider. The proposed robust techniques are evaluated on numerical data, the results of which demonstrate their effectiveness for European call option pricing.
This research investigates liquidity dynamics in fractional ownership markets, focusing on illiquid alternative investments traded on a FinTech platform. By leveraging empirical data and employing agent-based modeling (ABM), the study simulates trading behaviors in sell offer-driven systems, offering insights into the mechanisms that influence liquidity in these markets. The ABM-based simulation model provides a data augmentation environment which allows for the exploration of diverse trading architectures and rules, offering an alternative to direct experimentation. This approach bridges academic theory and practical application, supported by collaboration with industry and Swiss federal funding. The paper lays the foundation for extensions, including the identification of a liquidity-maximizing trading environment and the design of a market maker, by simulating modifications to the current design of the investment platform using an ABM specified with empirical data.(1)
This paper introduces a novel two-dimensional (2D) time series forecasting model that integrates cohort behavior over time, addressing challenges in small data environments. We demonstrate its efficacy using multiple real-world datasets, showcasing superior performance in accuracy and adaptability compared to reference models. The approach offers valuable insights for strategic decision-making across industries facing financial and marketing forecasting challenges.
Traditional volatility forecasting often relies on generalized autoregressive conditional heteroskedasticity (GARCH) models. Moreover, conditional variance forecasts are first obtained, and then the volatility forecasts are obtained as the square root of conditional variance forecasts. However, these methods assume that log returns are either normally distributed or $t$ distributed and are less efficient for heavy-tailed $t$ distributions with excess kurtosis having infinite asymptotic variance. Unlike the existing work, the novelty of this paper is to use recently proposed data-driven neuro volatility models to obtain direct volatility forecasts utilizing neural network models (NNAR, RNN, LSTM, and GRU). The proposed approach eliminates the need to model conditional variance, thereby addressing inefficiencies in traditional methods. The forecast accuracies of single-layer, twolayer, and bidirectional neural networks are compared against GARCH-based models using the CBOE Volatility Index (VIX) and Bitcoin (BTC). The results demonstrate that the data-driven neuro volatility models significantly outperform the GARCH models.
This work introduces a novel approach to study financial networks and portfolio optimization by addressing the limitations of Pearson correlation, which only captures linear relationships between assets. The novelty of this study is to propose pairwise neuro correlations, which emphasize nonlinear relationships using non-linear neural network autoregressive (NNAR) models. The methodology involves, modeling non-stationary/non-linear stock prices using ARIMA and NNAR models, defining neuro correlations as the correlations of the innovations from NNAR models and constructing adjacency matrices for financial networks using neuro correlations. The study employs network-based community detection method to uncover hidden market structures and applies them to portfolio selection. Experimental results demonstrate that combining neural networks of stock prices and financial networks (for capturing asset relationships) enhances portfolio optimization, yielding portfolios with improved cumulative returns. This approach highlights the strengths of both techniques: neural networks' ability to model non-linearity and financial networks' ability to capture the relationship structure of assets in modeling complex financial relationships.
This study addresses critical gaps in financial risk assessment and portfolio optimization by integrating advanced machine learning (ML) and deep learning (DL) techniques to handle the complexities of non-linearity, non-normality, and dynamic correlations among financial assets. This study comprehensively analyzes various dimensionality reduction techniques across different financial assets and time periods. By extracting non-linear features and constructing dynamic, data-driven covariance matrices, both linear and non-linear interactions among assets have been captured. This novel methodology hybridizes ML/DL and statistical approaches to enhance the robustness and resilience of portfolio optimization. The findings demonstrate significant improvements in profitability and stability under varying market conditions, offering a substantial advancement over prior studies. Therefore, this research provides a pioneering framework for more accurate and dynamic financial analysis, setting a novel standard in this research direction.
It is difficult to predict stock prices and many indicators are used for this purpose. This difficulty is even greater in short-term transactions. Regardless of the term, it should definitely be used for buying and selling timing. In this study, we propose an approach to predict this timing. This paper addresses challenges in time series forecasting using Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks, commonly employed in stock market forecasting, such as overfitting and extended learning times. To mitigate these issues, the PEC-W preprocessing framework, generally adapted from Convolutional Neural Networks (CNN), is applied to enhance forecasting accuracy without altering model parameters. This approach incorporates normalization and data augmentation to prevent overfitting, and utilizes Discrete Wavelet Transform (DWT) to reduce learning times while preserving temporal characteristics. Aggregation through averaging and mean subtraction further improves data visibility and model accuracy. The effectiveness of the PEC-W technique is validated using Explainable Artificial Intelligence (XAI) method, such as SHAP, which confirm the robustness of the enhanced forecasting approach.
