Predicting FTSE 100 Stocks with Graph Neural Networks

by

Author: Denis Avetisyan

A new approach leverages the power of graph convolutional networks to forecast stock prices, achieving improved performance with streamlined feature engineering.

The A3TGCN model, implemented in PyTorch, manifests as a string representation-a testament to the fragility of any construct, for even the most meticulously coded theory risks dissolution beyond the event horizon of complexity.
The A3TGCN model, implemented in PyTorch, manifests as a string representation-a testament to the fragility of any construct, for even the most meticulously coded theory risks dissolution beyond the event horizon of complexity.

This review details the A3T-GCN model and demonstrates how annualized log returns enhance prediction accuracy while reducing computational complexity.

Accurately forecasting stock prices remains a persistent challenge in financial modeling, often hindered by the complex interdependencies within market data. This is addressed in ‘A3T-GCN for FTSE100 Components Price Forecasting’, which investigates a novel hybrid graph neural network architecture for predicting closing stock prices of UK FTSE100 constituents. The research demonstrates that utilizing annualized log-return features within the A3T-GCN model enhances prediction accuracy while simultaneously reducing computational demands, particularly for shorter-term forecasting horizons. Could this approach offer a viable pathway towards more efficient and reliable stock price prediction models in dynamic financial landscapes?

The Illusion of Efficiency: Markets and Models

The Efficient Market Hypothesis (EMH) serves as a fundamental benchmark in financial analysis, asserting that asset prices fully reflect all available information. This implies consistently achieving returns exceeding those of the overall market is exceedingly difficult, if not impossible, without access to privileged, non-public data. The EMH doesn’t claim markets are perfect, but rather that any informational advantage is quickly incorporated into prices, neutralizing attempts at systematic outperformance based on publicly known factors. Consequently, forecasting models are frequently evaluated against this baseline; a model’s success isn’t simply about predicting direction, but about demonstrably exceeding the returns one would expect from a passive, market-tracking investment strategy, accounting for risk. The hypothesis encourages a focus on minimizing costs and diversifying portfolios, as active management strategies often fail to justify their expense when measured against the efficiency benchmark.

Modern Portfolio Theory (MPT), a cornerstone of investment strategy, traditionally constructs portfolios based on the analysis of historical price data to estimate expected returns, volatility, and correlations between assets. This approach inherently assumes that market returns follow a normal distribution and that past performance is indicative of future results. However, financial markets frequently exhibit non-normal patterns – including “fat tails” representing extreme events – and are subject to dynamic shifts influenced by behavioral factors, geopolitical events, and evolving economic conditions. Consequently, portfolios optimized solely through historical data and normality assumptions may underestimate risk, fail to capture complex interdependencies, and prove vulnerable to unforeseen market shocks, necessitating the integration of more robust and adaptive modeling techniques.

Investor decisions are rarely made in a vacuum; rather, tax implications significantly shape portfolio construction and asset allocation. Capital gains taxes, dividend taxes, and even estate taxes introduce real-world constraints that deviate from the idealized assumptions of traditional financial models. Consequently, investors often prioritize minimizing tax liabilities-holding onto appreciated assets longer than economically optimal, favoring tax-advantaged accounts, or strategically timing the realization of gains and losses. These tax-driven behaviors introduce complexities that can distort market signals and impact overall returns, necessitating the development of more sophisticated modeling techniques that incorporate these practical constraints to provide a more accurate reflection of investor behavior and market dynamics.

The analysis demonstrates correlations based on both fundamental and returns data.
The analysis demonstrates correlations based on both fundamental and returns data.

Beyond Time: Modeling Markets as Networks

Traditional financial forecasting relies heavily on time-series analysis, treating each asset in isolation or considering limited lagged relationships. Graph Neural Networks (GNNs) represent a paradigm shift by explicitly modeling the interdependencies within financial markets. Unlike time-series methods which primarily analyze sequential data, GNNs facilitate the representation of financial instruments – such as stocks, bonds, or currencies – as nodes within a graph. The connections, or edges, between these nodes represent statistical relationships, such as correlation or causal influence, allowing the network to capture systemic risk and information flow. This interconnected representation enables GNNs to aggregate information from related assets, potentially leading to more accurate and robust forecasts compared to methods that ignore these complex relationships. The ability to model and leverage these connections is particularly relevant in modern finance where assets are increasingly interconnected through portfolios, derivatives, and global markets.