This study assesses the Large Language Models (LLMs) in creating investment portfolios. We implement a few-shot learning technique, followed by Retrieval Augmented Generation (RAG) enhanced with comprehensive up-to-date financial data, using Meta's latest LLM, Llama 3.1-8b. In the first phase, We assess the models' efficacy using key financial indicators, including total returns, annualized volatility, riskadjusted performance (Sharpe ratio), potential loss estimates (value-at-risk), and their pre-training knowledge with the S&P 500 Index performance baseline. In the second phase, we enhance the LLM's knowledge base by RAG and the latest historical and statistical metrics (such as earnings per share (EPS), dividends per share (DPS), profit margin, and many more) for each asset from different classes. The study constrains model inputs to specific sets of financial assets, such as equities, exchangetraded funds (ETFs), commodities, cryptocurrencies, and bonds. To evaluate model performance and adaptability, we analyzed across two distinct time frames: (1) within the models' training data cutoff, and (2) from the cutoff date to the present. This approach enables the assessment of model generalization to past and present market conditions. The research quantifies LLMs’ capabilities in financial asset allocation, comparing baseline performance against RAG-augmented strategies. Our results demonstrate that RAG-enhanced LLM significantly outperforms vanilla LLM in portfolio construction across various asset classes. We contemplate that these results could influence AI-driven financial decision-making processes such as automated trading, real-time sentiment analysis, and investment management.
Recent advancements in large language models (LLMs) have demonstrated their potential to significantly impact finance trading, particularly through sentiment analysis. The cryptocurrency market, known for its volatility and unpredictability, often renders price-based trading approaches inadequate. This necessitates the adoption of more sophisticated techniques such as market sentiment analysis, which can benefit from the insights provided by LLMs. This study introduces an innovative method that integrates sentiment analysis derived from five distinct LLMs with deep reinforcement learning to devise a cryptocurrency trading strategy. Recognizing that LLM outputs cannot be guaranteed to be infallibly accurate, which contributing to the LLM hallucinations, this paper details the implementation of a stringent outlier detection and removal process. By adopting a “Trust-The-Majority” strategy, the research aims to ensure that trading decisions are informed by reliable sentiment data. In addition, sentiment scores are traditionally timestamped to the publication of news or social media posts. To more accurately reflect the actual impact of such information on market sentiment, this study applies the Ebbinghaus Forgetting Curve to model the waning influence of information over time. This allows for a more nuanced understanding of how news affects market dynamics. The enhanced sentiment scores, in conjunction with traditional market data such as OHLCV (Open, High, Low, Close, Volume), are utilized by a deep reinforcement learning model to make trading decisions. Experimental results demonstrate that the proposed multi-LLM sentiment-driven framework improves trading performance in the fast-paced cryptocurrency market. The methodology outlined in this paper offers a solid foundation for incorporating real-time market sentiment analysis into financial applications.
Financial sentiment analysis is the task of evaluating and quantifying the emotions and opinions expressed in financial news, reports, or social media to help investors and institutions make informed decisions. Financial institutions have been actively exploring the use of large language models (LLMs) to analyse market sentiment signals for a more nuanced understanding of a broader context. However, issues such as the scale of training data, model complexity, and the potential for human oversight can introduce or even amplify bias in these systems. Representation bias is a common challenge for LLMs as training data fail to properly represent the target groups, hence causes harmful bias in general-purpose use. Therefore, replacing current solutions with LLMs in financial organisations requires a robust evaluation methodology to ensure fairness. This paper investigates a three-level bias evaluation approach that specifically focuses on representation bias and presents a baseline evaluation of the FinBERT model. Step 1 uses a synthetic dataset that explicitly reveals sources of bias, structured as probability- and embedding-based evaluation recipes. Step 2 evaluates the model against data released by another country (e.g. Indian News dataset) to assess its performance in relation to more implicit biases. Step 3 examines individual problematic samples using token-based interpretability methods (e.g. integrated gradients). This paper presents the application of this structured bias evaluation process and its results on the FinBERT model. The evaluation code and dataset are available on GitHub (https://github.com/asabuncuoglu13/faid-test-financial-sentiment-analysis).
ESG (Environmental, Social, Governance)-thematics play an important role for investing as they can represent additional risk for a company and subsequently its stock price. These ESG-thematics can be extracted from alternative data sources such as news data. However, the extend to which an ESG-event influences a company can vary greatly depending on the ESG-aspect of the event and the current expectations of the affected company. In this paper, we present an approach where we leverage Dynamic Time Warping (DTW) in order to identify periods in which certain ESG-aspects are closely aligned with the stockprice movement of a company. We demonstrate this approach using more than 50,000 news articles mined for 42 german companies between January 2023 and September 2024, (i) answering the question of which aspects are important for a company in what time-frames and (ii) showing that this approach increases the performance of an ESG-sensitive portfolio by a large amount.