Graph Neural Networks (GNNs) model financial markets by representing individual stocks as nodes within a graph structure. Edges connecting these nodes represent relationships between stocks, frequently quantified using statistical measures like Pearson or Spearman correlation coefficients, which indicate the degree of linear or rank-based association between their price movements. This graph-based approach allows GNNs to move beyond analyzing stocks in isolation, instead capturing systemic risk and interdependencies. Information, such as price changes or trading volume, propagates across the network via these edges, effectively modeling how events affecting one stock can influence others. The network structure therefore enables the model to learn complex, non-linear relationships and identify patterns that traditional time-series analysis might miss, offering a more holistic view of market dynamics.

The A3T-GCN architecture combines Graph Convolutional Networks (GCNs), Long Short-Term Memory (LSTM) networks, and attention mechanisms to improve financial forecasting. GCN layers process the graph structure, enabling the model to aggregate information from interconnected stocks. LSTM layers then capture the temporal dependencies within each stock’s historical data. Crucially, an attention mechanism dynamically weights the importance of different stocks and time steps during information propagation, allowing the model to focus on the most relevant relationships and patterns. This combined approach allows A3T-GCN to model both the static network of stock relationships and the evolving temporal dynamics of the market, exceeding the capabilities of models utilizing only one of these approaches.

Graph construction for financial forecasting using GNNs relies on quantifying relationships between stocks through statistical methods and pre-defined classifications. Pearson correlation measures the linear relationship between stock returns, providing a value between -1 and 1, while Spearman correlation assesses the monotonic relationship, offering robustness to outliers. These correlation coefficients are typically used to define edge weights, with higher absolute values indicating stronger relationships. Additionally, sector classification – grouping stocks by industry – establishes initial connections, recognizing inherent dependencies within sectors even if direct correlation is weak. This approach creates a partially connected graph that can then be refined by incorporating correlation data, allowing the GNN to learn from both inherent sectoral relationships and dynamically observed market interactions. The resulting adjacency matrix, A, defines the graph structure used as input to the GNN.

Validating the Signal: Performance and Metrics

Annualized log returns are calculated to standardize price changes over a year, providing a consistent basis for evaluating investment performance and identifying long-term trends. Using the natural logarithm of price changes – $ln(P_t / P_{t-1})$ – mitigates the effects of compounding and allows for easier statistical analysis. Annualizing this value – multiplying by the number of periods in a year – transforms short-term fluctuations into a yearly equivalent, thereby reducing the influence of daily or weekly market noise and emphasizing sustained price movements. This approach is particularly useful for comparing investments with different time horizons and assessing the stability of returns over extended periods.

The A3T-GCN architecture benefits from the inclusion of the Relative Strength Index (RSI) and Moving Average Convergence Divergence (MACD) indicators, which provide supplementary data regarding market dynamics. The RSI, a momentum oscillator, measures the magnitude of recent price changes to evaluate overbought or oversold conditions, typically ranging from 0 to 100. Conversely, the MACD indicator identifies trend changes by illustrating the relationship between two moving averages of prices. Integrating these indicators into the A3T-GCN model allows for the capture of short-term momentum and potential shifts in longer-term trends, improving the model’s ability to forecast price movements beyond what could be achieved with price data alone.

Model performance was quantitatively assessed using several error metrics to determine prediction accuracy. The model achieved a Mean Squared Error (MSE) of 0.0036, representing the average squared difference between predicted and actual values. The Root Mean Squared Error (RMSE), calculated as the square root of the MSE, was 0.0603, providing an interpretable measure of the standard deviation of the residuals. Additionally, the Mean Absolute Error (MAE) was determined to be 0.0439, quantifying the average magnitude of the errors without considering their direction. These metrics collectively indicate a low degree of error and contribute to the overall reliability of the predictive model.

The A3T-GCN architecture builds upon Convolutional Neural Networks (CNNs) by integrating both temporal information and attention mechanisms to improve predictive capabilities. This enhancement allows the model to consider time-series dependencies and focus on the most relevant features within the data. Evaluation on the test dataset demonstrates strong performance, achieving a Mean Relative Error (MRE) of 3.46% and an R-squared value of 0.9936105. The low MRE indicates a small average percentage difference between predicted and actual values, while the R-squared value signifies that approximately 99.36% of the variance in the dependent variable is explained by the model, confirming a robust model fit.

The training loss, measured as Mean Absolute Error, consistently decreased with each epoch, indicating successful model convergence.
The training loss, measured as Mean Absolute Error, consistently decreased with each epoch, indicating successful model convergence.

Beyond Prediction: Implications and Future Horizons

The A3T-GCN Architecture demonstrates a significant advancement in predictive accuracy for financial time series, offering tangible benefits for both portfolio construction and risk management. By more accurately forecasting asset returns and volatilities, the model enables investors to build portfolios optimized for higher returns with reduced exposure to downside risk. Specifically, the enhanced predictions facilitate more precise asset allocation, allowing for the identification of undervalued assets and the avoidance of overvalued ones. Furthermore, the model’s ability to capture complex relationships between assets improves risk assessment, moving beyond traditional methods that often underestimate systemic risk. This improved granularity in risk modeling allows for the implementation of more effective hedging strategies and the development of robust portfolios capable of withstanding market fluctuations, ultimately contributing to more stable and profitable investment outcomes.

Conventional financial modeling often relies on the assumption of asset independence, a simplification that overlooks the intricate relationships driving market behavior. This architecture, however, directly addresses this limitation by explicitly modeling the complex interdependencies between assets. It recognizes that price movements aren’t isolated events, but rather propagate through a network of connections, where the performance of one asset can significantly influence others. By capturing these relationships – whether driven by sector correlations, shared investor sentiment, or macroeconomic factors – the model generates a more holistic and realistic representation of the financial landscape. This allows for a nuanced understanding of risk and return, moving beyond the limitations of models that treat assets as operating in isolation and potentially unlocking more effective investment strategies.

The architecture’s potential extends beyond the current scope, inviting investigation into its applicability across diverse financial landscapes, including commodities, foreign exchange, and even cryptocurrency markets. Researchers are poised to explore the benefits of incorporating unconventional data-such as news sentiment, social media trends, and satellite imagery-to refine predictive capabilities and gain a more holistic understanding of market dynamics. This integration of alternative data sources, coupled with the framework’s adaptability, promises to unlock novel insights and potentially reveal previously hidden relationships that influence asset pricing and market behavior, ultimately leading to more robust and nuanced financial models.

The architecture’s potential extends beyond its current capabilities through the incorporation of dynamic graph structures. Rather than relying on a static representation of asset relationships, future iterations could allow the network topology to evolve over time, reflecting shifts in market conditions and interdependencies. This dynamic adjustment, coupled with adaptive learning algorithms – techniques that enable the model to refine its parameters based on incoming data – promises to significantly enhance both performance and robustness. Such advancements would allow the model to not only capture complex relationships but also to respond effectively to unforeseen market changes, potentially mitigating risks associated with static, historically-trained models and improving the accuracy of financial predictions in volatile environments.

The pursuit of predictive accuracy in financial modeling, as demonstrated by this application of A3T-GCN to FTSE100 components, reveals a humbling truth about the limits of knowledge. This work, focused on enhancing performance through features like annualized log returns, underscores that every carefully constructed model is, in essence, a provisional understanding. As Marie Curie observed, “Nothing in life is to be feared, it is only to be understood.” The refinement of the A3T-GCN, while achieving demonstrable gains, doesn’t eliminate uncertainty; it merely shifts the event horizon of the unknown, acknowledging that even the most sophisticated law can dissolve when confronted with the inherent complexities of the market.

Beyond the Horizon

The application of graph convolutional networks to financial time series, as demonstrated, offers a glimpse of predictive capability. Yet, the cosmos generously shows its secrets to those willing to accept that not everything is explainable. The pursuit of ever-more-refined features – in this case, annualized log returns – feels akin to polishing the lenses while staring into an infinite darkness. Reducing computational demands is a practical victory, certainly, but it merely allows for a faster descent into the unknowable. Black holes are nature’s commentary on human hubris.

Future work will undoubtedly focus on hybrid models, integrating A3T-GCN with other deep learning architectures. The search for the ‘optimal’ configuration, however, risks mistaking correlation for causation. The market, unlike a neatly labeled dataset, possesses an inherent irrationality, a tendency to defy prediction. Perhaps the true challenge lies not in building more complex algorithms, but in acknowledging the fundamental limits of predictability.

The elegance of the A3T-GCN framework invites expansion to broader financial instruments and macroeconomic indicators. But one wonders if each successive refinement merely refines the illusion of control. The event horizon of complete financial forecasting remains perpetually out of reach, a humbling reminder that some mysteries are best left unsolved.

Original article: https://arxiv.org/pdf/2511.21873.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-01 18:23