This paper proposes a novel framework for applying reinforcement learning (RL) within agent-based models (ABMs) to study labor market dynamics in labor economics. ABMs provide a flexible platform for simulating economic systems, particularly by modelling heterogeneous agents and their complex interactions, which effectively capture non-linear and emergent phenomena in labor markets. We extend an existing labor market model by integrating it into an RL framework, using the Deep Deterministic Policy Gradient (DDPG) algorithm to train agents to maximize profits, and comparing their performance with bounded-rational agents governed by predefined policies. Our findings show that RL agents, depending on the level of competition and rationality in the market, spontaneously learn distinct strategies, which significantly impact outcomes, such as unemployment and wage distribution. This work underscores the importance of ABMs in analyzing labor market dynamics and illustrates how RL-equipped agents can learn optimal strategies in evolving economic environments, offering a robust tool for policy analysis and exploring the complexities of labor market behavior.
Skewed data distribution poses many challenges in various domains, including the financial sector. Information about a company's potential bankruptcy is crucial for financial institutions and decision-making managers. As bankruptcy prediction has been a key concern for practitioners and research workers for decades, advancements in machine learning offer numerous methods to address imbalanced data scenarios with promising results. Among these, ensembles and neural networks showed remarkable results regarding prediction effectiveness. In this study, we propose a novel deep ensemble boosting approach to overcome imbalanced scenarios called Boosting TabNet. Achieved results showed promising results, specifically in the case of real-world datasets characterized by significant class imbalance ratios. Boosting TabNet outperformed or achieved results comparable to other approaches in most utilized datasets. The highest achieved score in terms of geometric mean (GM) score was over 96%.
The over-the-counter (OTC) government bond markets are characterised by their bilateral trading structures, which pose unique challenges to understanding and ensuring market stability and liquidity. In this paper, we develop a bespoke ABM that simulates market-maker interactions within a stylised government bond market. The model focuses on the dynamics of liquidity and stability in the secondary trading of government bonds, particularly in concentrated markets like those found in Australia and the UK. Through this simulation, we test key hypotheses around improving market stability, focusing on the effects of agent diversity, business costs, and client base size. We demonstrate that greater agent diversity enhances market liquidity and that reducing the costs of market-making can improve overall market stability. The model offers insights into computational finance by simulating trading without price transparency, highlighting how micro-structural elements can affect macro-level market outcomes. This research contributes to the evolving field of computational finance by employing computational intelligence techniques to better understand the fundamental mechanics of government bond markets, providing actionable insights for both academics and practitioners.
Large models have shown unprecedented capabilities in natural language processing, image generation, and most recently, time series forecasting. This leads us to ask the question: treating market prices as a time series, can large models be used to predict the market? In this paper, we answer this by evaluating the performance of the latest time series foundation model TimesFM on price prediction. We find that due to the irregular nature of price data, directly applying TimesFM gives unsatisfactory results and propose to fine-tune TimeFM on financial data for the task of price prediction. This is done by continual pre-training of the latest time series foundation model TimesFM on price data containing 100 million time points, spanning a range of financial instruments spanning hourly and daily granularities. The fine-tuned model demonstrates higher price prediction accuracy than the baseline model. We conduct mock trading for our model in various financial markets and show that it outperforms various benchmarks in terms of returns, sharpe ratio, max drawdown and trading cost.
Inspired by the recently proposed Kolmogorov-Arnold Networks (KANs), we introduce the KAN-based Option Pricing (KANOP) model to value American-style options, building on the conventional Least Square Monte Carlo (LSMC) algorithm. KANs, which are based on Kolmogorov-Arnold representation theorem, offer a data-efficient alternative to traditional Multi-Layer Perceptrons, requiring fewer hidden layers to achieve a higher level of performance. By leveraging the flexibility of KANs, KANOP provides a learnable alternative to the conventional set of basis functions used in the LSMC model, allowing the model to adapt to the pricing task and effectively estimate the expected continuation value. Using examples of standard American and Asian-American options, we demonstrate that KANOP produces more reliable option value estimates, both for single-dimensional cases and in more complex scenarios involving multiple input variables. The delta estimated by the KANOP model is also more accurate than that obtained using conventional basis functions, which is crucial for effective option hedging. Graphical illustrations further validate KANOP's ability to accurately model the expected continuation value for American-style options.
